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

core/lrc.py

@@ -70,6 +70,35 @@ def is_lrc(content: str) -> bool:
 
 # ---------------------------------------------------------------- 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.
 
@@ -78,14 +107,18 @@ def read_embedded_lyrics(audio_path: Path) -> Optional[str]:
     suffix = audio_path.suffix.lower()
     try:
         if suffix in (".mp3", ".mp2"):
-            return _read_mp3(audio_path)
+            text = _read_mp3(audio_path)
-        if suffix in (".flac", ".ogg", ".opus"):
+        elif suffix in (".flac", ".ogg", ".opus"):
-            return _read_vorbis(audio_path)
+            text = _read_vorbis(audio_path)
-        if suffix in (".m4a", ".mp4", ".m4b", ".aac"):
+        elif suffix in (".m4a", ".mp4", ".m4b", ".aac"):
-            return _read_mp4(audio_path)
+            text = _read_mp4(audio_path)
-    except Exception:  # noqa: BLE001 — never break the pipeline on tag issues
+        else:
+            return None
+    except Exception:  # noqa: BLE001
         return None
-    return None
+    if text and _is_junk_lyrics(text):
+        return None
+    return text
 
 
 def _read_mp3(path: Path) -> Optional[str]:

core/lrc_pipeline.py

@@ -12,9 +12,12 @@ 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, _save_unrecognized, collect_audio_files,
+    FileResult, _artifact_belongs_to, _output_dir_for, _save_unrecognized,
+    collect_audio_files,
+)
+from .transcribe import (
+    CudaOOMError, align_text_to_audio, transcribe, transcribe_aligned,
 )
-from .transcribe import transcribe_aligned
 
 
 LogFn = Callable[[str], None]
@@ -23,11 +26,13 @@ OverallFn = Callable[[int, int], None]        # completed_files, total_files
 
 
 def _pick_audio_source(
-    f: Path, output_dir: Path, config: PipelineConfig, work_dir: Path, log: LogFn,
+    f: Path, input_root: Path, output_dir: Path, config: PipelineConfig,
+    work_dir: Path, log: LogFn,
 ) -> Path:
     mode = (config.lrc.audio_source or "auto").lower()
-    vocals = output_dir / f.stem / "vocals.wav"
+    track_out = _output_dir_for(f, input_root, output_dir)
-    convert = output_dir / f.stem / "convert.wav"
+    vocals = track_out / "vocals.wav"
+    convert = track_out / "convert.wav"
     if mode == "vocals":
         return vocals
     if mode == "convert":
@@ -83,7 +88,17 @@ def run_lrc_pipeline(
         log("Аудиофайлы не найдены.")
         return []
 
-    lang = config.lrc.language or config.transcribe.language
+    # 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}")
 
@@ -100,11 +115,28 @@ def run_lrc_pipeline(
             break
         log(f"\n▶ [{idx}/{total}] {f.name}")
 
-        track_out = output_dir / f.stem
+        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:
@@ -118,6 +150,7 @@ def run_lrc_pipeline(
 
         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")
@@ -125,36 +158,96 @@ def run_lrc_pipeline(
                 results.append(FileResult(f, None, track_out, None, "embedded+lrc"))
                 continue
 
-            audio_path = _pick_audio_source(f, output_dir, config, tmp, log)
+            # 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 аудио не найден: {audio_path}. "
-                    "Включите conversion/separation в основном пайплайне "
+                    f"Источник={config.lrc.audio_source}, "
-                    "или выберите audio_source=input."
+                    f"convert.enabled={config.convert.enabled}, "
+                    f"separate.enabled={config.separate.enabled}"
                 )
             log(f"  [align] {audio_path.name}")
 
-            segments, detected = transcribe_aligned(
+            # 3) готовый текст → forced alignment (быстро, без ASR)
-                audio_path,
+            if text is not None:
-                model=config.transcribe.model,
+                log(f"  [align-only] text.txt найден ({len(text)} символов)")
-                language=lang,
+                segments, detected = align_text_to_audio(
-                device=config.lrc.device,
+                    audio_path,
-                compute_type=config.transcribe.compute_type,
+                    text,
-                batch_size=config.transcribe.batch_size,
+                    language=lang,
-                align_model=config.lrc.align_model,
+                    model=config.transcribe.model,
-                local_files_only=config.transcribe.local_files_only,
+                    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})")
 
-            # NEW: save to JSONL for LRCLIB submission
             _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}")
@@ -206,7 +299,8 @@ def run_embed_pipeline(
             log("⏹ Отменено")
             break
 
