File size: 18,357 Bytes
287f01b 05bd568 287f01b bed3a49 287f01b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 | # audio_tools.py (ASR delegated to remote HF Space "veureu/asr")
# -----------------------------------------------------------------------------
# Veureu — AUDIO utilities (orchestrator w/ remote ASR)
# - FFmpeg extraction (WAV)
# - Diarization (pyannote) [local]
# - Voice embeddings (SpeechBrain ECAPA) [local]
# - Speaker identification (KMeans + ChromaDB optional) [local]
# - ASR: delegated to HF Space `veureu/asr` (faster-whisper-large-v3-ca-3catparla)
# - SRT generation
# - Orchestrator: process_audio_for_video(...)
# -----------------------------------------------------------------------------
from __future__ import annotations
import json
import logging
import math
import os
import shlex
import subprocess
from pathlib import Path
from typing import List, Dict, Any, Tuple, Optional
import numpy as np
# Optional torchaudio for I/O and resampling (fallback to soundfile+librosa otherwise)
try:
import torch
import torchaudio as ta
import torchaudio.transforms as T
HAS_TORCHAUDIO = True
# Note: ta.set_audio_backend is deprecated in newer torchaudio versions
except Exception:
HAS_TORCHAUDIO = False
ta = None # type: ignore
import soundfile as sf
# Pyannote for diarization (local)
from pyannote.audio import Pipeline
# Speaker embeddings (local)
from speechbrain.inference.speaker import SpeakerRecognition # v1.0+
# Clustering
from sklearn.cluster import KMeans
from sklearn.metrics import silhouette_score
# Router to remote Spaces (asr)
from llm_router import load_yaml, LLMRouter
# -------------------------------- Logging ------------------------------------
log = logging.getLogger("audio_tools")
if not log.handlers:
_h = logging.StreamHandler()
_h.setFormatter(logging.Formatter("[%(levelname)s] %(message)s"))
log.addHandler(_h)
log.setLevel(logging.INFO)
# ------------------------------- Utilities -----------------------------------
def load_wav(path: str | Path, sr: int = 16000):
"""Load audio as mono float32 at the requested sample rate."""
if HAS_TORCHAUDIO:
wav, in_sr = ta.load(str(path))
if in_sr != sr:
wav = ta.functional.resample(wav, in_sr, sr)
if wav.dim() > 1:
wav = wav.mean(dim=0, keepdim=True)
return wav.squeeze(0).numpy(), sr
import librosa
y, in_sr = sf.read(str(path), dtype="float32", always_2d=False)
if y.ndim > 1:
y = y.mean(axis=1)
if in_sr != sr:
y = librosa.resample(y, orig_sr=in_sr, target_sr=sr)
return y.astype(np.float32), sr
def save_wav(path: str | Path, y, sr: int = 16000):
"""Save mono float32 wav."""
if HAS_TORCHAUDIO:
import torch
wav = torch.from_numpy(np.asarray(y, dtype=np.float32)).unsqueeze(0)
ta.save(str(path), wav, sr)
else:
sf.write(str(path), np.asarray(y, dtype=np.float32), sr)
def extract_audio_ffmpeg(
video_path: str,
audio_out: Path,
sr: int = 16000,
mono: bool = True,
) -> str:
"""Extract audio from video to WAV using ffmpeg."""
audio_out.parent.mkdir(parents=True, exist_ok=True)
cmd = f'ffmpeg -y -i "{video_path}" -vn {"-ac 1" if mono else ""} -ar {sr} -f wav "{audio_out}"'
subprocess.run(
shlex.split(cmd),
check=True,
stdout=subprocess.DEVNULL,
stderr=subprocess.DEVNULL,
)
return str(audio_out)
# ----------------------------------- ASR (REMOTE) -------------------------------------
def transcribe_audio_remote(audio_path: str | Path, cfg: Dict[str, Any]) -> Dict[str, Any]:
"""
Send the audio file to the remote ASR Space `veureu/asr` (Gradio or HTTP).
The remote model is 'faster-whisper-large-v3-ca-3catparla' (Aina).
Returns standardized dict: {'text': str, 'segments': list?}
"""
if not cfg:
cfg = load_yaml("config.yaml")
router = LLMRouter(cfg)
model_name = (cfg.get("models", {}).get("asr") or "whisper-catalan")
params = {
"language": "ca",
"model": "faster-whisper-large-v3-ca-3catparla",
"timestamps": True,
"diarization": False, # diarization stays local
}
result = router.asr_transcribe(str(audio_path), model=model_name, **params)
if isinstance(result, str):
return {"text": result, "segments": []}
if isinstance(result, dict):
if "text" not in result and "transcription" in result:
result["text"] = result["transcription"]
result.setdefault("segments", [])
return result
return {"text": str(result), "segments": []}
# -------------------------------- Diarization --------------------------------
def diarize_audio(
wav_path: str,
base_dir: Path,
clips_folder: str = "clips",
min_segment_duration: float = 20.0,
max_segment_duration: float = 50.0,
hf_token_env: str | None = None,
) -> Tuple[List[str], List[Dict[str, Any]]]:
"""Diarization with pyannote and clip export with pydub."""
