align LRC from ready text.txt via stable-ts and auto-detect language

- add align_text_to_audio(): forced alignment without ASR, fast
- add _StableTSModels cache; detect language via stable-ts when language=None
- LRC stage: use existing text.txt, copy it if already LRC, fall back to ASR only
- filter junk lyrics from tags and ASR hallucinations (promo phrases, tics)
- mark empty ASR results with metadata.empty_text and skip them on resume
- warn in GUI when LRC is run with transcription disabled
- add mark_no_speech.py one-shot script for existing tracks

.gitignore

@@ -14,3 +14,4 @@ output/
 
 preset.json
 *.log
+combined_files.txt

core/config.py

@@ -16,6 +16,7 @@ class ConvertConfig:
     sample_rate: int = 44100
     channels: int = 2
     ffmpeg_path: str = "ffmpeg"
+    convert_concurrency: int = 4
 
 
 @dataclass
@@ -25,6 +26,7 @@ class SeparateConfig:
     isolate_stem: str = "vocals"
     unblend_path: str = "unblend"
     custom_kernels: bool = False
+    save_vocals_copy: bool = False
 
 
 @dataclass
@@ -43,6 +45,7 @@ class TranscribeConfig:
 @dataclass
 class ShazamConfig:
     enabled: bool = False
+    identify_concurrency: int = 4
 
 
 @dataclass
@@ -53,8 +56,24 @@ class LyricsConfig:
     use_tags_first: bool = True
     use_shazam_result: bool = True
     try_after_asr: bool = True
-    fuzzy_threshold: float = 0.6          # difflib ratio, 0..1
+    fuzzy_threshold: float = 0.6
     transcribe_if_identified_no_lyrics: bool = True
+    # NEW:
+    use_embedded_lyrics: bool = True     # read lyrics already stored in file tags first
+    save_unrecognized: bool = False      # write unresolved texts to JSONL for LRCLIB submission
+    unrecognized_path: str = ""          # "" = <output>/unrecognized_lyrics.jsonl
+
+
+@dataclass
+class LrcConfig:
+    enabled: bool = True             # LRC generation stage
+    embed_enabled: bool = True       # LRC embedding stage
+    skip_existing: bool = True       # skip tracks with a ready text.lrc
+    language: Optional[str] = "ru"
+    align_model: Optional[str] = None
+    device: str = "auto"
+    audio_source: str = "auto"
+    embed_in_place: bool = False
 
 
 @dataclass
@@ -65,12 +84,16 @@ class PipelineConfig:
     keep_intermediate: bool = True
     overwrite: bool = True
     mode: str = "sequential"
+    chunk_size: int = 64
+    skip_existing: bool = False     # skip files that already have text.txt
+    resume_partial: bool = True     # reuse existing vocals.wav / convert.wav
 
     convert: ConvertConfig = field(default_factory=ConvertConfig)
     separate: SeparateConfig = field(default_factory=SeparateConfig)
     transcribe: TranscribeConfig = field(default_factory=TranscribeConfig)
     shazam: ShazamConfig = field(default_factory=ShazamConfig)
     lyrics: LyricsConfig = field(default_factory=LyricsConfig)
+    lrc: LrcConfig = field(default_factory=LrcConfig)
 
     def to_dict(self) -> dict:
         return asdict(self)

core/lrc.py

@@ -0,0 +1,232 @@
+"""LRC generation, parsing, embedding, and reading embedded lyrics.
+
+Reading embedded lyrics checks all tag locations for a given audio format:
+  - MP3       -> ID3 USLT / SYLT
+  - FLAC/OGG  -> LYRICS, UNSYNCEDLYRICS (Vorbis comments)
+  - M4A/MP4   -> `©lyr` atom
+"""
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional
+
+
+_LRC_LINE_RE = re.compile(r"^\s*\[\d{1,3}:\d{2}(?:\.\d{1,3})?\]")
+_LRC_PARSE_RE = re.compile(r"\[(\d+):(\d+)\.(\d+)\](.*)$")
+
+
+@dataclass
+class LrcLine:
+    start: float
+    text: str
+
+
+def format_timestamp(seconds: float) -> str:
+    if seconds < 0:
+        seconds = 0.0
+    minutes = int(seconds // 60)
+    secs = seconds - minutes * 60
+    return f"[{minutes:02d}:{secs:05.2f}]"
+
+
+def segments_to_lrc(segments: list[dict]) -> str:
+    lines: list[str] = []
+    for seg in segments:
+        text = (seg.get("text") or "").strip()
+        if not text:
+            continue
+        # WhisperX can return multi-line text on long segments. An LRC line
+        # must be a single line: any newline would produce a broken line
+        # without a timestamp. Collapse all whitespace into single spaces.
+        text = re.sub(r"\s+", " ", text)
+        start = float(seg.get("start", 0.0))
+        lines.append(f"{format_timestamp(start)}{text}")
+    return "\n".join(lines) + "\n"
+
+
+def parse_lrc(content: str) -> list[LrcLine]:
+    out: list[LrcLine] = []
+    for line in content.splitlines():
+        m = _LRC_PARSE_RE.match(line.strip())
+        if not m:
+            continue
+        mm, ss, frac, text = m.groups()
+        frac_sec = float(f"0.{frac}") if frac else 0.0
+        out.append(LrcLine(start=int(mm) * 60 + int(ss) + frac_sec, text=text.strip()))
+    return out
+
+
+def is_lrc(content: str) -> bool:
+    """True if content contains at least one LRC-style timestamp line."""
+    if not content:
+        return False
+    for line in content.splitlines():
+        if _LRC_LINE_RE.match(line):
+            return True
+    return False
+
+
+# ---------------------------------------------------------------- read embedded
+
+_JUNK_LYRIC_MARKERS = (
+    "created with",
+    "russian-area.de",
+    "http://", "https://",
+    "www.",
+    "mp3tag",
+    "tag&rename",
+    "foobar2000",
+    "itunes",
+)
+
+
+def _is_junk_lyrics(text: str) -> bool:
+    """Reject tag values that are really URL/signature junk, not lyrics."""
+    if not text or len(text.strip()) < 3:
+        return True
+    low = text.lower()
+    if any(m in low for m in _JUNK_LYRIC_MARKERS):
+        return True
+    # A "lyrics" blob that is just an email or URL — reject.
+    import re
+    stripped = re.sub(r"\s+", "", low)
+    if re.fullmatch(r"[a-z0-9.\-_/]+", stripped) and (
+        "." in stripped or "/" in stripped
+    ):
+        return True
+    return False
+
+
+def read_embedded_lyrics(audio_path: Path) -> Optional[str]:
+    """Return lyrics stored in the file's tags, or None.
+
+    Order: prefer synced (LRC-like) -> plain. Never raises.
+    """
+    suffix = audio_path.suffix.lower()
+    try:
+        if suffix in (".mp3", ".mp2"):
+            text = _read_mp3(audio_path)
+        elif suffix in (".flac", ".ogg", ".opus"):
+            text = _read_vorbis(audio_path)
+        elif suffix in (".m4a", ".mp4", ".m4b", ".aac"):
+            text = _read_mp4(audio_path)
+        else:
+            return None
+    except Exception:  # noqa: BLE001
+        return None
+    if text and _is_junk_lyrics(text):
+        return None
+    return text
+
+
+def _read_mp3(path: Path) -> Optional[str]:
+    from mutagen.id3 import ID3, ID3NoHeaderError
+    try:
+        tags = ID3(str(path))
+    except ID3NoHeaderError:
+        return None
+
+    # SYLT = synced lyrics (rare); USLT = unsynced (common, may contain LRC text)
+    sylt = tags.getall("SYLT")
+    for tag in sylt:
+        if tag.text:
+            try:
+                content = "\n".join(
+                    f"{format_timestamp(ms / 1000.0)}{txt}"
+                    for txt, ms in tag.text
+                )
+                if content.strip():
+                    return content
+            except Exception:  # noqa: BLE001
+                continue
+
+    uslt = tags.getall("USLT")
+    for tag in uslt:
+        if tag.text and str(tag.text).strip():
+            return str(tag.text)
+
+    return None
+
+
+def _read_vorbis(path: Path) -> Optional[str]:
+    suffix = path.suffix.lower()
+    if suffix == ".flac":
+        from mutagen.flac import FLAC as Audio
+    elif suffix == ".ogg":
+        from mutagen.oggvorbis import OggVorbis as Audio
+    else:
+        from mutagen.oggopus import OggOpus as Audio
+    audio = Audio(str(path))
+    for key in ("LYRICS", "UNSYNCEDLYRICS", "LYRIC", "SYNCEDLYRICS"):
+        if key in audio:
+            values = audio[key]
+            if isinstance(values, list) and values:
+                text = str(values[0])
+            else:
+                text = str(values)
+            if text.strip():
+                return text
+    return None
+
+
+def _read_mp4(path: Path) -> Optional[str]:
+    from mutagen.mp4 import MP4
+    audio = MP4(str(path))
+    for key in ("\xa9lyr", "©lyr"):
+        if key in audio:
+            values = audio[key]
+            if isinstance(values, list) and values:
+                text = str(values[0])
+            else:
+                text = str(values)
+            if text.strip():
+                return text
+    return None
+
+
+# ---------------------------------------------------------------- embedding
+
+def embed_lrc(audio_path: Path, lrc_content: str) -> None:
+    suffix = audio_path.suffix.lower()
+    if suffix in (".mp3", ".mp2"):
+        _embed_mp3(audio_path, lrc_content)
+    elif suffix in (".flac", ".ogg", ".opus"):
+        _embed_vorbis(audio_path, lrc_content)
+    elif suffix in (".m4a", ".mp4", ".m4b", ".aac"):
+        _embed_mp4(audio_path, lrc_content)
+    else:
+        raise ValueError(f"Unsupported format for LRC embedding: {suffix}")
+
+
+def _embed_mp3(path: Path, lrc_content: str) -> None:
+    from mutagen.id3 import ID3, USLT, ID3NoHeaderError
+    try:
+        tags = ID3(str(path))
+    except ID3NoHeaderError:
+        tags = ID3()
+    tags.delall("USLT")
+    tags.add(USLT(encoding=3, lang="rus", desc="", text=lrc_content))
+    tags.save(str(path), v2_version=3)
+
+
+def _embed_vorbis(path: Path, lrc_content: str) -> None:
+    suffix = path.suffix.lower()
+    if suffix == ".flac":
+        from mutagen.flac import FLAC as Audio
+    elif suffix == ".ogg":
+        from mutagen.oggvorbis import OggVorbis as Audio
+    else:
+        from mutagen.oggopus import OggOpus as Audio
+    audio = Audio(str(path))
+    audio["LYRICS"] = lrc_content
+    audio["UNSYNCEDLYRICS"] = lrc_content
+    audio.save()
+
+
+def _embed_mp4(path: Path, lrc_content: str) -> None:
+    from mutagen.mp4 import MP4
+    audio = MP4(str(path))
+    audio["\xa9lyr"] = [lrc_content]
+    audio.save()