-        lrc_path = output_dir / f.stem / "text.lrc"
+        track_out = _output_dir_for(f, input_dir, output_dir)
+        lrc_path = track_out / "text.lrc"
         error: Optional[str] = None
 
         try:
@@ -220,7 +314,7 @@ def run_embed_pipeline(
             if config.lrc.embed_in_place:
                 target = f
             else:
-                target = output_dir / f.stem / f.name
+                target = track_out / f.name
                 target.parent.mkdir(parents=True, exist_ok=True)
                 if not target.exists() or config.overwrite:
                     shutil.copy2(f, target)

core/pipeline.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 import difflib
 import json
 import shutil
+import sys
 import tempfile
 import threading
 from concurrent.futures import ThreadPoolExecutor, as_completed
@@ -18,7 +19,9 @@ from .lyrics import LyricsMatch, search_by_identity, search_by_text
 from .metadata import TrackMetadata, read_metadata
 from .recognize import ShazamResult, recognize
 from .separate import separate_vocals, separate_vocals_batch
-from .transcribe import transcribe
+from .transcribe import (
+    CudaOOMError, preload_transcription_model, transcribe,
+)
 from .lrc import is_lrc, read_embedded_lyrics
 
 
@@ -48,11 +51,18 @@ class Cancelled(Exception):
 # ---------------------------------------------------------------- helpers
 
 def _free_memory() -> None:
-    """Free unreferenced RAM and VRAM. Keeps model caches alive."""
+    """Free unreferenced RAM and VRAM. Keeps model caches alive.
+
+    Never import torch here: if the current run does not use ASR, importing it
+    would initialize a CUDA context and steal VRAM from separation. If torch is
+    already loaded by the ASR stage, we can safely empty its cache.
+    """
     import gc
     gc.collect()
+    torch = sys.modules.get("torch")
+    if torch is None:
+        return
     try:
-        import torch
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
     except Exception:  # noqa: BLE001 — never fail cleanup on a broken CUDA context
@@ -150,6 +160,130 @@ def _write_metadata_json(path: Path, data: Optional[dict]) -> None:
     path.write_text(json.dumps(data, indent=2, ensure_ascii=False), encoding="utf-8")
 
