jads/AudioTranscriberCombine
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix some bugs, and fix models

Browse Source
This commit is contained in:
jadssss
2026-10-24 08:39:33 +03:00
parent 78636699dc
commit fe338e6880
5 changed files with 355 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

195
combined_files.txt
View File

@@ -113,19 +113,14 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
# Файл: core/separate.py
# =========================================
"""Vocal separation via unblend (Mel-Band RoFormer).
Two entry points:
- separate_vocals() single file, subprocess per call (sequential mode)
- separate_vocals_batch() many files in ONE subprocess, model loaded once
(staged mode — big speedup on batches)
"""
"""Vocal separation via unblend (Mel-Band RoFormer)."""
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:
@@ -150,7 +145,6 @@ def separate_vocals(
custom_kernels: bool = False,
device: str = "cuda",
) -> Path:
"""Separate one file. Returns path to the isolated stem WAV."""
work_dir.mkdir(parents=True, exist_ok=True)
output_pattern = work_dir / "{track}" / "{stem}.wav"
cmd = [
@@ -179,14 +173,13 @@ def separate_vocals_batch(
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.
Input WAVs are expected to live in distinct directories
(e.g. <work_root>/<stem>/convert.wav). The output pattern uses {parent}
so each result lands next to its input as <parent>/vocals.wav.
Returns: {input_path: vocal_path}
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 {}
@@ -203,8 +196,27 @@ def separate_vocals_batch(
"--device", device,
]
result = subprocess.run(cmd, capture_output=True, text=True, env=_build_env(custom_kernels))
_check_result(result, f"{len(input_wavs)} files")
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:
if progress_callback and processed < total - 1:
if proc.wait(2.0) is not None:
break
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] = []
@@ -216,7 +228,6 @@ def separate_vocals_batch(
missing.append(src)
if missing:
# Fallback: search anywhere under work_root
found = sorted(work_root.rglob(f"{isolate_stem}.wav"))
if len(found) == len(input_wavs):
for src, dst in zip(input_wavs, found):
@@ -360,6 +371,7 @@ import json
import shutil
import tempfile
import threading
from concurrent.futures import ThreadPoolExecutor, as_completed
import re
from dataclasses import dataclass, field
from pathlib import Path
@@ -653,6 +665,129 @@ def _persist_artifacts(
# ---------------------------------------------------------------- entry
def _run_identification_parallel(
config: PipelineConfig,
files: list[Path],
log: LogFn,
on_progress: Optional[ProgressFn],
stop_event: Optional[threading.Event],
work_root: Path,
) -> dict[Path, Optional[IdentifyResult]]:
"""Parallel identification pass (tags → LRCLIB, Shazam, etc.)."""
if not (config.shazam.enabled or config.lyrics.enabled):
return {}
concurrency = max(1, int(config.shazam.identify_concurrency))
total = len(files)
if concurrency == 1 or total == 1:
results: dict[Path, Optional[IdentifyResult]] = {}
for i, f in enumerate(files, 1):
if stop_event is not None and stop_event.is_set():
break
if on_progress:
on_progress(i, total, "identify")
try:
results[f] = _identify_file(f, work_root / f.stem, config, log)
except Exception as e: # noqa: BLE001
log(f" ✗ {f.name} (identify): {e}")
results[f] = None
return results
log(f" Идентификация: {total} файлов, конкурентность {concurrency}")
def worker(f: Path) -> tuple[Path, Optional[IdentifyResult]]:
if stop_event is not None and stop_event.is_set():
return f, None
def prefixed(msg: str) -> None:
log(f" [{f.name}] {msg.strip()}")
try:
ident = _identify_file(f, work_root / f.stem, config, prefixed)
return f, ident
except Exception as e: # noqa: BLE001
log(f" ✗ {f.name} (identify): {e}")
return f, None