core/lrc_pipeline.py

@@ -0,0 +1,335 @@
+"""LRC-generation and LRC-embedding pipelines."""
+from __future__ import annotations
+
+import shutil
+import tempfile
+import threading
+from pathlib import Path
+from typing import Callable, Optional
+
+from .config import PipelineConfig
+from .convert import convert_to_wav
+from .lrc import embed_lrc, is_lrc, read_embedded_lyrics, segments_to_lrc
+from .metadata import read_metadata
+from .pipeline import (
+    FileResult, _artifact_belongs_to, _output_dir_for, _save_unrecognized,
+    collect_audio_files,
+)
+from .transcribe import (
+    CudaOOMError, align_text_to_audio, transcribe, transcribe_aligned,
+)
+
+
+LogFn = Callable[[str], None]
+ProgressFn = Callable[[str, int, int], None]  # stage, current, total
+OverallFn = Callable[[int, int], None]        # completed_files, total_files
+
+
+def _pick_audio_source(
+    f: Path, input_root: Path, output_dir: Path, config: PipelineConfig,
+    work_dir: Path, log: LogFn,
+) -> Path:
+    mode = (config.lrc.audio_source or "auto").lower()
+    track_out = _output_dir_for(f, input_root, output_dir)
+    vocals = track_out / "vocals.wav"
+    convert = track_out / "convert.wav"
+    if mode == "vocals":
+        return vocals
+    if mode == "convert":
+        return convert
+    if mode == "input":
+        return f
+    if vocals.exists():
+        return vocals
+    if convert.exists():
+        return convert
+    target = work_dir / "input_16k.wav"
+    log(f"  [convert] {f.name} -> 16k mono WAV (для align)")
+    convert_to_wav(
+        f, target, sample_rate=16000, channels=1,
+        ffmpeg_path=config.convert.ffmpeg_path,
+    )
+    return target
+
+
+def _ready_lrc(track_out: Path, source: Path) -> Optional[str]:
+    """Return existing valid LRC text for this exact source track, if any."""
+    if not _artifact_belongs_to(track_out, source):
+        return None
+    lrc_path = track_out / "text.lrc"
+    if not lrc_path.is_file():
+        return None
+    try:
+        content = lrc_path.read_text(encoding="utf-8")
+    except OSError:
+        return None
+    return content if content.strip() and is_lrc(content) else None
+
+
+def run_lrc_pipeline(
+    config: PipelineConfig,
+    log: LogFn = print,
+    on_progress: Optional[ProgressFn] = None,
+    stop_event: Optional[threading.Event] = None,
+    on_overall_progress: Optional[OverallFn] = None,
+) -> list[FileResult]:
+    input_dir = Path(config.input_dir)
+    if not input_dir.is_dir():
+        raise FileNotFoundError(f"Input directory not found: {input_dir}")
+    output_dir = Path(config.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    if not config.lrc.enabled:
+        log("LRC: генерация отключена — пропуск этапа")
+        return []
+
+    files = collect_audio_files(input_dir, config.recursive)
+    if not files:
+        log("Аудиофайлы не найдены.")
+        return []
+
+    # sanity: если ASR и конвертация/отделение выключены — alignment имеет смысл
+    # только для треков, у которых УЖЕ есть text.txt + vocals/convert.wav.
+    # Иначе почти всё упадёт. Предупредить заранее.
+    if not config.transcribe.enabled and not config.convert.enabled and not config.separate.enabled:
+        log("  ⚠ ASR, конвертация и отделение выключены. "
+            "LRC будет построен только для треков, где уже есть text.txt + vocals.wav "
+            "(или convert.wav). Остальные упадут с ошибкой.")
+
+    # `None` means "auto" — the language will be auto-detected during ASR
+    # (transcribe_aligned) or by the alignment model (align_text_to_audio).
+    lang = config.lrc.language
+    log(f"LRC: {len(files)} файлов   |   язык={lang or 'auto'}   |   "
+        f"источник={config.lrc.audio_source}")
+
+    results: list[FileResult] = []
+    total = len(files)
+    if on_progress:
+        on_progress("lrc", 0, total)
+    if on_overall_progress:
+        on_overall_progress(0, total)
+
+    for idx, f in enumerate(files, 1):
+        if stop_event is not None and stop_event.is_set():
+            log("⏹ Отменено")
+            break
+        log(f"\n▶ [{idx}/{total}] {f.name}")
+
+        track_out = _output_dir_for(f, input_dir, output_dir)
+        track_out.mkdir(parents=True, exist_ok=True)
+        lrc_path = track_out / "text.lrc"
+        error: Optional[str] = None
+
+        if config.lrc.skip_existing:
+            ready = _ready_lrc(track_out, f)
+            if ready is not None:
+                log("  ⊘ text.lrc уже готов — пропуск")
+                results.append(FileResult(f, None, track_out, None, "existing+lrc"))
+                if on_progress:
+                    on_progress("lrc", idx, total)
+                if on_overall_progress:
+                    on_overall_progress(idx, total)
+                continue
+
+        track_out = _output_dir_for(f, input_dir, output_dir)
+        track_out.mkdir(parents=True, exist_ok=True)
+        lrc_path = track_out / "text.lrc"
+        text_file = track_out / "text.txt"
+        error: Optional[str] = None
+
+        if config.lrc.skip_existing:
+            ready = _ready_lrc(track_out, f)
+            if ready is not None:
+                log("  ⊘ text.lrc уже готов — пропуск")
+                results.append(FileResult(f, None, track_out, None, "existing+lrc"))
+                if on_progress:
+                    on_progress("lrc", idx, total)
+                if on_overall_progress:
+                    on_overall_progress(idx, total)
+                continue
+
+        tmp = Path(tempfile.mkdtemp(prefix=f"at_lrc_{f.stem}_"))
+        try:
+            # 0) embedded LRC from tags — бесплатно
+            embedded = read_embedded_lyrics(f) if config.lyrics.use_embedded_lyrics else None
+            if embedded and is_lrc(embedded):
+                lrc_path.write_text(embedded, encoding="utf-8")
+                log("  ✓ text.lrc (взят из тегов аудио)")
+                results.append(FileResult(f, None, track_out, None, "embedded+lrc"))
+                continue
+
+            # 1) готовый текст из text.txt
+            text: Optional[str] = None
+            if text_file.is_file():
+                try:
+                    text = text_file.read_text(encoding="utf-8").strip()
+                except OSError as e:
+                    raise RuntimeError(f"не читается text.txt: {e}") from e
+            if not text:
+                text = None
+
+            # 1a) text.txt уже LRC → писать его как есть, alignment не нужен
+            if text is not None and is_lrc(text):
+                lrc_path.write_text(text, encoding="utf-8")
+                log("  ✓ text.lrc (text.txt уже содержит LRC-таймкоды)")
+                results.append(FileResult(f, None, track_out, None, "text_already_lrc"))
+                continue
+
+            # 2) аудио для alignment
+            audio_path = _pick_audio_source(
+                f, input_dir, output_dir, config, tmp, log,
+            )
+            if not audio_path.exists():
+                # Если convert/separate выключены и vocals/convert.wav отсутствуют —
+                # понятная ошибка вместо «файл не найден»
+                if not config.convert.enabled and not config.separate.enabled:
+                    raise FileNotFoundError(
+                        f"для alignment нет аудио: {audio_path}. "
+                        "Включите конвертацию/отделение в основном пайплайне, "
+                        "либо запустите их один раз, либо положите vocals.wav/convert.wav."
+                    )
+                raise FileNotFoundError(
+                    f"источник аудио не найден: {audio_path}. "
+                    f"Источник={config.lrc.audio_source}, "
+                    f"convert.enabled={config.convert.enabled}, "
+                    f"separate.enabled={config.separate.enabled}"
+                )
+            log(f"  [align] {audio_path.name}")
+
+            # 3) готовый текст → forced alignment (быстро, без ASR)
+            if text is not None:
+                log(f"  [align-only] text.txt найден ({len(text)} символов)")
+                segments, detected = align_text_to_audio(
+                    audio_path,
+                    text,
+                    language=lang,
+                    model=config.transcribe.model,
+                    device=config.lrc.device,
+                    compute_type=config.transcribe.compute_type,
+                    local_files_only=config.transcribe.local_files_only,
+                )
+            else:
+                # 3a) нет text.txt — ASR разрешён только если transcribe.enabled
+                if not config.transcribe.enabled:
+                    raise FileNotFoundError(
+                        "text.txt отсутствует, а транскрибация выключена. "
+                        "Включите транскрибацию в основном пайплайне или "
+                        "положите text.txt в папку трека вручную."
+                    )
+                log("  [asr+align] text.txt нет — выполняю ASR заново")
+                segments, detected = transcribe_aligned(
+                    audio_path,
+                    model=config.transcribe.model,
+                    language=lang,
+                    device=config.lrc.device,
+                    compute_type=config.transcribe.compute_type,
+                    batch_size=config.transcribe.batch_size,
+                    align_model=config.lrc.align_model,
+                    local_files_only=config.transcribe.local_files_only,
+                )
+
+            lrc_content = segments_to_lrc(segments)
+            if not lrc_content.strip():
+                log(f"  ⚠ {f.name}: пустой LRC (нет сегментов)")
+                results.append(FileResult(
+                    f, None, None,
+                    "EmptyResult: alignment returned no text", "lrc+empty",
+                ))
+                continue
+            lrc_path.write_text(lrc_content, encoding="utf-8")
+            log(f"  ✓ text.lrc ({len(segments)} строк, язык: {detected})")
+
+            _save_unrecognized(
+                config, f, lrc_content, source="lrc",
+                metadata=read_metadata(f),
+            )
+            results.append(FileResult(f, None, track_out, None, "lrc"))
+        except CudaOOMError as e:
+            error = f"CudaOOMError: {e}"
+            log(f"  ⊘ {f.name}: OOM — пропущен, идём дальше")
+            results.append(FileResult(f, None, None, error, "lrc+oom"))
+        except Exception as e:  # noqa: BLE001
+            error = f"{type(e).__name__}: {e}"
+            log(f"  ✗ Ошибка: {error}")
+            results.append(FileResult(f, None, None, error, "lrc"))
+        finally:
+            shutil.rmtree(tmp, ignore_errors=True)
+            if on_progress:
+                on_progress("lrc", idx, total)
+            if on_overall_progress:
+                on_overall_progress(idx, total)
+
+    return results
+
+
+def run_embed_pipeline(
+    config: PipelineConfig,
+    log: LogFn = print,
+    on_progress: Optional[ProgressFn] = None,
+    stop_event: Optional[threading.Event] = None,
+    on_overall_progress: Optional[OverallFn] = None,
+) -> list[FileResult]:
+    input_dir = Path(config.input_dir)
+    if not input_dir.is_dir():
+        raise FileNotFoundError(f"Input directory not found: {input_dir}")
+    output_dir = Path(config.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    if not config.lrc.embed_enabled:
+        log("LRC embed: встраивание отключено — пропуск этапа")
+        return []
+
+    files = collect_audio_files(input_dir, config.recursive)
+    if not files:
+        log("Аудиофайлы не найдены.")
+        return []
+
+    log(f"Embed LRC: {len(files)} файлов   |   "
+        f"режим={'in-place' if config.lrc.embed_in_place else 'копия в output'}")
+
+    results: list[FileResult] = []
+    total = len(files)
+    if on_progress:
+        on_progress("embed", 0, total)
+    if on_overall_progress:
+        on_overall_progress(0, total)
+
+    for idx, f in enumerate(files, 1):
+        if stop_event is not None and stop_event.is_set():
+            log("⏹ Отменено")
+            break
+
+        track_out = _output_dir_for(f, input_dir, output_dir)
+        lrc_path = track_out / "text.lrc"
+        error: Optional[str] = None
+
+        try:
+            if not lrc_path.exists():
+                log(f"  ⊘ {f.name}: text.lrc отсутствует, пропуск")
+                results.append(FileResult(f, None, None, "no LRC", "embed"))
+                continue
+
+            lrc_content = lrc_path.read_text(encoding="utf-8")
+
+            if config.lrc.embed_in_place:
+                target = f
+            else:
+                target = track_out / f.name
+                target.parent.mkdir(parents=True, exist_ok=True)
+                if not target.exists() or config.overwrite:
+                    shutil.copy2(f, target)
+
+            embed_lrc(target, lrc_content)
+            log(f"  ✓ {f.name} → {target if config.lrc.embed_in_place else target.name}")
+            results.append(FileResult(f, None, target.parent, None, "embed"))
+        except Exception as e:  # noqa: BLE001
+            error = f"{type(e).__name__}: {e}"
+            log(f"  ✗ {f.name}: {error}")
+            results.append(FileResult(f, None, None, error, "embed"))
+        finally:
+            if on_progress:
+                on_progress("embed", idx, total)
+            if on_overall_progress:
+                on_overall_progress(idx, total)
+
+    return results