 
+def _merge_metadata_file(existing: Path, data: Optional[dict]) -> Optional[dict]:
+    """Keep tags/shazam from an earlier run when the new data has less info.
+
+    Reprocessing a track for one stage only (e.g. separation) must not replace
+    a rich metadata.json with a path-only dict.
+    """
+    if not data or not existing.is_file():
+        return data
+    try:
+        old = json.loads(existing.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        return data
+    if not isinstance(old, dict):
+        return data
+    merged = dict(old)
+    merged.update(data)
+    return merged
+
+
+def _output_dir_for(f: Path, input_root: Path, output_root: Path) -> Path:
+    """Path-based track folder: output/<relpath_without_ext>/."""
+    try:
+        rel = f.relative_to(input_root)
+    except ValueError:
+        rel = Path(f.name)
+    return output_root / rel.with_suffix("")
+
+
+def _existing_artifacts(
+    input_root: Path, output_root: Path, input_path: Path,
+) -> tuple[bool, bool, bool]:
+    """Return (has_text, has_vocals, has_convert) for the matching output folder.
+
+    Artifacts are ignored when ``metadata.json`` says the folder belongs to a
+    different input file with the same stem.
+    """
+    track_out = _output_dir_for(input_path, input_root, output_root)
+    if not _artifact_belongs_to(track_out, input_path):
+        return False, False, False
+    return (
+        (track_out / "text.txt").is_file(),
+        (track_out / "vocals.wav").is_file(),
+        (track_out / "convert.wav").is_file(),
+    )
+
+
+def _final_artifact_present(track_out: Path, config: PipelineConfig) -> bool:
+    """True if the artifact of the last enabled stage already exists.
+
+    ``skip_existing`` must not treat an old ``text.txt`` as "done" when the
+    current run only needs separation: the final stage here is ``vocals.wav``,
+    and without it the track still has work to do.
+    """
+    if config.transcribe.enabled:
+        return (track_out / "text.txt").is_file()
+    if config.separate.enabled:
+        return (track_out / "vocals.wav").is_file()
+    if config.convert.enabled:
+        return (track_out / "convert.wav").is_file()
+    return False
+
+
+def _is_track_finished(
+    track_out: Path, config: PipelineConfig, belongs: bool = True,
+) -> bool:
+    """True when ``skip_existing`` may skip this output folder entirely."""
+    if not belongs:
+        return False
+    # If metadata says this track had an empty ASR result, treat it as
+    # "finished" so skip_existing does not keep re‑running it.
+    meta = track_out / "metadata.json"
+    if meta.is_file():
+        try:
+            data = json.loads(meta.read_text(encoding="utf-8"))
+            if isinstance(data, dict) and data.get("empty_text"):
+                return True
+        except (OSError, ValueError):
+            pass
+    if _final_artifact_present(track_out, config):
+        return True
+    # Identification-only runs still finish with text.txt (no other stage
+    # produces an artifact of its own).
+    if (not config.convert.enabled and not config.separate.enabled
+            and not config.transcribe.enabled
+            and (config.shazam.enabled or config.lyrics.enabled)):
+        return (track_out / "text.txt").is_file()
+    return False
+
+
+def _needs_work(
+    config: PipelineConfig, input_root: Path, output_root: Path, f: Path,
+) -> bool:
+    """True if at least one enabled stage still has to do something for ``f``.
+
+    This mirrors the resume rules used by the staged/sequential pipelines:
+      * final artifact present + skip_existing -> resolved, nothing to do
+      * vocals.wav + resume_partial     -> convert/separate are skipped
+      * convert.wav + resume_partial    -> convert is skipped
+    Files with no pending work are filtered out before batching, so batches
+    stay full instead of being filled by already-finished tracks.
+    """
+    track_out = _output_dir_for(f, input_root, output_root)
+    belongs = _artifact_belongs_to(track_out, f)
+    has_text, has_vocals, has_convert = _existing_artifacts(
+        input_root, output_root, f,
+    )
+
+    if config.skip_existing and _is_track_finished(track_out, config, belongs):
+        return False
+
+    vocals_ready = config.resume_partial and has_vocals
+    convert_ready = config.resume_partial and (has_vocals or has_convert)
+
+    if config.shazam.enabled or config.lyrics.enabled:
+        return True  # identification still has to run
+    if config.convert.enabled and not convert_ready:
+        return True
+    if config.separate.enabled and not vocals_ready:
+        return True
+    if config.transcribe.enabled:
+        return True  # text is missing or skip_existing is disabled
+    return False
+
+
 def _save_unrecognized(
     config: PipelineConfig,
     f: Path,
@@ -532,7 +666,8 @@ def run_pipeline(
         log("Аудиофайлы не найдены.")
         return []
     mode = (config.mode or "sequential").lower()
-    log(f"Найдено треков: {len(files)}   |   режим: {mode}   |   "
+    total_all = len(files)
+    log(f"Найдено треков: {total_all}   |   режим: {mode}   |   "
         f"keep_intermediate={config.keep_intermediate}")
     if config.separate.save_vocals_copy:
         log("  ✓ Вокал будет сохраняться всегда (save_vocals_copy=True)")
@@ -541,17 +676,37 @@ def run_pipeline(
     if config.lyrics.enabled:
         log(f"  LRCLIB: включён (порог fuzzy={config.lyrics.fuzzy_threshold:.2f})")
     if config.skip_existing:
-        log("  Resume: готовые файлы (text.txt) будут пропущены")
+        log("  Resume: готовые файлы (результат последнего этапа) будут пропущены")
     if config.resume_partial:
         log("  Resume: существующие vocals.wav / convert.wav будут переиспользованы")
+
+    # Pre-scan before chunking: remove files that are already complete for the
+    # current set of enabled stages. Without this, skipped tracks consumed
+    # batch slots and a staged run produced many half-empty batches.
+    pending = [f for f in files if _needs_work(config, input_dir, output_dir, f)]
+    already_done = total_all - len(pending)
+    if already_done:
+        log(f"  ✓ Уже готово: {already_done} — в батчи не попадут")
+        log(f"  → К обработке: {len(pending)}")
+    files = pending
+    if not files:
+        log("  Все файлы уже готовы — работа не требуется.")
+        if on_overall_progress:
+            on_overall_progress(total_all, total_all)
+        return []
+
     if on_overall_progress:
-        on_overall_progress(0, len(files))
+        on_overall_progress(already_done, total_all)
     if mode == "staged":
         return _run_staged(
             config, files, log, on_progress, stop_event, on_overall_progress,
+            overall_base=already_done, overall_total=total_all,
+            input_root=input_dir,
         )
     return _run_sequential(
         config, files, log, on_progress, stop_event, on_overall_progress,
+        overall_base=already_done, overall_total=total_all,
+        input_root=input_dir,
     )
 
 
@@ -559,16 +714,24 @@ def _run_sequential(
     config: PipelineConfig, files: list[Path], log: LogFn,
     on_progress: Optional[ProgressFn], stop_event: Optional[threading.Event],
     on_overall_progress: Optional[OverallFn] = None,
+    overall_base: int = 0,
+    overall_total: Optional[int] = None,
+    input_root: Optional[Path] = None,
 ) -> list[FileResult]:
     total = len(files)
+    if overall_total is None:
+        overall_total = overall_base + total
+    if input_root is None:
+        input_root = Path(config.input_dir)
     results: list[FileResult] = []
     output_root = Path(config.output_dir)
+    log("Раскладка: output/<relpath_from_input>/<stem>/")
     ident_tmp = Path(tempfile.mkdtemp(prefix="at_ident_"))
     ident_files = files
     if config.skip_existing:
         ident_files = [
             f for f in files
-            if not (output_root / f.stem / "text.txt").is_file()
+            if not (_output_dir_for(f, input_root, output_root) / "text.txt").is_file()
         ]
     try:
         ident_map = _run_identification_parallel(
@@ -585,24 +748,32 @@ def _run_sequential(
             on_progress("file", idx - 1, total)
         log(f"\n▶ [{idx}/{total}] {f.name}")
 
-        track_out = output_root / f.stem
+        track_out = _output_dir_for(f, input_root, output_root)
         track_out.mkdir(parents=True, exist_ok=True)
         belongs = _artifact_belongs_to(track_out, f)
+        if not belongs and any(
+            (track_out / name).exists()
+            for name in ("text.txt", "vocals.wav", "convert.wav")
+        ):
+            log(f"  ⚠ коллизия раскладки: {track_out} занят другим треком — обрабатываю заново")
 