results: dict[Path, Optional[IdentifyResult]] = {}
with ThreadPoolExecutor(max_workers=concurrency) as pool:
futures = {pool.submit(worker, f): f for f in files}
done = 0
for fut in as_completed(futures):
f = futures[fut]
try:
_, ident = fut.result()
results[f] = ident
except Exception as e: # noqa: BLE001
log(f" ✗ {f.name} (identify): {e}")
results[f] = None
done += 1
if on_progress:
on_progress(done, total, "identify")
return results
def _run_convert_parallel(
config: PipelineConfig,
files: list[Path],
state: dict,
log: LogFn,
on_progress: Optional[ProgressFn],
stop_event: Optional[threading.Event],
) -> None:
"""Parallel convert stage. Writes state[f]['audio_path'] or state[f]['error']."""
concurrency = max(1, int(config.convert.convert_concurrency))
total = len(files)
def worker(f: Path) -> tuple[Path, Optional[Path], Optional[str]]:
if stop_event is not None and stop_event.is_set():
return f, None, None
try:
target = Path(tempfile.mkdtemp(prefix=f"at_{f.stem}_")) / "convert.wav"
if config.overwrite or not target.exists():
convert_to_wav(
f, target,
sample_rate=config.convert.sample_rate,
channels=config.convert.channels,
ffmpeg_path=config.convert.ffmpeg_path,
)
return f, target, None
except Exception as e: # noqa: BLE001
return f, None, f"{type(e).__name__}: {e}"
if concurrency == 1 or total == 1:
for i, f in enumerate(files, 1):
if stop_event is not None and stop_event.is_set():
break
if on_progress:
on_progress(i, total, "convert")
f2, target, err = worker(f)
if err:
state[f2]["error"] = err
log(f" ✗ {f2.name}: {err}")
else:
state[f2]["audio_path"] = target
log(f" ✓ {f2.name}")
return
log(f" Конвертация: {total} файлов, конкурентность {concurrency}")
with ThreadPoolExecutor(max_workers=concurrency) as pool:
futures = {pool.submit(worker, f): f for f in files}
done = 0
for fut in as_completed(futures):
f2, target, err = fut.result()
if err:
state[f2]["error"] = err
log(f" ✗ {f2.name}: {err}")
else:
state[f2]["audio_path"] = target
log(f" ✓ {f2.name}")
done += 1
if on_progress:
on_progress(done, total, "convert")
def run_pipeline(
config: PipelineConfig,
log: LogFn = print,
@@ -1476,6 +1611,7 @@ class ConvertConfig:
sample_rate: int = 44100
channels: int = 2
ffmpeg_path: str = "ffmpeg"
convert_concurrency: int = 4
@dataclass
@@ -1503,6 +1639,7 @@ class TranscribeConfig:
@dataclass
class ShazamConfig:
enabled: bool = False
identify_concurrency: int = 4
@dataclass
@@ -1513,7 +1650,7 @@ 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
@@ -1803,6 +1940,13 @@ class MainWindow(QMainWindow):
form.setSpacing(10)
self.shazam_enabled_cb = QCheckBox("Включён")
self.shazam_concurrency = QSpinBox()
self.shazam_concurrency.setRange(1, 16)
self.shazam_concurrency.setToolTip(
"Сколько файлов опознавать параллельно.\n"
"1 — последовательно, 4–8 — оптимум для Shazam."
)
form.addRow("Параллельных запросов:", self.shazam_concurrency)
form.addRow("", self.shazam_enabled_cb)
info = QLabel(
@@ -1874,6 +2018,13 @@ class MainWindow(QMainWindow):
form.addRow("Каналы:", self.convert_ch)
self.convert_ffmpeg = QLineEdit()
self.convert_concurrency = QSpinBox()
self.convert_concurrency.setRange(1, 16)
self.convert_concurrency.setToolTip(
"Сколько ffmpeg-процессов запускать одновременно.\n"
"Для SSD 4–8 нормально; для HDD лучше 1–2."
)
form.addRow("Параллельных конвертаций:", self.convert_concurrency)
form.addRow("ffmpeg:", self.convert_ffmpeg)
return w
@@ -1992,6 +2143,7 @@ class MainWindow(QMainWindow):
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)
@@ -2011,6 +2163,7 @@ 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)
@@ -2037,7 +2190,10 @@ class MainWindow(QMainWindow):
cfg.overwrite = self.overwrite_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(),
@@ -2054,6 +2210,7 @@ class MainWindow(QMainWindow):
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(),