core/separate.py

@@ -3,7 +3,22 @@ from __future__ import annotations
 
 import os
 import subprocess
+import time
 from pathlib import Path
+from typing import Callable, Optional
+
+
+def _build_env(custom_kernels: bool) -> dict:
+    env = os.environ.copy()
+    env["UNBLEND_CUSTOM_KERNELS"] = "1" if custom_kernels else "0"
+    return env
+
+
+def _check_result(result: subprocess.CompletedProcess, what: str) -> None:
+    if result.returncode != 0:
+        raise RuntimeError(
+            f"unblend failed ({what}):\n{result.stderr.strip()[:1200]}"
+        )
 
 
 def separate_vocals(
@@ -13,8 +28,8 @@ def separate_vocals(
     isolate_stem: str = "vocals",
     unblend_path: str = "unblend",
     custom_kernels: bool = False,
+    device: str = "cuda",
 ) -> Path:
-    """Separate the isolated stem from input_wav. Returns path to the stem WAV."""
     work_dir.mkdir(parents=True, exist_ok=True)
     output_pattern = work_dir / "{track}" / "{stem}.wav"
     cmd = [
@@ -22,15 +37,10 @@ def separate_vocals(
         "-m", model,
         "-o", str(output_pattern),
         "--isolate-stem", isolate_stem,
+        "--device", device,
     ]
-    env = os.environ.copy()
-    env["UNBLEND_CUSTOM_KERNELS"] = "1" if custom_kernels else "0"
-
-    result = subprocess.run(cmd, capture_output=True, text=True, env=env)
-    if result.returncode != 0:
-        raise RuntimeError(
-            f"unblend failed for {input_wav.name}:\n{result.stderr.strip()[:800]}"
-        )
+    result = subprocess.run(cmd, capture_output=True, text=True, env=_build_env(custom_kernels))
+    _check_result(result, input_wav.name)
 
     candidates = sorted(work_dir.rglob(f"{isolate_stem}.wav"))
     if not candidates:
@@ -38,3 +48,78 @@ def separate_vocals(
             f"unblend finished but stem '{isolate_stem}.wav' was not found in {work_dir}"
         )
     return candidates[0]
+
+
+def separate_vocals_batch(
+    input_wavs: list[Path],
+    work_root: Path,
+    model: str = "bs_roformer_anvuew",
+    isolate_stem: str = "vocals",
+    unblend_path: str = "unblend",
+    custom_kernels: bool = False,
+    device: str = "cuda",
+    progress_callback: Optional[Callable[[int], None]] = None,
+) -> dict[Path, Path]:
+    """Separate many files in one unblend invocation.
+
+    unblend doesn't emit per-file progress, so progress_callback is called
+    with an incrementing counter roughly every 2 seconds, up to N-1.
+    The final N is always reported after the process exits.
+    """
+    if not input_wavs:
+        return {}
+
+    work_root.mkdir(parents=True, exist_ok=True)
+    output_pattern = work_root / "{parent}" / "{stem}.wav"
+
+    cmd = [
+        unblend_path, "separate",
+        *[str(p) for p in input_wavs],
+        "-m", model,
+        "-o", str(output_pattern),
+        "--isolate-stem", isolate_stem,
+        "--device", device,
+    ]
+
+    proc = subprocess.Popen(
+        cmd,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        env=_build_env(custom_kernels),
+    )
+
+    total = len(input_wavs)
+    processed = 0
+    while proc.poll() is None:
+        time.sleep(2.0)
+        if progress_callback and processed < total - 1:
+            processed += 1
+            progress_callback(processed)
+
+    if proc.returncode != 0:
+        raise RuntimeError(f"unblend batch failed with exit code {proc.returncode}")
+
+    if progress_callback:
+        progress_callback(total)
+
+    mapping: dict[Path, Path] = {}
+    missing: list[Path] = []
+    for src in input_wavs:
+        candidate = src.parent / f"{isolate_stem}.wav"
+        if candidate.exists():
+            mapping[src] = candidate
+        else:
+            missing.append(src)
+
+    if missing:
+        found = sorted(work_root.rglob(f"{isolate_stem}.wav"))
+        if len(found) == len(input_wavs):
+            for src, dst in zip(input_wavs, found):
+                mapping[src] = dst
+        else:
+            names = ", ".join(p.name for p in missing[:3])
+            raise RuntimeError(
+                f"unblend finished but outputs missing for {len(missing)} file(s): {names}"
+            )
+
+    return mapping

core/transcribe.py

@@ -1,16 +1,152 @@
 """Speech recognition via WhisperX or GigaAM (onnx-asr).
 
-Models are cached per (name, device, compute_type) so loading happens once
-per process, not once per file. Loading is retried a few times to survive
-transient HuggingFace / SSL hiccups.
+Also exposes `transcribe_aligned()` for word-level alignment (used by LRC).
+Models are cached; aggressive VRAM cleanup happens after each file.
 """
 from __future__ import annotations
 
+import gc
+import threading
 import time
 from pathlib import Path
 from typing import Optional
 
 
+# The model caches below are process-wide. They are intentionally never
+# cleared between files or batches: reloading WhisperX takes minutes and is
+# what made runs look like they reloaded the model every batch. The lock makes
+# the "load once, reuse forever" guarantee safe even if ASR is ever called
+# from multiple threads.
+_MODEL_LOCK = threading.RLock()
+_REUSE_LOGGED: set = set()
+
+# CUDA context failures that cannot be recovered by retrying or empty_cache().
+# They require restarting the app (or switching to CPU).
+_DEAD_CUDA_MARKERS = (
+    "invalid device ordinal",
+    "cudaerrorinvaliddevice",
+    "device-side assert",
+    "unspecified launch failure",
+)
+
+
+def _is_dead_cuda(exc: BaseException) -> bool:
+    err = str(exc).lower()
+    return any(m in err for m in _DEAD_CUDA_MARKERS)
+
+
+    _WhisperXModels._cache.clear()
+    _AlignModels._cache.clear()
+    _GigaAMModels._cache.clear()
+    _StableTSModels._cache.clear()
+    _REUSE_LOGGED.clear()
+
+
+# Phrases that are strong signals of ASR hallucination (end credits)
+_HALLUCINATION_PHRASES = (
+    "продолжение следует",
+    "спасибо за внимание",
+    "спасибо за просмотр",
+    "спасибо за подписку",
+    "подписывайтесь на канал",
+    "ставьте лайки",
+    "не забудьте подписаться",
+    "субтитры создавал",
+    "субтитры сделал",
+    "субтитры подготовил",
+    "редактор субтитров",
+    "корректор",
+    "thank you for watching",
+    "thanks for watching",
+    "please subscribe",
+    "subscribe to my channel",
+    "like and subscribe",
+)
+
+
+def _is_hallucination(text: str) -> bool:
+    """Return True if the text looks like an ASR hallucination.
+
+    Checks:
+      - known call-to-action / promo phrases;
+      - repeated-word tic ("Музыка Музыка").
+    """
+    if not text:
+        return False
+    t = " ".join(text.lower().split()).strip(" .!?…,;:\"'«»")
+    for phrase in _HALLUCINATION_PHRASES:
+        if phrase in t:
+            return True
+    words = t.split()
+    if len(words) == 2 and words[0] == words[1]:
+        return True
+    return False
+
+
+def _detect_language(model_obj, audio_or_path, fallback: str = "ru") -> str:
+    """Detect the spoken language from audio using the ASR model itself.
+
+    Returns a lowercase ISO code or `fallback` on failure.
+    """
+    try:
+        lang, _prob = model_obj.detect_language(audio_or_path)
+        if lang:
+            return lang.lower()
+    except Exception:
+        pass
+    return fallback
+
+
+class _StableTSModels:
+    _cache: dict = {}
+
+    @classmethod
+    def get(cls, model_name: str, device: str, compute_type: Optional[str],
+            local_files_only: bool = False) -> Any:
+        if compute_type is None:
+            compute_type = "int8_float16" if device == "cuda" else "int8"
+        key = (model_name, device, compute_type, local_files_only)
+        with _MODEL_LOCK:
+            if key not in cls._cache:
+                import stable_whisper
+
+                print(f"  [ALIGN] loading stable-ts {model_name} on {device} "
+                      f"({compute_type}, local_files_only={local_files_only})...")
+
+                def load() -> 'faster_whisper.WhisperModel':
+                    return stable_whisper.load_faster_whisper(
+                        model_name, device=device, compute_type=compute_type,
+                        local_files_only=local_files_only,
+                    )
+
+                cls._cache[key] = _retry(load)
+                print("  [ALIGN] stable-ts model loaded once — reuse.")
+            return cls._cache[key]
+
+
+def _clear_model_caches() -> None:
+    _WhisperXModels._cache.clear()
+    _AlignModels._cache.clear()
+    _GigaAMModels._cache.clear()
+    _StableTSModels._cache.clear()
+    _REUSE_LOGGED.clear()
+
+
+class CudaOOMError(RuntimeError):
+    """CUDA OOM, переживший все стратегии понижения нагрузки."""
+
+
+def _is_cuda_error(exc: BaseException) -> bool:
+    err = str(exc).lower()
+    return (
+        "out of memory" in err
+        or "cuda" in err
+        or "device ordinal" in err
+        or "cublas" in err
+        or "cudnn" in err
+    )
+
+
 def _resolve_device(device: str) -> str:
     if device in ("cuda", "cpu"):
         return device
@@ -26,13 +162,76 @@ def _retry(fn, attempts: int = 3, base_delay: float = 2.0):
     for i in range(attempts):
         try:
             return fn()
-        except Exception as e:  # noqa: BLE001
+        except Exception as e:
             last = e
             if i < attempts - 1:
                 time.sleep(base_delay * (i + 1))
     raise last
 