-        if config.skip_existing:
+        if config.skip_existing and _is_track_finished(track_out, config, belongs):
             existing_text = track_out / "text.txt"
-            if existing_text.is_file() and belongs:
+            existing: Optional[str] = None
+            if existing_text.is_file():
                 try:
                     existing = existing_text.read_text(encoding="utf-8")
                 except OSError:
                     existing = None
-                log("  ⊘ text.txt уже есть — пропуск")
+            log("  ⊘ результат последнего этапа уже есть — пропуск")
-                if on_progress:
+            if on_progress:
-                    on_progress("file", idx, total)
+                on_progress("file", idx, total)
-                if on_overall_progress:
+            if on_overall_progress:
-                    on_overall_progress(idx, total)
+                on_overall_progress(
-                results.append(FileResult(f, existing, track_out, None, "existing"))
+                    min(overall_base + idx, overall_total), overall_total,
-                continue
+                )
+            results.append(FileResult(f, existing, track_out, None, "existing"))
+            continue
 
         tmp_root = Path(tempfile.mkdtemp(prefix=f"at_{f.stem}_"))
 
@@ -699,18 +870,31 @@ def _run_sequential(
                     gigaam_provider=config.transcribe.gigaam_provider,
                     local_files_only=config.transcribe.local_files_only,
                 )
+                text = (text or "").strip()
                 source = config.transcribe.engine
-                log(f"  ✓ распознано {len(text)} символов")
+                if not text:
-                if (config.lyrics.enabled and config.lyrics.try_after_asr):
+                    text = None
-                    matched = _try_post_asr_match(text, config, log)
+                    source = "empty_result"
-                    if matched:
+                    error = "EmptyResult: ASR returned no text"
-                        text = matched
+                    log(f"  ⚠ {f.name}: текст не извлечён (пустой результат ASR)")
-                        source = f"{source}+lrclib"
+                else:
-                (tmp_root / "text.txt").write_text(text, encoding="utf-8")
+                    log(f"  ✓ распознано {len(text)} символов")
+                    if (config.lyrics.enabled and config.lyrics.try_after_asr):
+                        matched = _try_post_asr_match(text, config, log)
+                        if matched:
+                            text = matched
+                            source = f"{source}+lrclib"
+                    (tmp_root / "text.txt").write_text(text, encoding="utf-8")
 
             metadata_dict = _build_metadata_dict(
                 ident.metadata, ident.shazam, source, input_path=f,
             )
+            if metadata_dict.get("empty_text") is None:
+                metadata_dict["empty_text"] = False
+            if belongs:
+                metadata_dict = _merge_metadata_file(
+                    track_out / "metadata.json", metadata_dict,
+                )
             if metadata_dict:
                 _write_metadata_json(tmp_root / "metadata.json", metadata_dict)
 
@@ -729,6 +913,10 @@ def _run_sequential(
             log("⏹ Отменено")
             was_cancelled = True
             break
+        except CudaOOMError as e:
+            error = f"CudaOOMError: {e}"
+            source = "oom_skipped"
+            log(f"  ⊘ {f.name}: OOM после ретраев — пропущен, идём дальше")
         except Exception as e:  # noqa: BLE001
             error = f"{type(e).__name__}: {e}"
             log(f"  ✗ Ошибка: {error}")
@@ -758,7 +946,9 @@ def _run_sequential(
                 if on_progress:
                     on_progress("file", idx, total)
                 if on_overall_progress:
-                    on_overall_progress(idx, total)
+                    on_overall_progress(
+                        min(overall_base + idx, overall_total), overall_total,
+                    )
 
     _log_summary(results, log)
     return results
@@ -768,19 +958,28 @@ def _run_staged(
     config: PipelineConfig, files: list[Path], log: LogFn,
     on_progress: Optional[ProgressFn], stop_event: Optional[threading.Event],
     on_overall_progress: Optional[OverallFn] = None,
+    overall_base: int = 0,
+    overall_total: Optional[int] = None,
+    input_root: Optional[Path] = None,
 ) -> list[FileResult]:
     """Chunked staged mode: process `chunk_size` files, free memory, repeat."""
     output_root = Path(config.output_dir)
+    if input_root is None:
+        input_root = Path(config.input_dir)
     total = len(files)
+    if overall_total is None:
+        overall_total = overall_base + total
     chunk_size = max(1, int(config.chunk_size or total))
+    log("Раскладка: output/<relpath_from_input>/<stem>/")
 
     if chunk_size >= total:
         log(f"Один батч: {total} файлов (chunk_size={chunk_size})")
         results = _run_staged_batch(
             config, files, output_root, log, on_progress, stop_event,
-            overall_offset=0, overall_total=total,
+            overall_offset=overall_base, overall_total=overall_total,
             on_overall_progress=on_overall_progress,
-            overall_counter=[0],
+            overall_counter=[overall_base],
+            input_root=input_root,
         )
         _free_memory()
         _log_summary(results, log)
@@ -790,7 +989,7 @@ def _run_staged(
     log(f"Батчей: {n_batches} × {chunk_size} файлов (всего {total})")
 