4
core/config.py
View File

@@ -16,6 +16,7 @@ class ConvertConfig:
sample_rate: int = 44100
channels: int = 2
ffmpeg_path: str = "ffmpeg"
convert_concurrency: int = 4
@dataclass
@@ -43,6 +44,7 @@ class TranscribeConfig:
@dataclass
class ShazamConfig:
enabled: bool = False
identify_concurrency: int = 4
@dataclass
@@ -53,7 +55,7 @@ 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

129
core/pipeline.py
View File

@@ -19,6 +19,7 @@ import json
import shutil
import tempfile
import threading
from concurrent.futures import ThreadPoolExecutor, as_completed
import re
from dataclasses import dataclass, field
from pathlib import Path
@@ -312,6 +313,129 @@ def _persist_artifacts(
# ---------------------------------------------------------------- entry
def _run_identification_parallel(
config: PipelineConfig,
files: list[Path],
log: LogFn,
on_progress: Optional[ProgressFn],
stop_event: Optional[threading.Event],
work_root: Path,
) -> dict[Path, Optional[IdentifyResult]]:
"""Parallel identification pass (tags → LRCLIB, Shazam, etc.)."""
if not (config.shazam.enabled or config.lyrics.enabled):
return {}
concurrency = max(1, int(config.shazam.identify_concurrency))
total = len(files)
if concurrency == 1 or total == 1:
results: dict[Path, Optional[IdentifyResult]] = {}
for i, f in enumerate(files, 1):
if stop_event is not None and stop_event.is_set():
break
if on_progress:
on_progress(i, total, "identify")
try:
results[f] = _identify_file(f, work_root / f.stem, config, log)
except Exception as e: # noqa: BLE001
log(f" ✗ {f.name} (identify): {e}")
results[f] = None
return results
log(f" Идентификация: {total} файлов, конкурентность {concurrency}")
def worker(f: Path) -> tuple[Path, Optional[IdentifyResult]]:
if stop_event is not None and stop_event.is_set():
return f, None
def prefixed(msg: str) -> None:
log(f" [{f.name}] {msg.strip()}")
try:
ident = _identify_file(f, work_root / f.stem, config, prefixed)
return f, ident
except Exception as e: # noqa: BLE001
log(f" ✗ {f.name} (identify): {e}")
return f, None
results: dict[Path, Optional[IdentifyResult]] = {}
with ThreadPoolExecutor(max_workers=concurrency) as pool:
futures = {pool.submit(worker, f): f for f in files}
done = 0
for fut in as_completed(futures):
f = futures[fut]
try:
_, ident = fut.result()
results[f] = ident
except Exception as e: # noqa: BLE001
log(f" ✗ {f.name} (identify): {e}")
results[f] = None
done += 1
if on_progress:
on_progress(done, total, "identify")
return results
def _run_convert_parallel(
config: PipelineConfig,
files: list[Path],
state: dict,
log: LogFn,
on_progress: Optional[ProgressFn],
stop_event: Optional[threading.Event],
) -> None:
"""Parallel convert stage. Writes state[f]['audio_path'] or state[f]['error']."""
concurrency = max(1, int(config.convert.convert_concurrency))
total = len(files)
def worker(f: Path) -> tuple[Path, Optional[Path], Optional[str]]:
if stop_event is not None and stop_event.is_set():
return f, None, None
try:
target = Path(tempfile.mkdtemp(prefix=f"at_{f.stem}_")) / "convert.wav"
if config.overwrite or not target.exists():
convert_to_wav(
f, target,
sample_rate=config.convert.sample_rate,
channels=config.convert.channels,
ffmpeg_path=config.convert.ffmpeg_path,
)
return f, target, None
except Exception as e: # noqa: BLE001
return f, None, f"{type(e).__name__}: {e}"
if concurrency == 1 or total == 1:
for i, f in enumerate(files, 1):
if stop_event is not None and stop_event.is_set():
break
if on_progress:
on_progress(i, total, "convert")
f2, target, err = worker(f)
if err:
state[f2]["error"] = err
log(f" ✗ {f2.name}: {err}")
else:
state[f2]["audio_path"] = target
log(f" ✓ {f2.name}")
return
log(f" Конвертация: {total} файлов, конкурентность {concurrency}")
with ThreadPoolExecutor(max_workers=concurrency) as pool:
futures = {pool.submit(worker, f): f for f in files}
done = 0
for fut in as_completed(futures):
f2, target, err = fut.result()
if err:
state[f2]["error"] = err
log(f" ✗ {f2.name}: {err}")
else:
state[f2]["audio_path"] = target
log(f" ✓ {f2.name}")
done += 1
if on_progress:
on_progress(done, total, "convert")
def run_pipeline(
config: PipelineConfig,
log: LogFn = print,
@@ -668,8 +792,6 @@ def _run_staged(
if cancelled():
log("⏹ Отменено")
break
if on_progress:
on_progress(i, len(todo), "transcribe")
try:
text = transcribe(
state[f]["audio_path"],
@@ -688,6 +810,9 @@ def _run_staged(
except Exception as e: # noqa: BLE001
state[f]["error"] = f"{type(e).__name__}: {e}"
log(f" ✗ {f.name}: {e}")
finally:
if on_progress:
on_progress(i, len(todo), "transcribe")
# --- Stage 4: post-ASR lyrics match + persist ---
if not cancelled():