 
+def _free_vram() -> None:
+    gc.collect()
+    try:
+        import torch
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+    except Exception:  # noqa: BLE001 — cleanup must never mask the original error
+        pass
+
+
+def _enable_tf32() -> None:
+    """Enable TF32 matmul on Ampere+ GPUs (RTX 30/40).
+
+    Called lazily right before a WhisperX model is loaded, so runs that do not
+    use ASR never import Torch or create a CUDA context at startup.
+    """
+    try:
+        import torch
+        if torch.cuda.is_available():
+            torch.backends.cuda.matmul.allow_tf32 = True
+            torch.backends.cudnn.allow_tf32 = True
+    except Exception:  # noqa: BLE001 — purely an optimization
+        pass
+
+
+def _model_transcribe_with_retry(
+    model_obj, audio, language, batch_size, name="<audio>",
+):
+    """Try progressively lighter settings so one long file can still finish.
+
+    The loaded model is intentionally kept in the cache between attempts:
+    reloading it costs minutes and does not help with an OOM caused by a long
+    audio file. VAD is not touched here — WhisperX expects it at model-load
+    time (``vad_model``/``vad_options``), not per transcription call. A dead
+    CUDA context is never retried: caches are dropped and CudaOOMError is
+    raised immediately. If even batch=1 fails with a normal OOM, raise
+    CudaOOMError so the caller can skip this file and continue with the rest
+    of the batch.
+    """
+    attempts = [batch_size, 1]
+    last: Optional[Exception] = None
+    for i, bs in enumerate(attempts, 1):
+        try:
+            return model_obj.transcribe(
+                audio, batch_size=bs, language=language,
+            )
+        except Exception as e:  # noqa: BLE001
+            if _is_dead_cuda(e):
+                _clear_model_caches()
+                raise CudaOOMError(
+                    f"CUDA context died on {name}: {e}. "
+                    "GPU state is unrecoverable — restart the app or switch to CPU."
+                ) from e
+            if not _is_cuda_error(e):
+                raise
+            last = e
+            print(f"  [ASR] OOM on {name} (attempt {i}/{len(attempts)}, bs={bs}): {e}")
+            _free_vram()
+    raise CudaOOMError(
+        f"OOM after {len(attempts)} attempts on {name}: {last}"
+    ) from last
+
+
 class _WhisperXModels:
     _cache: dict = {}
 
@@ -42,7 +241,10 @@ class _WhisperXModels:
         if compute_type is None:
             compute_type = "int8_float16" if device == "cuda" else "int8"
         key = (model_name, device, compute_type, local_files_only)
+        with _MODEL_LOCK:
             if key not in cls._cache:
+                if device == "cuda":
+                    _enable_tf32()
                 import whisperx
                 print(f"  [ASR] loading WhisperX {model_name} on {device} "
                       f"({compute_type}, local_files_only={local_files_only})...")
@@ -55,7 +257,36 @@ class _WhisperXModels:
                     )
 
                 cls._cache[key] = _retry(load)
-            print("  [ASR] model loaded.")
+                print("  [ASR] model loaded once — subsequent files will reuse it.")
+            elif ("whisperx",) + key not in _REUSE_LOGGED:
+                _REUSE_LOGGED.add(("whisperx",) + key)
+                print(f"  [ASR] reusing loaded WhisperX {model_name} ({device}).")
+            return cls._cache[key]
+
+
+class _AlignModels:
+    _cache: dict = {}
+
+    @classmethod
+    def get(cls, language_code: str, device: str, model_name: Optional[str] = None):
+        key = (language_code, device, model_name or "")
+        with _MODEL_LOCK:
+            if key not in cls._cache:
+                import whisperx
+                print(f"  [ALIGN] loading align model for {language_code} "
+                      f"({model_name or 'default'}) on {device}...")
+
+                def load():
+                    kwargs = {"language_code": language_code, "device": device}
+                    if model_name:
+                        kwargs["model_name"] = model_name
+                    return whisperx.load_align_model(**kwargs)
+
+                cls._cache[key] = _retry(load)
+                print("  [ALIGN] align model loaded once — reusing it.")
+            elif ("align",) + key not in _REUSE_LOGGED:
+                _REUSE_LOGGED.add(("align",) + key)
+                print(f"  [ALIGN] reusing loaded align model for {language_code}.")
             return cls._cache[key]
 
 
@@ -70,6 +301,7 @@ class _GigaAMModels:
             "auto": ["CUDAExecutionProvider", "CPUExecutionProvider"],
         }.get(provider, ["CPUExecutionProvider"])
         key = tuple(providers)
+        with _MODEL_LOCK:
             if key not in cls._cache:
                 import onnx_asr
 
@@ -80,6 +312,28 @@ class _GigaAMModels:
             return cls._cache[key]
 
 
+def preload_transcription_model(
+    engine: str = "whisperx",
+    model: str = "large-v3",
+    device: str = "auto",
+    compute_type: Optional[str] = None,
+    gigaam_provider: str = "auto",
+    local_files_only: bool = False,
+) -> None:
+    """Load the ASR model once so the per-file loop only reuses it.
+
+    All model caches are process-wide and are never cleared between files,
+    batches or pipeline runs. This helper just makes the first load explicit.
+    """
+    device = _resolve_device(device)
+    if engine == "whisperx":
+        _WhisperXModels.get(model, device, compute_type, local_files_only)
+    elif engine == "gigaam":
+        _GigaAMModels.get(gigaam_provider)
+    else:
+        raise ValueError(f"Unknown ASR engine: {engine!r}")
+
+
 def transcribe(
     audio_path: Path,
     engine: str = "whisperx",
@@ -93,11 +347,13 @@ def transcribe(
     vad_filter: bool = True,
 ) -> str:
     device = _resolve_device(device)
+
     if engine == "whisperx":
         return _transcribe_whisperx(
             audio_path, model, language, device, compute_type,
             batch_size, local_files_only, vad_filter,
         )
+
     if engine == "gigaam":
         return _transcribe_gigaam(audio_path, gigaam_provider)
     raise ValueError(f"Unknown ASR engine: {engine!r}")
@@ -116,14 +372,112 @@ def _transcribe_whisperx(
     import whisperx
     model = _WhisperXModels.get(model_name, device, compute_type, local_files_only)
     audio = whisperx.load_audio(str(audio_path))
-    result = model.transcribe(
-        audio,
-        batch_size=batch_size,
-        language=language,
+    result = _model_transcribe_with_retry(
+        model, audio, language, batch_size, name=audio_path.name,
     )
-    return " ".join(seg["text"].strip() for seg in result["segments"]).strip()
+
+    # Filter out hallucinated / spammy segments.
+    segments = result.get("segments") or []
+    kept = []
+    for seg in segments:
+        t = (seg.get("text") or "").strip()
+        if not t or _is_hallucination(t):
+            continue
+        if seg.get("avg_logprob", 0.0) < -1.2:
+            continue
+        kept.append(seg)
+    text = " ".join(s["text"].strip() for s in kept).strip()
+    if _is_hallucination(text):
+        text = ""
+
+    del audio, result
+    _free_vram()
+    return text
 
 
 def _transcribe_gigaam(audio_path: Path, provider: str) -> str:
     model = _GigaAMModels.get(provider)
     return model.recognize(str(audio_path), channel="mean").strip()
+
+
+def transcribe_aligned(
+    audio_path: Path,
+    model: str = "large-v3",
+    language: Optional[str] = "ru",
+    device: str = "auto",
+    compute_type: Optional[str] = None,
+    batch_size: int = 8,
+    align_model: Optional[str] = None,
+    local_files_only: bool = False,
+    vad_filter: bool = True,
+) -> tuple[list[dict], str]:
+    """Transcribe + word-level align. Returns (aligned_segments, detected_language).
+
+    Each segment has 'start', 'end', 'text', and optionally 'words'.
+    Requires whisperx engine.
+    """
+    import whisperx
+    device = _resolve_device(device)
+    model_obj = _WhisperXModels.get(model, device, compute_type, local_files_only)
+    audio = whisperx.load_audio(str(audio_path))
+    result = _model_transcribe_with_retry(
+        model_obj, audio, language, batch_size, name=audio_path.name,
+    )
+
+    detected = result.get("language") or language or "en"
+
+    try:
+        align_model_obj, align_metadata = _AlignModels.get(detected, device, align_model)
+        aligned = whisperx.align(
+            result["segments"], align_model_obj, align_metadata,
+            audio, device, return_char_alignments=False,
+        )
+        segments = aligned["segments"]
+    except Exception as e:
+        print(f"  [ALIGN] failed: {e}. Falling back to raw WhisperX segments.")
+        segments = result["segments"]
+
+    # Filter out hallucinated segments.
+    segments = [s for s in segments if not _is_hallucination(s.get("text", ""))]
+
+    del audio, result
+    _free_vram()
+    return segments, detected
+
+
+def align_text_to_audio(
+    audio_path: Path,
+    text: str,
+    language: Optional[str] = "ru",
+    model: str = "large-v3",
+    device: str = "auto",
+    compute_type: Optional[str] = None,
+    local_files_only: bool = False,
+) -> tuple[list[dict], str]:
+    """Forced alignment of EXISTING text to audio. No ASR.
+
+    When `language` is None, the language is auto-detected from the audio
+    using the alignment model itself (which is a faster-whisper model and
+    exposes `detect_language`).
+    """
+    import stable_whisper as _sw
+    device = _resolve_device(device)
+    model_obj = _StableTSModels.get(model, device, compute_type, local_files_only)
+
+    if language:
+        lang = language
+    else:
+        lang = _detect_language(model_obj, str(audio_path), fallback="ru")
+
+    try:
+        result = model_obj.align(str(audio_path), text, language=lang)
+    except Exception as e:  # noqa: BLE001
+        raise RuntimeError(f"align failed on {audio_path.name}: {e}") from e
+
+    segments = [
+        {"start": float(seg.start), "end": float(seg.end),
+         "text": (seg.text or "").strip()}
+        for seg in result.segments
+        if (seg.text or "").strip()
+    ]
+    return segments, lang

gui/main_window.py

@@ -1,21 +1,22 @@
-"""Main application window — dark theme, mode selector, staged/sequential pipelines."""
+"""Main application window — dark theme, mode selector, LRC tab."""
 from __future__ import annotations
 
 import shutil
+import subprocess
 from pathlib import Path
 from typing import Optional
 
 from PySide6.QtCore import Qt, QUrl
-from PySide6.QtGui import QColor, QDesktopServices
+from PySide6.QtGui import QColor, QDesktopServices, QKeySequence, QShortcut
 from PySide6.QtWidgets import (
-    QCheckBox, QComboBox, QDoubleSpinBox, QFileDialog, QFormLayout, QGroupBox,
-    QHBoxLayout, QLabel, QLineEdit, QMainWindow, QMessageBox,
-    QPlainTextEdit, QProgressBar, QPushButton, QSpinBox, QTabWidget,
-    QVBoxLayout, QWidget,
+    QCheckBox, QComboBox, QDoubleSpinBox, QFileDialog, QFormLayout, QFrame,
+    QGroupBox, QHBoxLayout, QLabel, QLineEdit, QMainWindow, QMessageBox,
+    QPlainTextEdit, QProgressBar, QPushButton, QScrollArea, QSpinBox,
+    QTabWidget, QVBoxLayout, QWidget,
 )
 
 from core.config import (
-    ConvertConfig, LyricsConfig, PipelineConfig, SeparateConfig,
+    ConvertConfig, LrcConfig, LyricsConfig, PipelineConfig, SeparateConfig,
     ShazamConfig, TranscribeConfig,
 )
 from gui.worker import PipelineWorker
@@ -31,6 +32,8 @@ STAGE_LABELS = {
     "convert": "Конвертация",
     "separate": "Отделение",
     "transcribe": "Транскрибация",
+    "lrc": "LRC",
+    "embed": "Встраивание LRC",
 }
 
 MODE_LABELS = {
@@ -40,19 +43,30 @@ MODE_LABELS = {
 
 
 def _detect_device_info() -> tuple[str, str]:
-    try:
-        import torch
-    except ImportError:
-        return "⚠ PyTorch не установлен", "#f38ba8"
+    """Report the GPU via nvidia-smi.
 