     all_results: list[FileResult] = []
-    overall_counter = [0]
+    overall_counter = [overall_base]
     for b in range(n_batches):
         if stop_event is not None and stop_event.is_set():
             log("⏹ Отменено")
@@ -805,9 +1004,10 @@ def _run_staged(
         try:
             batch_results = _run_staged_batch(
                 config, chunk, output_root, log, on_progress, stop_event,
-                overall_offset=start, overall_total=total,
+                overall_offset=overall_base + start, overall_total=overall_total,
                 on_overall_progress=on_overall_progress,
                 overall_counter=overall_counter,
+                input_root=input_root,
             )
             all_results.extend(batch_results)
         except Exception as e:  # noqa: BLE001
@@ -831,9 +1031,12 @@ def _run_staged_batch(
     overall_total: Optional[int] = None,
     on_overall_progress: Optional[OverallFn] = None,
     overall_counter: Optional[list[int]] = None,
+    input_root: Optional[Path] = None,
 ) -> list[FileResult]:
     """One chunk of the staged pipeline. All temps are tracked and removed
     in the `finally` block at the end of this call."""
+    if input_root is None:
+        input_root = Path(config.input_dir)
     if overall_total is None:
         overall_total = len(batch_files)
 
@@ -890,7 +1093,7 @@ def _run_staged_batch(
     resumed_partial = 0
     mismatched_folders = 0
     for f in batch_files:
-        track_out = output_root / f.stem
+        track_out = _output_dir_for(f, input_root, output_root)
         audio_path: Path = f
         skip_convert = False
         skip_separate = False
@@ -906,17 +1109,18 @@ def _run_staged_batch(
             )
             if has_artifacts:
                 mismatched_folders += 1
-                log(f"  ⚠ {f.name}: папка {track_out.name} относится к другому треку — обрабатываю заново")
+                log(f"  ⚠ коллизия раскладки: {track_out} занят другим треком — обрабатываю заново")
 
         text_file = track_out / "text.txt"
-        if config.skip_existing and belongs and text_file.is_file():
+        if config.skip_existing and _is_track_finished(track_out, config, belongs):
             resolved = True
             skipped_existing += 1
             source = "existing"
-            try:
+            if text_file.is_file():
-                text = text_file.read_text(encoding="utf-8")
+                try:
-            except OSError:
+                    text = text_file.read_text(encoding="utf-8")
-                text = None
+                except OSError:
+                    text = None
         elif config.resume_partial and belongs:
             vocals = track_out / "vocals.wav"
             convert = track_out / "convert.wav"
@@ -938,7 +1142,7 @@ def _run_staged_batch(
         }
 
     if skipped_existing:
-        log(f"  ⊘ Пропущено готовых (text.txt): {skipped_existing}")
+        log(f"  ⊘ Пропущено готовых (результат последнего этапа): {skipped_existing}")
     if resumed_partial:
         log(f"  ↻ Подхвачено промежуточных (vocals/convert): {resumed_partial}")
     if mismatched_folders:
@@ -979,10 +1183,13 @@ def _run_staged_batch(
                             ident.metadata, ident.shazam, ident.source,
                             input_path=f,
                         )
-                        track_out = output_root / f.stem
+                        track_out = _output_dir_for(f, input_root, output_root)
                         track_out.mkdir(parents=True, exist_ok=True)
                         if ident.text:
                             (track_out / "text.txt").write_text(ident.text, encoding="utf-8")
+                            md = state[f].setdefault("metadata_dict", {})
+                            if md.get("empty_text") is None:
+                                md["empty_text"] = False
                         if state[f]["metadata_dict"]:
                             _write_metadata_json(track_out / "metadata.json", state[f]["metadata_dict"])
                         _save_unrecognized(
@@ -1010,7 +1217,22 @@ def _run_staged_batch(
                 def separate_single(f: Path) -> None:
                     try:
                         src = state[f]["audio_path"]
-                        target = src.parent / "vocals.wav"
+                        # Never write next to the original input: with
+                        # convert.enabled=False src == f, so every track would
+                        # overwrite input_dir/vocals.wav. Only reuse the track
+                        # output folder when the source already lives there.
+                        track_dir = _output_dir_for(f, input_root, output_root)
+                        if src.parent == track_dir:
+                            # Already in the persistent track folder (resume).
+                            target = src.parent / "vocals.wav"
+                        elif src.parent == f.parent:
+                            # src is the original input file (convert disabled).
+                            # Never write next to it: all tracks share input_dir
+                            # and would overwrite input_dir/vocals.wav.
+                            target = _mkdtemp("at_sep_") / "vocals.wav"
+                        else:
+                            # Freshly converted file in its own unique temp dir.
+                            target = src.parent / "vocals.wav"
                         with tempfile.TemporaryDirectory() as tmp:
                             raw = separate_vocals(
                                 src, Path(tmp), model=config.separate.model,
@@ -1027,11 +1249,15 @@ def _run_staged_batch(
 