45
core/separate.py
View File

@@ -1,16 +1,11 @@
"""Vocal separation via unblend (Mel-Band RoFormer).
Two entry points:
- separate_vocals() single file, subprocess per call (sequential mode)
- separate_vocals_batch() many files in ONE subprocess, model loaded once
(staged mode — big speedup on batches)
"""
"""Vocal separation via unblend (Mel-Band RoFormer)."""
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:
@@ -35,7 +30,6 @@ def separate_vocals(
custom_kernels: bool = False,
device: str = "cuda",
) -> Path:
"""Separate one file. Returns path to the isolated stem WAV."""
work_dir.mkdir(parents=True, exist_ok=True)
output_pattern = work_dir / "{track}" / "{stem}.wav"
cmd = [
@@ -64,14 +58,13 @@ def separate_vocals_batch(
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.
Input WAVs are expected to live in distinct directories
(e.g. <work_root>/<stem>/convert.wav). The output pattern uses {parent}
so each result lands next to its input as <parent>/vocals.wav.
Returns: {input_path: vocal_path}
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 {}
@@ -88,8 +81,27 @@ def separate_vocals_batch(
"--device", device,
]
result = subprocess.run(cmd, capture_output=True, text=True, env=_build_env(custom_kernels))
_check_result(result, f"{len(input_wavs)} files")
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:
if progress_callback and processed < total - 1:
if proc.wait(2.0) is not None:
break
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] = []
@@ -101,7 +113,6 @@ def separate_vocals_batch(
missing.append(src)
if missing:
# Fallback: search anywhere under work_root
found = sorted(work_root.rglob(f"{isolate_stem}.wav"))
if len(found) == len(input_wavs):
for src, dst in zip(input_wavs, found):

22
gui/main_window.py
View File

@@ -182,6 +182,13 @@ class MainWindow(QMainWindow):
form.setSpacing(10)
self.shazam_enabled_cb = QCheckBox("Включён")
self.shazam_concurrency = QSpinBox()
self.shazam_concurrency.setRange(1, 16)
self.shazam_concurrency.setToolTip(
"Сколько файлов опознавать параллельно.\n"
"1 — последовательно, 4–8 — оптимум для Shazam."
)
form.addRow("Параллельных запросов:", self.shazam_concurrency)
form.addRow("", self.shazam_enabled_cb)
info = QLabel(
@@ -253,6 +260,13 @@ class MainWindow(QMainWindow):
form.addRow("Каналы:", self.convert_ch)
self.convert_ffmpeg = QLineEdit()
self.convert_concurrency = QSpinBox()
self.convert_concurrency.setRange(1, 16)
self.convert_concurrency.setToolTip(
"Сколько ffmpeg-процессов запускать одновременно.\n"
"Для SSD 4–8 нормально; для HDD лучше 1–2."
)
form.addRow("Параллельных конвертаций:", self.convert_concurrency)
form.addRow("ffmpeg:", self.convert_ffmpeg)
return w
@@ -371,6 +385,7 @@ class MainWindow(QMainWindow):
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)
@@ -390,6 +405,7 @@ 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)
@@ -416,7 +432,10 @@ class MainWindow(QMainWindow):
cfg.overwrite = self.overwrite_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(),
@@ -433,6 +452,7 @@ class MainWindow(QMainWindow):
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(),