-    if torch.cuda.is_available():
+    Torch must not be imported here: importing it initializes a CUDA context
+    and reserves VRAM even for runs that only need separation. The ASR stage
+    imports Torch lazily when it actually loads a model.
+    """
     try:
-            name = torch.cuda.get_device_name(0)
-            vram = torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)
+        out = subprocess.run(
+            ["nvidia-smi", "--query-gpu=name,memory.total",
+             "--format=csv,noheader,nounits"],
+            capture_output=True, text=True, timeout=5, check=True,
+        ).stdout.strip()
+    except (OSError, subprocess.SubprocessError):
+        out = ""
+    if out:
+        parts = [p.strip() for p in out.splitlines()[0].split(",")]
+        name = parts[0] or "CUDA GPU"
+        if len(parts) > 1:
+            try:
+                vram = float(parts[1]) / 1024  # MiB -> GiB
                 return f"✓ CUDA: {name} · {vram:.1f} GB VRAM", "#a6e3a1"
-        except Exception:  # noqa: BLE001
-            return "✓ CUDA доступна", "#a6e3a1"
-
+            except ValueError:
+                pass
+        return f"✓ CUDA: {name}", "#a6e3a1"
     return "⚠ CUDA не обнаружена — расчёт пойдёт на CPU (медленно)", "#f9e2af"
 
 
@@ -60,7 +74,7 @@ class MainWindow(QMainWindow):
     def __init__(self) -> None:
         super().__init__()
         self.setWindowTitle("Song Transcriber")
-        self.resize(1140, 880)
+        self.resize(1140, 920)
         self._worker: Optional[PipelineWorker] = None
         self._build_ui()
         self._apply_defaults()
@@ -71,8 +85,17 @@ class MainWindow(QMainWindow):
     # ---------------- UI ----------------
 
     def _build_ui(self) -> None:
+        # The whole page lives in a scroll area: on small screens the user can
+        # scroll down to the log instead of the console being cut off by the
+        # bottom edge of the window.
+        scroll = QScrollArea()
+        scroll.setWidgetResizable(True)
+        scroll.setFrameShape(QFrame.Shape.NoFrame)
+        scroll.setHorizontalScrollBarPolicy(Qt.ScrollBarPolicy.ScrollBarAlwaysOff)
+        self.setCentralWidget(scroll)
+
         central = QWidget()
-        self.setCentralWidget(central)
+        scroll.setWidget(central)
         root = QVBoxLayout(central)
         root.setContentsMargins(18, 16, 18, 16)
         root.setSpacing(12)
@@ -82,6 +105,10 @@ class MainWindow(QMainWindow):
         title.setObjectName("appTitle")
         header.addWidget(title)
         header.addStretch(1)
+        self.fullscreen_btn = QPushButton("⛶  На весь экран")
+        self.fullscreen_btn.setToolTip("Развернуть окно на весь экран (F11)")
+        self.fullscreen_btn.clicked.connect(self._toggle_fullscreen)
+        header.addWidget(self.fullscreen_btn)
         self.device_label = QLabel("")
         self.device_label.setObjectName("deviceInfo")
         header.addWidget(self.device_label)
@@ -93,7 +120,6 @@ class MainWindow(QMainWindow):
 
         self.input_edit = QLineEdit()
         top_layout.addRow("Входная папка:", self._with_browse(self.input_edit, self._pick_input_dir))
-
         self.output_edit = QLineEdit()
         top_layout.addRow("Выходная папка:", self._with_browse(self.output_edit, self._pick_output_dir))
 
@@ -102,6 +128,16 @@ class MainWindow(QMainWindow):
             self.mode_cb.addItem(label, key)
         top_layout.addRow("Режим:", self.mode_cb)
 
+        self.chunk_size_spin = QSpinBox()
+        self.chunk_size_spin.setRange(1, 10000)
+        self.chunk_size_spin.setToolTip(
+            "Размер батча в staged-режиме.\n"
+            "После каждого батча: удаляются временные файлы, освобождается VRAM.\n"
+            "Модели WhisperX остаются в памяти — перезагрузки не будет.\n"
+            "64 — разумный компромисс для 32 ГБ RAM и 8 ГБ VRAM."
+        )
+        top_layout.addRow("Размер батча:", self.chunk_size_spin)
+
         opts = QHBoxLayout()
         self.recursive_cb = QCheckBox("Рекурсивный обход")
         self.keep_intermediate_cb = QCheckBox("Сохранять промежуточные")
@@ -112,6 +148,22 @@ class MainWindow(QMainWindow):
         opts.addStretch(1)
         top_layout.addRow("Опции:", opts)
 
+        resume_opts = QHBoxLayout()
+        self.skip_existing_cb = QCheckBox("Пропускать готовые (text.txt)")
+        self.skip_existing_cb.setToolTip(
+            "Не трогать файлы, у которых в <output>/<relpath>/<stem>/ уже есть text.txt.\n"
+            "Позволяет продолжить большой прогон после сбоя без пересчёта.\n"
+            "Если в metadata.json записан path другого файла — папка не считается готовой."
+        )
+        self.resume_partial_cb = QCheckBox("Подхватывать vocals.wav / convert.wav")
+        self.resume_partial_cb.setToolTip(
+            "Если text.txt ещё нет, но есть vocals.wav — сразу переходить к\n"
+            "транскрибации. Если есть только convert.wav — пропустить конвертацию."
+        )
+        resume_opts.addWidget(self.skip_existing_cb)
+        resume_opts.addWidget(self.resume_partial_cb)
+        resume_opts.addStretch(1)
+        top_layout.addRow("Resume:", resume_opts)
         root.addWidget(top_box)
 
         self.tabs = QTabWidget()
@@ -120,6 +172,7 @@ class MainWindow(QMainWindow):
         self.tabs.addTab(self._build_convert_tab(), "Конвертация")
         self.tabs.addTab(self._build_separate_tab(), "Отделение вокала")
         self.tabs.addTab(self._build_transcribe_tab(), "Транскрибация")
+        self.tabs.addTab(self._build_lrc_tab(), "LRC")
         root.addWidget(self.tabs)
 
         self.shazam_enabled_cb.toggled.connect(self._sync_tab_indicators)
@@ -127,13 +180,14 @@ class MainWindow(QMainWindow):
         self.convert_enabled_cb.toggled.connect(self._sync_tab_indicators)
         self.separate_enabled_cb.toggled.connect(self._sync_tab_indicators)
         self.transcribe_enabled_cb.toggled.connect(self._sync_tab_indicators)
+        self.lrc_enabled_cb.toggled.connect(self._sync_tab_indicators)
         self.engine_cb.currentTextChanged.connect(self._sync_engine_fields)
 
         row = QHBoxLayout()
         row.setSpacing(10)
-        self.start_btn = QPushButton("▶  Запустить")
+        self.start_btn = QPushButton("▶  Запустить пайплайн")
         self.start_btn.setObjectName("primary")
-        self.start_btn.clicked.connect(self._on_start)
+        self.start_btn.clicked.connect(lambda: self._run_task("main"))
         self.stop_btn = QPushButton("⏹  Стоп")
         self.stop_btn.setObjectName("danger")
         self.stop_btn.setEnabled(False)
@@ -157,12 +211,35 @@ class MainWindow(QMainWindow):
         self.progress.setFormat("Ожидание")
         root.addWidget(self.progress)
 
+        self.overall_progress = QProgressBar()
+        self.overall_progress.setObjectName("overallProgress")
+        self.overall_progress.setFormat("Общий прогресс: ожидание")
+        root.addWidget(self.overall_progress)
+
         self.log_view = QPlainTextEdit()
         self.log_view.setObjectName("log")
         self.log_view.setReadOnly(True)
+        # Keep at least ~12 log lines visible regardless of the UI font size.
+        line_height = self.log_view.fontMetrics().lineSpacing()
+        self.log_view.setMinimumHeight(max(280, line_height * 12 + 24))
         self.log_view.setPlaceholderText("Здесь появится лог выполнения…")
         root.addWidget(self.log_view, 1)
 
+        self._fs_shortcut = QShortcut(QKeySequence("F11"), self)
+        self._fs_shortcut.activated.connect(self._toggle_fullscreen)
+        self._esc_shortcut = QShortcut(QKeySequence("Escape"), self)
+        self._esc_shortcut.activated.connect(self._leave_fullscreen)
+
+    def _toggle_fullscreen(self) -> None:
+        if self.isFullScreen():
+            self.showNormal()
+        else:
+            self.showFullScreen()
+
+    def _leave_fullscreen(self) -> None:
+        if self.isFullScreen():
+            self.showNormal()
+
     def _with_browse(self, edit: QLineEdit, picker) -> QWidget:
         w = QWidget()
         h = QHBoxLayout(w)
@@ -180,19 +257,18 @@ class MainWindow(QMainWindow):
         form = QFormLayout(w)
         form.setContentsMargins(16, 16, 16, 16)
         form.setSpacing(10)
-
         self.shazam_enabled_cb = QCheckBox("Включён")
+        self.shazam_concurrency = QSpinBox()
+        self.shazam_concurrency.setRange(1, 16)
+        form.addRow("Параллельных запросов:", self.shazam_concurrency)
         form.addRow("", self.shazam_enabled_cb)
-
         info = QLabel(
             "Идентификация трека через Shazam (unofficial API).\n"
-            "При совпадении сохраняются метаданные и (если возможно) текст из LRCLIB.\n"
-            "Требует ffmpeg (для конвертации в 16 кГц моно)."
+            "При совпадении сохраняются метаданные и (если возможно) текст из LRCLIB."
         )
         info.setWordWrap(True)
         info.setStyleSheet("color: #a6adc8;")
         form.addRow(info)
-
         return w
 