                 # Resumed convert.wav files already live in output_root/<stem>,
                 # while freshly converted files live in unique temp dirs.
+                # Original inputs (convert disabled) must also go through the
+                # one-by-one path: batching them shares input_dir and can
+                # collide on vocals.wav.
                 # unblend's `{parent}` batch layout would mix those roots, so
-                # separate resumed files one by one and batch the rest.
+                # separate resumed/original files one by one and batch the rest.
                 resumed = [
                     f for f in todo
-                    if state[f]["audio_path"].parent == output_root / f.stem
+                    if state[f]["audio_path"].parent == _output_dir_for(f, input_root, output_root)
+                    or state[f]["audio_path"].parent == f.parent
                 ]
                 resumed_set = set(resumed)
                 batchable = [f for f in todo if f not in resumed_set]
@@ -1092,6 +1318,20 @@ def _run_staged_batch(
         if config.transcribe.enabled and not cancelled():
             todo = active()
             log(f"\n══ Этап 3 · Транскрибация ({len(todo)} файлов) ══")
+            if todo:
+                # Load once, after separation has finished. Every later file
+                # and every later batch reuses this instance.
+                try:
+                    preload_transcription_model(
+                        engine=config.transcribe.engine,
+                        model=config.transcribe.model,
+                        device=config.transcribe.device,
+                        compute_type=config.transcribe.compute_type,
+                        gigaam_provider=config.transcribe.gigaam_provider,
+                        local_files_only=config.transcribe.local_files_only,
+                    )
+                except Exception as e:  # noqa: BLE001
+                    log(f"  ⚠ Не удалось загрузить модель заранее: {e}")
             progress("transcribe", 0, len(todo))
             for i, f in enumerate(todo, 1):
                 if cancelled():
@@ -1108,9 +1348,20 @@ def _run_staged_batch(
                         gigaam_provider=config.transcribe.gigaam_provider,
                         local_files_only=config.transcribe.local_files_only,
                     )
-                    state[f]["text"] = text
+                    text = (text or "").strip()
-                    state[f]["source"] = config.transcribe.engine
+                    if not text:
-                    log(f"  ✓ {f.name} ({len(text)} символов)")
+                        state[f]["text"] = None
+                        state[f]["source"] = "empty_result"
+                        state[f]["error"] = "EmptyResult: ASR returned no text"
+                        log(f"  ⚠ {f.name}: текст не извлечён (пустой результат ASR)")
+                    else:
+                        state[f]["text"] = text
+                        state[f]["source"] = config.transcribe.engine
+                        log(f"  ✓ {f.name} ({len(text)} символов)")
+                except CudaOOMError as e:
+                    state[f]["error"] = f"CudaOOMError: {e}"
+                    state[f]["source"] = "oom_skipped"
+                    log(f"  ⊘ {f.name}: OOM — пропущен")
                 except Exception as e:  # noqa: BLE001
                     state[f]["error"] = f"{type(e).__name__}: {e}"
                     log(f"  ✗ {f.name}: {e}")
@@ -1133,7 +1384,7 @@ def _run_staged_batch(
                     if isinstance(src_audio, Path):
                         vocal = src_audio.parent / "vocals.wav"
                         if vocal.exists():
-                            track_out = output_root / f.stem
+                            track_out = _output_dir_for(f, input_root, output_root)
                             track_out.mkdir(parents=True, exist_ok=True)
                             dst = track_out / "vocals.wav"
                             if dst.exists():
@@ -1174,8 +1425,13 @@ def _run_staged_batch(
             )
             s["metadata_dict"] = metadata_dict
 
-            track_out = output_root / f.stem
+            track_out = _output_dir_for(f, input_root, output_root)
             track_out.mkdir(parents=True, exist_ok=True)
+            if _artifact_belongs_to(track_out, f):
+                metadata_dict = _merge_metadata_file(
+                    track_out / "metadata.json", metadata_dict,
+                )
+                s["metadata_dict"] = metadata_dict
 
             audio_work = s["audio_path"]
             if isinstance(audio_work, Path) and audio_work != f:
@@ -1193,6 +1449,9 @@ def _run_staged_batch(
             else:
                 if text:
                     (track_out / "text.txt").write_text(text, encoding="utf-8")
+                    md = metadata_dict.setdefault("empty_text", False)
+                else:
+                    md = metadata_dict.setdefault("empty_text", True)
                 if metadata_dict:
                     _write_metadata_json(track_out / "metadata.json", metadata_dict)
 