     def _build_lrclib_tab(self) -> QWidget:
@@ -200,37 +276,56 @@ class MainWindow(QMainWindow):
         form = QFormLayout(w)
         form.setContentsMargins(16, 16, 16, 16)
         form.setSpacing(10)
-
         self.lyrics_enabled_cb = QCheckBox("Включён")
         form.addRow("", self.lyrics_enabled_cb)
-
+        self.lyrics_use_embedded_cb = QCheckBox("Использовать встроенный текст из тегов (USLT/LYRICS/©lyr)")
+        self.lyrics_use_embedded_cb.setToolTip(
+            "Сначала проверяем, есть ли текст (LRC или обычный) уже в тегах аудио.\n"
+            "Если есть — ASR/Shazam/LRCLIB не запускаются."
+        )
+        form.addRow("", self.lyrics_use_embedded_cb)
         self.lyrics_user_agent = QLineEdit()
         form.addRow("User-Agent:", self.lyrics_user_agent)
-
-        self.lyrics_use_tags_cb = QCheckBox("Искать по тегам (artist + title)")
+        self.lyrics_use_tags_cb = QCheckBox("Искать по тегам")
+        self.lyrics_use_tags_cb.setToolTip(
+            "Использовать artist/title из тегов файла для поиска на LRCLIB.\n"
+            "Срабатывает как fallback, если Shazam ничего не нашёл\n"
+            "или его результат отключён."
+        )
         form.addRow("", self.lyrics_use_tags_cb)
-
-        self.lyrics_use_shazam_cb = QCheckBox("Искать по результату Shazam")
+        self.lyrics_use_shazam_cb = QCheckBox("Искать по Shazam")
+        self.lyrics_use_shazam_cb.setToolTip(
+            "Приоритетный источник для поиска на LRCLIB: artist/title из Shazam.\n"
+            "Теги файла используются только если Shazam не дал результата."
+        )
         form.addRow("", self.lyrics_use_shazam_cb)
-
-        self.lyrics_after_asr_cb = QCheckBox("Проверять после ASR (fuzzy-match)")
+        self.lyrics_after_asr_cb = QCheckBox("Проверять после ASR (fuzzy)")
         form.addRow("", self.lyrics_after_asr_cb)
-
         self.lyrics_transcribe_unresolved_cb = QCheckBox(
-            "Запускать ASR, если Shazam опознал, но текста нет"
+            "ASR, если Shazam опознал без текста"
         )
         form.addRow("", self.lyrics_transcribe_unresolved_cb)
-
         self.lyrics_threshold = QDoubleSpinBox()
         self.lyrics_threshold.setRange(0.0, 1.0)
         self.lyrics_threshold.setSingleStep(0.05)
         self.lyrics_threshold.setDecimals(2)
-        self.lyrics_threshold.setToolTip(
-            "Насколько должен совпасть распознанный ASR-текст с текстом из LRCLIB,\n"
-            "чтобы считать это тем же треком. 1.0 = точное совпадение, 0.0 = любое."
-        )
         form.addRow("Порог fuzzy:", self.lyrics_threshold)
 
+        self.lyrics_save_unrecognized_cb = QCheckBox(
+            "Сохранять тексты без LRCLIB в JSONL (для добавления в LRCLIB)"
+        )
+        self.lyrics_save_unrecognized_cb.setToolTip(
+            "Если текст получен НЕ из LRCLIB (WhisperX / GigaAM / embedded), "
+            "записать его в <output>/unrecognized_lyrics.jsonl.\n"
+            "Формат совместим с LRCLIB API для ручной отправки."
+        )
+        form.addRow("", self.lyrics_save_unrecognized_cb)
+
+        self.lyrics_unrecognized_path_edit = QLineEdit()
+        self.lyrics_unrecognized_path_edit.setPlaceholderText(
+            "(пусто — <output>/unrecognized_lyrics.jsonl)"
+        )
+        form.addRow("Файл unrecognized:", self.lyrics_unrecognized_path_edit)
         return w
 
     def _build_convert_tab(self) -> QWidget:
@@ -238,21 +333,20 @@ class MainWindow(QMainWindow):
         form = QFormLayout(w)
         form.setContentsMargins(16, 16, 16, 16)
         form.setSpacing(10)
-
         self.convert_enabled_cb = QCheckBox("Включён")
         form.addRow("", self.convert_enabled_cb)
-
         self.convert_sr = QComboBox()
         for sr in SAMPLE_RATES:
             self.convert_sr.addItem(f"{sr} Hz", sr)
         form.addRow("Sample rate:", self.convert_sr)
-
         self.convert_ch = QComboBox()
         self.convert_ch.addItem("Моно", 1)
         self.convert_ch.addItem("Стерео", 2)
         form.addRow("Каналы:", self.convert_ch)
-
         self.convert_ffmpeg = QLineEdit()
+        self.convert_concurrency = QSpinBox()
+        self.convert_concurrency.setRange(1, 16)
+        form.addRow("Параллельных конвертаций:", self.convert_concurrency)
         form.addRow("ffmpeg:", self.convert_ffmpeg)
         return w
 
@@ -261,21 +355,22 @@ class MainWindow(QMainWindow):
         form = QFormLayout(w)
         form.setContentsMargins(16, 16, 16, 16)
         form.setSpacing(10)
-
         self.separate_enabled_cb = QCheckBox("Включён")
         form.addRow("", self.separate_enabled_cb)
-
         self.separate_model = QLineEdit()
         form.addRow("Модель:", self.separate_model)
-
         self.separate_stem = QLineEdit()
         form.addRow("Isolate stem:", self.separate_stem)
-
         self.separate_unblend = QLineEdit()
         form.addRow("unblend CLI:", self.separate_unblend)
-
         self.separate_kernels_cb = QCheckBox("Использовать кастомные ядра")
         form.addRow("", self.separate_kernels_cb)
+        self.separate_save_vocals_cb = QCheckBox("Всегда сохранять vocals.wav в output")
+        self.separate_save_vocals_cb.setToolTip(
+            "Независимо от 'Сохранять промежуточные': вокал всегда будет лежать "
+            "в <output>/<relpath>/<stem>/vocals.wav — полезно для LRC-этапа."
+        )
+        form.addRow("", self.separate_save_vocals_cb)
         return w
 
     def _build_transcribe_tab(self) -> QWidget:
@@ -283,53 +378,109 @@ class MainWindow(QMainWindow):
         form = QFormLayout(w)
         form.setContentsMargins(16, 16, 16, 16)
         form.setSpacing(10)
-
         self.transcribe_enabled_cb = QCheckBox("Включён")
         form.addRow("", self.transcribe_enabled_cb)
-
         self.engine_cb = QComboBox()
         self.engine_cb.addItems(ENGINES)
         form.addRow("Движок:", self.engine_cb)
-
         self.model_edit = QLineEdit()
         form.addRow("Модель:", self.model_edit)
-
         self.lang_cb = QComboBox()
         self.lang_cb.setEditable(True)
         self.lang_cb.addItems(LANGUAGES)
-        self.lang_cb.setToolTip(
-            "'auto' — WhisperX определяет язык на каждом файле.\n"
-            "Если знаешь язык — выбери его явно, будет быстрее."
-        )
         form.addRow("Язык:", self.lang_cb)
-
         self.device_cb = QComboBox()
         self.device_cb.addItems(["auto", "cuda", "cpu"])
         form.addRow("Device:", self.device_cb)
-
         self.compute_type_cb = QComboBox()
         self.compute_type_cb.addItems(["auto", "int8_float16", "int8", "float16", "float32"])
         form.addRow("Compute type:", self.compute_type_cb)
-
         self.batch_spin = QSpinBox()
         self.batch_spin.setRange(1, 512)
-        self.batch_spin.setToolTip(
-            "Размер батча сегментов ВНУТРИ одного файла (не файлов!).\n"
-            "8–16 для 8 ГБ VRAM, 32–64 для 24 ГБ."
-        )
         form.addRow("Batch size:", self.batch_spin)
-
         self.gigaam_provider_cb = QComboBox()
         self.gigaam_provider_cb.addItems(["auto", "cuda", "cpu"])
         form.addRow("GigaAM provider:", self.gigaam_provider_cb)
-
         self.local_files_only_cb = QCheckBox("Не проверять HuggingFace (offline)")
-        self.local_files_only_cb.setToolTip(
-            "Если модель уже скачана — включите, чтобы не ждать сетевую проверку."
-        )
         form.addRow("", self.local_files_only_cb)
         return w
 
+    def _build_lrc_tab(self) -> QWidget:
+        w = QWidget()
+        form = QFormLayout(w)
+        form.setContentsMargins(16, 16, 16, 16)
+        form.setSpacing(10)
+
+        info = QLabel(
+            "1) Сгенерировать LRC: WhisperX + align → <output>/<relpath>/<stem>/text.lrc.\n"
+            "2) Встроить LRC: записывает LRC в тег lyrics аудио (USLT/LYRICS/©lyr).\n"
+            "Если 'in-place' выключено — оригинал не трогается, создаётся копия в output."
+        )
+        info.setWordWrap(True)
+        info.setStyleSheet("color: #a6adc8;")
+        form.addRow(info)
+
+        self.lrc_enabled_cb = QCheckBox("Включён (генерация LRC)")
+        self.lrc_enabled_cb.setToolTip(
+            "Отключите, чтобы пропустить этап генерации text.lrc.\n"
+            "Остальные этапы пайплайна продолжат работать."
+        )
+        form.addRow("", self.lrc_enabled_cb)
+        self.lrc_embed_enabled_cb = QCheckBox("Включён (встраивание LRC в аудио)")
+        self.lrc_embed_enabled_cb.setToolTip(
+            "Отключите, чтобы пропустить этап встраивания LRC в теги аудио."
+        )
+        form.addRow("", self.lrc_embed_enabled_cb)
+        self.lrc_skip_existing_cb = QCheckBox("Пропускать готовые text.lrc")
+        self.lrc_skip_existing_cb.setToolTip(
+            "Если <output>/<relpath>/<stem>/text.lrc уже существует и содержит корректный LRC,\n"
+            "файл не будет пересчитан. Папка с metadata.json от другого трека\n"
+            "готовой не считается."
+        )
+        form.addRow("", self.lrc_skip_existing_cb)
+
+        self.lrc_audio_source_cb = QComboBox()
+        self.lrc_audio_source_cb.addItems(["auto", "vocals", "convert", "input"])
+        self.lrc_audio_source_cb.setToolTip(
+            "auto: vocals.wav → convert.wav → 16k WAV из исходника.\n"
+            "vocals/convert: только соответствующий файл.\n"
+            "input: оригинальный mp3/flac."
+        )
+        form.addRow("Источник аудио:", self.lrc_audio_source_cb)
+
+        self.lrc_language_cb = QComboBox()
+        self.lrc_language_cb.setEditable(True)
+        self.lrc_language_cb.addItems(LANGUAGES)
+        form.addRow("Язык (для align):", self.lrc_language_cb)
+
+        self.lrc_align_model_edit = QLineEdit()
+        self.lrc_align_model_edit.setPlaceholderText("(пусто — whisperx по языку)")
+        form.addRow("Align модель:", self.lrc_align_model_edit)
+
+        self.lrc_device_cb = QComboBox()
+        self.lrc_device_cb.addItems(["auto", "cuda", "cpu"])
+        form.addRow("Device:", self.lrc_device_cb)
+
+        self.lrc_embed_in_place_cb = QCheckBox("Встраивать in-place (перезаписывать оригинал)")
+        self.lrc_embed_in_place_cb.setToolTip(
+            "Если выключено (по умолчанию) — создаётся копия в <output>/<relpath>/<stem>/<name>.\n"
+            "Если включено — оригинальный файл модифицируется."
+        )
+        form.addRow("", self.lrc_embed_in_place_cb)
+
+        btns = QHBoxLayout()
+        self.lrc_generate_btn = QPushButton("🎵  Сгенерировать LRC")
+        self.lrc_generate_btn.setObjectName("primary")
+        self.lrc_generate_btn.clicked.connect(lambda: self._run_task("lrc"))
+        self.lrc_embed_btn = QPushButton("📝  Встроить LRC в аудио")
+        self.lrc_embed_btn.clicked.connect(lambda: self._run_task("embed"))
+        btns.addWidget(self.lrc_generate_btn)
+        btns.addWidget(self.lrc_embed_btn)
+        btns.addStretch(1)
+        form.addRow("", btns)
+
+        return w
+
     # ---------------- visual sync ----------------
 
     def _log_device_info(self) -> None:
@@ -347,6 +498,7 @@ class MainWindow(QMainWindow):
         bar.setTabTextColor(2, active if self.convert_enabled_cb.isChecked() else inactive)
         bar.setTabTextColor(3, active if self.separate_enabled_cb.isChecked() else inactive)
         bar.setTabTextColor(4, active if self.transcribe_enabled_cb.isChecked() else inactive)
+        bar.setTabTextColor(5, active if self.lrc_enabled_cb.isChecked() else inactive)
 
     def _sync_engine_fields(self) -> None:
         is_whisperx = self.engine_cb.currentText() == "whisperx"
@@ -365,22 +517,25 @@ class MainWindow(QMainWindow):
         self.recursive_cb.setChecked(cfg.recursive)
         self.keep_intermediate_cb.setChecked(cfg.keep_intermediate)
         self.overwrite_cb.setChecked(cfg.overwrite)
-
+        self.skip_existing_cb.setChecked(cfg.skip_existing)
+        self.resume_partial_cb.setChecked(cfg.resume_partial)
         i = self.mode_cb.findData(cfg.mode)
         if i >= 0:
             self.mode_cb.setCurrentIndex(i)
 
         self.shazam_enabled_cb.setChecked(cfg.shazam.enabled)
+        self.shazam_concurrency.setValue(cfg.shazam.identify_concurrency)
 
         self.lyrics_enabled_cb.setChecked(cfg.lyrics.enabled)
         self.lyrics_user_agent.setText(cfg.lyrics.user_agent)
         self.lyrics_use_tags_cb.setChecked(cfg.lyrics.use_tags_first)
         self.lyrics_use_shazam_cb.setChecked(cfg.lyrics.use_shazam_result)
         self.lyrics_after_asr_cb.setChecked(cfg.lyrics.try_after_asr)
-        self.lyrics_transcribe_unresolved_cb.setChecked(
-            cfg.lyrics.transcribe_if_identified_no_lyrics
-        )
+        self.lyrics_transcribe_unresolved_cb.setChecked(cfg.lyrics.transcribe_if_identified_no_lyrics)
         self.lyrics_threshold.setValue(cfg.lyrics.fuzzy_threshold)
+        self.lyrics_use_embedded_cb.setChecked(cfg.lyrics.use_embedded_lyrics)
+        self.lyrics_save_unrecognized_cb.setChecked(cfg.lyrics.save_unrecognized)
+        self.lyrics_unrecognized_path_edit.setText(cfg.lyrics.unrecognized_path)
 
         self.convert_enabled_cb.setChecked(cfg.convert.enabled)
         i = self.convert_sr.findData(cfg.convert.sample_rate)
@@ -390,12 +545,14 @@ class MainWindow(QMainWindow):
         if i >= 0:
             self.convert_ch.setCurrentIndex(i)
         self.convert_ffmpeg.setText(cfg.convert.ffmpeg_path)
+        self.convert_concurrency.setValue(cfg.convert.convert_concurrency)
 
         self.separate_enabled_cb.setChecked(cfg.separate.enabled)
         self.separate_model.setText(cfg.separate.model)
         self.separate_stem.setText(cfg.separate.isolate_stem)
         self.separate_unblend.setText(cfg.separate.unblend_path)
         self.separate_kernels_cb.setChecked(cfg.separate.custom_kernels)
+        self.separate_save_vocals_cb.setChecked(cfg.separate.save_vocals_copy)
 
         self.transcribe_enabled_cb.setChecked(cfg.transcribe.enabled)
         self.engine_cb.setCurrentText(cfg.transcribe.engine)
@@ -407,6 +564,16 @@ class MainWindow(QMainWindow):
         self.gigaam_provider_cb.setCurrentText(cfg.transcribe.gigaam_provider)
         self.local_files_only_cb.setChecked(cfg.transcribe.local_files_only)
 
+        self.lrc_audio_source_cb.setCurrentText(cfg.lrc.audio_source)
+        self.lrc_language_cb.setCurrentText(cfg.lrc.language or "auto")
+        self.lrc_align_model_edit.setText(cfg.lrc.align_model or "")
+        self.lrc_device_cb.setCurrentText(cfg.lrc.device)
+        self.lrc_embed_in_place_cb.setChecked(cfg.lrc.embed_in_place)
+        self.lrc_enabled_cb.setChecked(cfg.lrc.enabled)
+        self.lrc_embed_enabled_cb.setChecked(cfg.lrc.embed_enabled)
+        self.lrc_skip_existing_cb.setChecked(cfg.lrc.skip_existing)
+        self.chunk_size_spin.setValue(cfg.chunk_size)
+
     def _collect_config(self) -> PipelineConfig:
         cfg = PipelineConfig()
         cfg.input_dir = self.input_edit.text().strip()
@@ -414,10 +581,14 @@ class MainWindow(QMainWindow):
         cfg.recursive = self.recursive_cb.isChecked()
         cfg.keep_intermediate = self.keep_intermediate_cb.isChecked()
         cfg.overwrite = self.overwrite_cb.isChecked()
+        cfg.skip_existing = self.skip_existing_cb.isChecked()
+        cfg.resume_partial = self.resume_partial_cb.isChecked()
         cfg.mode = self.mode_cb.currentData()
 
-        cfg.shazam = ShazamConfig(enabled=self.shazam_enabled_cb.isChecked())
-
+        cfg.shazam = ShazamConfig(
+            enabled=self.shazam_enabled_cb.isChecked(),
+            identify_concurrency=self.shazam_concurrency.value(),
+        )
         cfg.lyrics = LyricsConfig(
             enabled=self.lyrics_enabled_cb.isChecked(),
             user_agent=self.lyrics_user_agent.text().strip() or "audio-transcriber/1.0",
@@ -426,13 +597,16 @@ class MainWindow(QMainWindow):
             try_after_asr=self.lyrics_after_asr_cb.isChecked(),
             transcribe_if_identified_no_lyrics=self.lyrics_transcribe_unresolved_cb.isChecked(),
             fuzzy_threshold=float(self.lyrics_threshold.value()),
+            use_embedded_lyrics=self.lyrics_use_embedded_cb.isChecked(),
+            save_unrecognized=self.lyrics_save_unrecognized_cb.isChecked(),
+            unrecognized_path=self.lyrics_unrecognized_path_edit.text().strip(),
         )
-
         cfg.convert = ConvertConfig(
             enabled=self.convert_enabled_cb.isChecked(),
             sample_rate=int(self.convert_sr.currentData()),
             channels=int(self.convert_ch.currentData()),
             ffmpeg_path=self.convert_ffmpeg.text().strip() or "ffmpeg",
+            convert_concurrency=self.convert_concurrency.value(),
         )
         cfg.separate = SeparateConfig(
             enabled=self.separate_enabled_cb.isChecked(),
@@ -440,8 +614,8 @@ class MainWindow(QMainWindow):
             isolate_stem=self.separate_stem.text().strip() or "vocals",
             unblend_path=self.separate_unblend.text().strip() or "unblend",
             custom_kernels=self.separate_kernels_cb.isChecked(),
+            save_vocals_copy=self.separate_save_vocals_cb.isChecked(),
         )
-
         lang = self.lang_cb.currentText().strip()
         compute = self.compute_type_cb.currentText().strip()
         cfg.transcribe = TranscribeConfig(
@@ -455,6 +629,18 @@ class MainWindow(QMainWindow):
             gigaam_provider=self.gigaam_provider_cb.currentText(),
             local_files_only=self.local_files_only_cb.isChecked(),
         )
+        lrc_lang = self.lrc_language_cb.currentText().strip()
+        cfg.lrc = LrcConfig(
+            enabled=self.lrc_enabled_cb.isChecked(),
+            embed_enabled=self.lrc_embed_enabled_cb.isChecked(),
+            skip_existing=self.lrc_skip_existing_cb.isChecked(),
+            language=None if lrc_lang in ("", "auto") else lrc_lang,
+            align_model=self.lrc_align_model_edit.text().strip() or None,
+            device=self.lrc_device_cb.currentText(),
+            audio_source=self.lrc_audio_source_cb.currentText(),
+            embed_in_place=self.lrc_embed_in_place_cb.isChecked(),
+        )
+        cfg.chunk_size = self.chunk_size_spin.value()
         return cfg
 
     # ---------------- handlers ----------------
@@ -492,11 +678,19 @@ class MainWindow(QMainWindow):
         except Exception as exc:  # noqa: BLE001
             QMessageBox.critical(self, "Ошибка", str(exc))
 
-    def _on_start(self) -> None:
+    def _set_busy(self, busy: bool) -> None:
+        self.start_btn.setEnabled(not busy)
+        self.lrc_generate_btn.setEnabled(not busy)
+        self.lrc_embed_btn.setEnabled(not busy)
+        self.stop_btn.setEnabled(busy)
+
+    def _run_task(self, task: str) -> None:
         cfg = self._collect_config()
         if not cfg.input_dir or not Path(cfg.input_dir).is_dir():
             QMessageBox.warning(self, "Нет входной папки", "Выберите существующую входную папку.")
             return
+
+        if task == "main":
             has_work = any([
                 cfg.shazam.enabled, cfg.lyrics.enabled,
                 cfg.convert.enabled, cfg.separate.enabled, cfg.transcribe.enabled,
@@ -514,19 +708,52 @@ class MainWindow(QMainWindow):
                                     f"Команда '{cfg.separate.unblend_path}' не найдена в PATH.")
                 return
 
+        if task == "lrc" and not shutil.which(cfg.convert.ffmpeg_path):
+            QMessageBox.warning(
+                self, "ffmpeg не найден",
+                f"'{cfg.convert.ffmpeg_path}' не найден в PATH.\n"
+                "LRC-этапу нужен ffmpeg для подготовки аудио."
+            )
+            return
+
+        if task == "lrc" and cfg.transcribe.engine != "whisperx":
+            QMessageBox.warning(
+                self, "LRC требует WhisperX",
+                "Align-модель доступна только для WhisperX. Переключите движок на whisperx."
+            )
+            return
+        if task == "lrc" and not cfg.lrc.enabled:
+            QMessageBox.warning(
+                self, "LRC отключён",
+                "Этап генерации LRC выключен. Включите его на вкладке LRC."
+            )
+            return
+        if task == "lrc" and not cfg.transcribe.enabled:
+            QMessageBox.information(
+                self, "Транскрибация выключена",
+                "Транскрибация отключена, ASR запускаться не будет."
+                "LRC будет построен только по готовым text.txt."
+            )
+        if task == "embed" and not cfg.lrc.embed_enabled:
+            QMessageBox.warning(
+                self, "Встраивание LRC отключено",
+                "Этап встраивания LRC выключен. Включите его на вкладке LRC."
+            )
+            return
+
         self.log_view.clear()
         self._log_device_info()
-        if cfg.transcribe.enabled and cfg.transcribe.language is None:
-            self._append_log("⚠ Язык = auto: WhisperX определит язык на каждом файле.")
-
         self.progress.setValue(0)
         self.progress.setFormat("Запуск…")
-        self.start_btn.setEnabled(False)
-        self.stop_btn.setEnabled(True)
+        self.overall_progress.setMaximum(1)
+        self.overall_progress.setValue(0)
+        self.overall_progress.setFormat("Общий прогресс: запуск…")
+        self._set_busy(True)
 