@@ -1200,7 +1459,10 @@ def _run_staged_batch(
                 config, f, text, source,
                 metadata=(ident.metadata if ident else None),
             )
-            log(f"  ✓ {f.name}")
+            if text is not None or s["resolved"]:
+                log(f"  ✓ {f.name}")
+            else:
+                log(f"  ⊘ {f.name}: пропущен (нет текста)")
             mark_overall(is_complete(s))
 
     finally:
@@ -1216,7 +1478,7 @@ def _run_staged_batch(
     return [
         FileResult(
             f, state[f]["text"],
-            output_root / f.stem if state[f]["error"] is None else None,
+            _output_dir_for(f, input_root, output_root) if state[f]["error"] is None else None,
             state[f]["error"], state[f]["source"], state[f]["metadata_dict"],
         )
         for f in batch_files

core/transcribe.py

@@ -6,15 +6,134 @@ 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
 
 
-# Set permanently once the CUDA context has died (e.g. after an unrecoverable
+# The model caches below are process-wide. They are intentionally never
-# OOM). Every later CUDA transcription fails instantly instead of re-loading
+# cleared between files or batches: reloading WhisperX takes minutes and is
-# the model forever and hanging the pipeline.
+# what made runs look like they reloaded the model every batch. The lock makes
-_CUDA_BROKEN = False
+# 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:
@@ -28,14 +147,6 @@ def _is_cuda_error(exc: BaseException) -> bool:
     )
 
 
-def _check_cuda_usable(device: str) -> None:
-    if device == "cuda" and _CUDA_BROKEN:
-        raise RuntimeError(
-            "CUDA context is broken (previous OOM was unrecoverable). "
-            "Restart the app or switch device to CPU."
-        )
-
-
 def _resolve_device(device: str) -> str:
     if device in ("cuda", "cpu"):
         return device
@@ -68,6 +179,59 @@ def _free_vram() -> None:
         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 = {}
 
@@ -77,21 +241,27 @@ 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)
-        if key not in cls._cache:
+        with _MODEL_LOCK:
-            import whisperx
+            if key not in cls._cache:
-            print(f"  [ASR] loading WhisperX {model_name} on {device} "
+                if device == "cuda":
-                  f"({compute_type}, local_files_only={local_files_only})...")
+                    _enable_tf32()
+                import whisperx
+                print(f"  [ASR] loading WhisperX {model_name} on {device} "
+                      f"({compute_type}, local_files_only={local_files_only})...")
 
-            def load():
+                def load():
-                return whisperx.load_model(
+                    return whisperx.load_model(
-                    model_name, device,
+                        model_name, device,
-                    compute_type=compute_type,
+                        compute_type=compute_type,
-                    local_files_only=local_files_only,
+                        local_files_only=local_files_only,
-                )
+                    )
 
-            cls._cache[key] = _retry(load)
+                cls._cache[key] = _retry(load)
-            print("  [ASR] model loaded.")
+                print("  [ASR] model loaded once — subsequent files will reuse it.")
-        return cls._cache[key]
+            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:
@@ -100,20 +270,24 @@ class _AlignModels:
     @classmethod
     def get(cls, language_code: str, device: str, model_name: Optional[str] = None):
         key = (language_code, device, model_name or "")
-        if key not in cls._cache:
+        with _MODEL_LOCK:
-            import whisperx
+            if key not in cls._cache:
-            print(f"  [ALIGN] loading align model for {language_code} "
+                import whisperx
-                  f"({model_name or 'default'}) on {device}...")
+                print(f"  [ALIGN] loading align model for {language_code} "
+                      f"({model_name or 'default'}) on {device}...")
 
-            def load():
+                def load():
-                kwargs = {"language_code": language_code, "device": device}
+                    kwargs = {"language_code": language_code, "device": device}
-                if model_name:
+                    if model_name:
-                    kwargs["model_name"] = model_name
+                        kwargs["model_name"] = model_name
-                return whisperx.load_align_model(**kwargs)
+                    return whisperx.load_align_model(**kwargs)
 
-            cls._cache[key] = _retry(load)
+                cls._cache[key] = _retry(load)
-            print("  [ALIGN] align model loaded.")
+                print("  [ALIGN] align model loaded once — reusing it.")
-        return cls._cache[key]
+            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]
 
 
 class _GigaAMModels:
@@ -127,14 +301,37 @@ class _GigaAMModels:
             "auto": ["CUDAExecutionProvider", "CPUExecutionProvider"],
         }.get(provider, ["CPUExecutionProvider"])
         key = tuple(providers)
-        if key not in cls._cache:
+        with _MODEL_LOCK:
-            import onnx_asr
+            if key not in cls._cache:
+                import onnx_asr
 
-            def load():
+                def load():
-                return onnx_asr.load_model("gigaam-v3-ctc", providers=list(providers))
+                    return onnx_asr.load_model("gigaam-v3-ctc", providers=list(providers))
 