-        self._worker = PipelineWorker(cfg)
+        self._worker = PipelineWorker(cfg, task=task)
         self._worker.log.connect(self._append_log)
         self._worker.progress.connect(self._on_progress)
+        self._worker.overall_progress.connect(self._on_overall_progress)
         self._worker.finished_ok.connect(self._on_finished)
         self._worker.failed.connect(self._on_failed)
         self._worker.start()
@@ -541,23 +768,61 @@ class MainWindow(QMainWindow):
         sb = self.log_view.verticalScrollBar()
         sb.setValue(sb.maximum())
 
-    def _on_progress(self, done: int, total: int, stage: str) -> None:
-        self.progress.setMaximum(total)
-        self.progress.setValue(done)
+    def _on_progress(self, stage: str, done: int, total: int) -> None:
         label = STAGE_LABELS.get(stage, stage)
-        self.progress.setFormat(f"{label} — %v / %m")
+        if total > 0:
+            self.progress.setMaximum(total)
+            self.progress.setValue(min(done, total))
+            self.progress.setFormat(f"{label}: %v / %m")
+        else:
+            # Nothing to do for this stage — still show which stage we are on
+            # instead of leaving the previous stage's label frozen.
+            self.progress.setMaximum(1)
+            self.progress.setValue(0)
+            self.progress.setFormat(f"{label}: 0 / 0")
+
+    def _on_overall_progress(self, done: int, total: int) -> None:
+        """Overall job progress: completed files / all files in the run."""
+        if total <= 0:
+            self.overall_progress.setMaximum(1)
+            self.overall_progress.setValue(0)
+            self.overall_progress.setFormat("Общий прогресс: 0 / 0")
+            return
+
+        done = max(0, min(done, total))
+        remaining = total - done
+        percent = int(done * 100 / total)
+        self.overall_progress.setMaximum(total)
+        self.overall_progress.setValue(done)
+        self.overall_progress.setFormat(
+            f"Общий прогресс: {done} / {total} ({percent}%) · осталось {remaining}"
+        )
 
     def _on_finished(self, results: list) -> None:
-        self.start_btn.setEnabled(True)
-        self.stop_btn.setEnabled(False)
+        self._set_busy(False)
         ok = sum(1 for r in results if r.error is None)
         fail = len(results) - ok
         self.progress.setFormat(f"Готово — успешно {ok}, с ошибкой {fail}")
+        total = self.overall_progress.maximum()
+        done = self.overall_progress.value()
+        if total > 0 and done >= total:
+            self.overall_progress.setFormat(
+                f"Общий прогресс: {done} / {total} — завершено"
+            )
+        elif total > 0:
+            remaining = total - done
+            self.overall_progress.setFormat(
+                f"Общий прогресс: {done} / {total} — остановлено, осталось {remaining}"
+            )
         self._append_log(f"\n✅ Готово. Успешно: {ok}, с ошибкой: {fail}.")
 
     def _on_failed(self, message: str) -> None:
-        self.start_btn.setEnabled(True)
-        self.stop_btn.setEnabled(False)
+        self._set_busy(False)
         self.progress.setFormat("Ошибка")
+        if self.overall_progress.maximum() > 0:
+            self.overall_progress.setFormat(
+                f"Общий прогресс: {self.overall_progress.value()} / "
+                f"{self.overall_progress.maximum()} — ошибка"
+            )
         self._append_log(f"\n✗ Ошибка пайплайна: {message}")
         QMessageBox.critical(self, "Ошибка", message)

gui/theme.py

@@ -170,20 +170,32 @@ QCheckBox::indicator:checked {
 }
 
 QProgressBar {
-    background-color: #181825;
-    border: 1px solid #313244;
+    background-color: #0b0b12;
+    border: 1px solid #7f849c;
     border-radius: 8px;
     text-align: center;
-    color: #cdd6f4;
-    min-height: 24px;
-    max-height: 24px;
-    font-weight: 500;
+    color: #ffffff;
+    min-height: 26px;
+    max-height: 26px;
+    font-weight: 700;
 }
 QProgressBar::chunk {
-    background-color: #89b4fa;
+    background-color: #1e66f5;
     border-radius: 7px;
     margin: 1px;
 }
+QProgressBar#overallProgress {
+    min-height: 30px;
+    max-height: 30px;
+    font-weight: 700;
+    color: #ffffff;
+    border: 1px solid #a6adc8;
+}
+QProgressBar#overallProgress::chunk {
+    background-color: #2f9e44;
+    border-radius: 8px;
+    margin: 2px;
+}
 
 QScrollBar:vertical {
     background: #181825;

gui/worker.py

@@ -1,4 +1,4 @@
-"""QThread wrapper around core.pipeline.run_pipeline."""
+"""QThread wrapper. Supports three task types: main, lrc, embed."""
 from __future__ import annotations
 
 import threading
@@ -6,18 +6,19 @@ import threading
 from PySide6.QtCore import QThread, Signal
 
 from core.config import PipelineConfig
-from core.pipeline import run_pipeline
 
 
 class PipelineWorker(QThread):
     log = Signal(str)
-    progress = Signal(int, int, str)   # done, total, stage
+    progress = Signal(str, int, int)   # stage, current, total_for_stage
+    overall_progress = Signal(int, int)  # completed_files, total_files
     finished_ok = Signal(list)
     failed = Signal(str)
 
-    def __init__(self, config: PipelineConfig, parent=None):
+    def __init__(self, config: PipelineConfig, task: str = "main", parent=None):
         super().__init__(parent)
         self._config = config
+        self._task = task
         self._stop_event = threading.Event()
 
     def stop(self) -> None:
@@ -25,11 +26,19 @@ class PipelineWorker(QThread):
 
     def run(self) -> None:
         try:
-            results = run_pipeline(
+            if self._task == "lrc":
+                from core.lrc_pipeline import run_lrc_pipeline as runner
+            elif self._task == "embed":
+                from core.lrc_pipeline import run_embed_pipeline as runner
+            else:
+                from core.pipeline import run_pipeline as runner
+
+            results = runner(
                 self._config,
                 log=self.log.emit,
-                on_progress=lambda d, t, s: self.progress.emit(d, t, s),
+                on_progress=lambda s, d, t: self.progress.emit(s, d, t),
                 stop_event=self._stop_event,
+                on_overall_progress=lambda d, t: self.overall_progress.emit(d, t),
             )
             self.finished_ok.emit(results)
         except Exception as exc:  # noqa: BLE001

main.py

@@ -26,20 +26,8 @@ def preload_cudnn() -> None:
                 pass
 
 
-def enable_tf32() -> None:
-    """Enable TF32 for faster matmul on Ampere+ GPUs (RTX 30/40 series)."""
-    try:
-        import torch
-        if torch.cuda.is_available():
-            torch.backends.cuda.matmul.allow_tf32 = True
-            torch.backends.cudnn.allow_tf32 = True
-    except ImportError:
-        pass
-
-
 def main() -> int:
     preload_cudnn()
-    enable_tf32()
 
     from PySide6.QtWidgets import QApplication
     from gui.theme import DARK_STYLESHEET

mark_no_speech.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python3
+"""Mark tracks with metadata.empty_text=true if their text is missing or hallucinated."""
+import json, sys
+from pathlib import Path
+
+HALLUCINATION = (
+    "продолжение следует", "субтитры сделан", "субтитры создава",
+    "created with", "russian-area.de", "музыка музыка",
+)
+
+
+def main(root: Path) -> int:
+    count = 0
+    for meta in root.rglob("metadata.json"):
+        track = meta.parent
+        txt = track / "text.txt"
+        if txt.is_file():
+            body = txt.read_text(encoding="utf-8", errors="ignore").strip().lower()
+            # If the text is real (not hallucination) and not empty, skip marking.
+            if body and not any(h in body for h in HALLUCINATION):
+                continue
+            # Remove the text.txt so the next run will re-process it.
+            txt.unlink()
+        # Set empty_text=True in metadata.
+        try:
+            data = json.loads(meta.read_text(encoding="utf-8"))
+        except (OSError, ValueError):
+            data = {}
+        data["empty_text"] = True
+        meta.write_text(
+            json.dumps(data, indent=2, ensure_ascii=False), encoding="utf-8",
+        )
+        print(f"marked: {track}")
+        count += 1
+    print(f"\nmarked {count} tracks")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(Path(sys.argv[1] if len(sys.argv) > 1 else ".")))

migrate_output_to_paths.py

@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+"""Migrate flat output/<stem>/ to output/<relpath>/<stem>/ (one-shot)."""
+from __future__ import annotations
+import argparse, json, shutil, sys
+from pathlib import Path
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--input-root", required=True, type=Path)
+    ap.add_argument("--output-root", required=True, type=Path)
+    ap.add_argument("--dry-run", action="store_true")
+    args = ap.parse_args()
+
+    input_root = args.input_root.resolve()
+    output_root = args.output_root.resolve()
+
+    if not input_root.is_dir():
+        print(f"input root not found: {input_root}", file=sys.stderr); return 1
+    if not output_root.is_dir():
+        print(f"output root not found: {output_root}", file=sys.stderr); return 1
+
+    moved = legacy = conflicts = errors = skipped = 0
+    artifact_names = (
+        "metadata.json", "text.txt", "text.lrc",
+        "vocals.wav", "convert.wav",
+    )
+    for track_dir in sorted(output_root.iterdir()):
+        if not track_dir.is_dir():
+            continue
+        if track_dir.name.startswith("_"):
+            continue
+        # Never treat an album/parent folder as a flat track folder.
+        if not any((track_dir / name).is_file() for name in artifact_names):
+            skipped += 1
+            continue
+
+        target = output_root / "_legacy" / track_dir.name
+        meta_file = track_dir / "metadata.json"
+        if meta_file.is_file():
+            try:
+                data = json.loads(meta_file.read_text(encoding="utf-8"))
+                stored = data.get("path") if isinstance(data, dict) else None
+            except (OSError, ValueError):
+                stored = None
+            if stored:
+                try:
+                    rel = Path(stored).relative_to(input_root)
+                    target = output_root / rel.with_suffix("")
+                except ValueError:
+                    target = output_root / "_legacy" / track_dir.name
+
+        if target == track_dir:
+            continue
+        if track_dir in target.parents:
+            print(f"skip: {track_dir.name} -> target inside itself")
+            skipped += 1
+            continue
+        if target.exists():
+            print(f"conflict: {track_dir.name} -> "
+                  f"{target.relative_to(output_root)}")
+            conflicts += 1
+            continue
+
+        if args.dry_run:
+            print(f"DRY: {track_dir.name} -> {target.relative_to(output_root)}")
+            continue
+        try:
+            target.parent.mkdir(parents=True, exist_ok=True)
+            shutil.move(str(track_dir), str(target))
+            if "_legacy" in target.parts:
+                legacy += 1
+            else:
+                moved += 1
+        except OSError as e:
+            print(f"error: {track_dir.name}: {e}")
+            errors += 1
+
+    print(f"\nmoved: {moved}  quarantined: {legacy}  "
+          f"conflicts: {conflicts}  skipped: {skipped}  errors: {errors}  "
+          f"(dry_run={args.dry_run})")
+    return 0 if errors == 0 else 2
+
+
+if __name__ == "__main__":
+    sys.exit(main())

requirements.txt

@@ -3,6 +3,7 @@ PySide6>=6.6
 
 # ASR
 whisperx>=3.1
+stable-ts>=2.16
 onnx-asr>=0.4
 
 # Recognition & lyrics