-            cls._cache[key] = _retry(load)
+                cls._cache[key] = _retry(load)
-        return cls._cache[key]
+            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(
@@ -149,37 +346,13 @@ def transcribe(
     local_files_only: bool = False,
     vad_filter: bool = True,
 ) -> str:
-    global _CUDA_BROKEN
     device = _resolve_device(device)
-    _check_cuda_usable(device)
 
     if engine == "whisperx":
-        try:
+        return _transcribe_whisperx(
-            return _transcribe_whisperx(
+            audio_path, model, language, device, compute_type,
-                audio_path, model, language, device, compute_type,
+            batch_size, local_files_only, vad_filter,
-                batch_size, local_files_only, vad_filter,
+        )
-            )
-        except Exception as e:
-            if not _is_cuda_error(e):
-                raise
-            print(f"  [ASR] CUDA error: {e}. Clearing model cache and retrying...")
-            _WhisperXModels._cache.clear()
-            _free_vram()
-            try:
-                return _transcribe_whisperx(
-                    audio_path, model, language, device, compute_type,
-                    batch_size, local_files_only, vad_filter,
-                )
-            except Exception as retry_error:
-                if not _is_cuda_error(retry_error):
-                    raise
-                _CUDA_BROKEN = True
-                print("  [ASR] CUDA context is unrecoverable — "
-                      "GPU disabled for the rest of this run.")
-                raise RuntimeError(
-                    "CUDA context is broken after OOM recovery failed. "
-                    "Restart the app or switch device to CPU."
-                ) from retry_error
 
     if engine == "gigaam":
         return _transcribe_gigaam(audio_path, gigaam_provider)
@@ -199,8 +372,24 @@ 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(
-    text = " ".join(seg["text"].strip() for seg in result["segments"]).strip()
+        model, audio, language, batch_size, name=audio_path.name,
+    )
+
+    # 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
@@ -220,6 +409,7 @@ def transcribe_aligned(
     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).
 
@@ -228,10 +418,11 @@ def transcribe_aligned(
     """
     import whisperx
     device = _resolve_device(device)
-    _check_cuda_usable(device)
     model_obj = _WhisperXModels.get(model, device, compute_type, local_files_only)
     audio = whisperx.load_audio(str(audio_path))
-    result = model_obj.transcribe(audio, batch_size=batch_size, language=language)
+    result = _model_transcribe_with_retry(
+        model_obj, audio, language, batch_size, name=audio_path.name,
+    )
 
     detected = result.get("language") or language or "en"
 
@@ -246,6 +437,47 @@ def transcribe_aligned(
         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

@@ -2,6 +2,7 @@
 from __future__ import annotations
 
 import shutil
+import subprocess
 from pathlib import Path
 from typing import Optional
 
@@ -42,17 +43,30 @@ MODE_LABELS = {
 
 
 def _detect_device_info() -> tuple[str, str]:
+    """Report the GPU via nvidia-smi.
+
+    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:
-        import torch
+        out = subprocess.run(
-    except ImportError:
+            ["nvidia-smi", "--query-gpu=name,memory.total",
-        return "⚠ PyTorch не установлен", "#f38ba8"
+             "--format=csv,noheader,nounits"],
-    if torch.cuda.is_available():
+            capture_output=True, text=True, timeout=5, check=True,
-        try:
+        ).stdout.strip()
-            name = torch.cuda.get_device_name(0)
+    except (OSError, subprocess.SubprocessError):
-            vram = torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)
+        out = ""
-            return f"✓ CUDA: {name} · {vram:.1f} GB VRAM", "#a6e3a1"
+    if out:
-        except Exception:  # noqa: BLE001
+        parts = [p.strip() for p in out.splitlines()[0].split(",")]
-            return "✓ CUDA доступна", "#a6e3a1"
+        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 ValueError:
+                pass
+        return f"✓ CUDA: {name}", "#a6e3a1"
     return "⚠ CUDA не обнаружена — расчёт пойдёт на CPU (медленно)", "#f9e2af"
 
 
@@ -137,7 +151,7 @@ class MainWindow(QMainWindow):
         resume_opts = QHBoxLayout()
         self.skip_existing_cb = QCheckBox("Пропускать готовые (text.txt)")
         self.skip_existing_cb.setToolTip(
-            "Не трогать файлы, у которых в <output>/<stem>/ уже есть text.txt.\n"
+            "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."
         )
@@ -205,7 +219,9 @@ class MainWindow(QMainWindow):
         self.log_view = QPlainTextEdit()
         self.log_view.setObjectName("log")
         self.log_view.setReadOnly(True)
-        self.log_view.setMinimumHeight(240)
+        # 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)
 
@@ -352,7 +368,7 @@ class MainWindow(QMainWindow):
         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>/<stem>/vocals.wav — полезно для LRC-этапа."
+            "в <output>/<relpath>/<stem>/vocals.wav — полезно для LRC-этапа."
         )
         form.addRow("", self.separate_save_vocals_cb)
         return w
@@ -396,7 +412,7 @@ class MainWindow(QMainWindow):
         form.setSpacing(10)
 
         info = QLabel(
-            "1) Сгенерировать LRC: WhisperX + align → <output>/<stem>/text.lrc.\n"
+            "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."
         )
@@ -417,7 +433,7 @@ class MainWindow(QMainWindow):
         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>/<stem>/text.lrc уже существует и содержит корректный LRC,\n"
+            "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."
         )
@@ -447,7 +463,7 @@ class MainWindow(QMainWindow):
 
         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>/<stem>/<name>.\n"
+            "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)
@@ -692,6 +708,14 @@ 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",
@@ -704,6 +728,12 @@ class MainWindow(QMainWindow):
                 "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",

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