from pydub import AudioSegment
audio = AudioSegment.from_wav(wav_path)
duration = len(audio) / 1000.0
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-3.1",
use_auth_token=(hf_token_env or os.getenv("HF_TOKEN"))
)
diarization = pipeline(wav_path)
clips_dir = (base_dir / clips_folder)
clips_dir.mkdir(parents=True, exist_ok=True)
clip_paths: List[str] = []
segments: List[Dict[str, Any]] = []
spk_map: Dict[str, int] = {}
prev_end = 0.0
for i, (turn, _, speaker) in enumerate(diarization.itertracks(yield_label=True)):
start, end = max(0.0, float(turn.start)), min(duration, float(turn.end))
if start < prev_end:
start = prev_end
if end <= start:
continue
seg_dur = end - start
if seg_dur < min_segment_duration:
continue
if seg_dur > max_segment_duration:
n = int(math.ceil(seg_dur / max_segment_duration))
sub_d = seg_dur / n
for j in range(n):
s = start + j * sub_d
e = min(end, start + (j + 1) * sub_d)
if e <= s:
continue
clip = audio[int(s * 1000):int(e * 1000)]
cp = clips_dir / f"segment_{i:03d}_{j:02d}.wav"
clip.export(cp, format="wav")
if speaker not in spk_map:
spk_map[speaker] = len(spk_map)
segments.append({"start": s, "end": e, "speaker": f"SPEAKER_{spk_map[speaker]:02d}"})
clip_paths.append(str(cp))
prev_end = e
else:
clip = audio[int(start * 1000):int(end * 1000)]
cp = clips_dir / f"segment_{i:03d}.wav"
clip.export(cp, format="wav")
if speaker not in spk_map:
spk_map[speaker] = len(spk_map)
segments.append({"start": start, "end": end, "speaker": f"SPEAKER_{spk_map[speaker]:02d}"})
clip_paths.append(str(cp))
prev_end = end
if not segments:
cp = clips_dir / "segment_000.wav"
audio.export(cp, format="wav")
return [str(cp)], [{"start": 0.0, "end": duration, "speaker": "SPEAKER_00"}]
pairs = sorted(zip(clip_paths, segments), key=lambda x: x[1]["start"])
clip_paths, segments = [p[0] for p in pairs], [p[1] for p in pairs]
return clip_paths, segments
# ------------------------------ Voice embeddings -----------------------------
class VoiceEmbedder:
def __init__(self):
self.model = SpeakerRecognition.from_hparams(
source="speechbrain/spkrec-ecapa-voxceleb",
savedir="pretrained_models/spkrec-ecapa-voxceleb",
)
self.model.eval()
def embed(self, wav_path: str) -> List[float]:
if HAS_TORCHAUDIO:
waveform, sr = ta.load(wav_path)
target_sr = 16000
if sr != target_sr:
waveform = T.Resample(orig_freq=sr, new_freq=target_sr)(waveform)
if waveform.shape[0] > 1:
waveform = waveform.mean(dim=0, keepdim=True)
min_samples = int(0.2 * target_sr)
if waveform.shape[1] < min_samples:
pad = min_samples - waveform.shape[1]
import torch
waveform = torch.cat([waveform, torch.zeros((1, pad))], dim=1)
with torch.no_grad(): # type: ignore
emb = self.model.encode_batch(waveform).squeeze().cpu().numpy().astype(float)
return emb.tolist()
else:
y, sr = load_wav(wav_path, sr=16000)
min_len = int(0.2 * 16000)
if len(y) < min_len:
y = np.pad(y, (0, min_len - len(y)))
import torch
w = torch.from_numpy(y).unsqueeze(0).unsqueeze(0)
with torch.no_grad(): # type: ignore
emb = self.model.encode_batch(w).squeeze().cpu().numpy().astype(float)
return emb.tolist()
def embed_voice_segments(clip_paths: List[str]) -> List[List[float]]:
ve = VoiceEmbedder()
out: List[List[float]] = []
for cp in clip_paths:
try:
out.append(ve.embed(cp))
except Exception as e:
log.warning(f"Embedding error in {cp}: {e}")
out.append([])
return out
# --------------------------- Speaker identification --------------------------
def identify_speakers(
embeddings: List[List[float]],
voice_collection,
cfg: Dict[str, Any],
) -> List[str]:
voice_cfg = cfg.get("voice_processing", {}).get("speaker_identification", {})
if not embeddings or sum(1 for e in embeddings if e) < 2:
return ["SPEAKER_00" for _ in embeddings]
valid = [e for e in embeddings if e and len(e) > 0]
if len(valid) < 2:
return ["SPEAKER_00" for _ in embeddings]
min_clusters = max(1, int(voice_cfg.get("min_speakers", 1)))
max_clusters = min(int(voice_cfg.get("max_speakers", 5)), len(valid) - 1)
if voice_cfg.get("find_optimal_clusters", True) and len(valid) > 2:
best_score, best_k = -1.0, min_clusters
for k in range(min_clusters, max_clusters + 1):
if k >= len(valid):
break
km = KMeans(n_clusters=k, random_state=42, n_init="auto")
labels = km.fit_predict(valid)
if len(set(labels)) > 1:
score = silhouette_score(valid, labels)
if score > best_score:
best_score, best_k = score, k
else:
best_k = min(max_clusters, max(min_clusters, int(voice_cfg.get("num_speakers", 2))))
best_k = max(1, min(best_k, len(valid) - 1))
km = KMeans(n_clusters=best_k, random_state=42, n_init="auto", init="k-means++")
labels = km.fit_predict(np.array(valid))
centers = km.cluster_centers_
cluster_to_name: Dict[int, str] = {}
unknown_counter = 0
for cid in range(best_k):
center = centers[cid].tolist()
name = f"SPEAKER_{cid:02d}"
if voice_collection is not None:
try:
q = voice_collection.query(query_embeddings=[center], n_results=1)
metas = q.get("metadatas", [[]])[0]
dists = q.get("distances", [[]])[0]
thr = voice_cfg.get("distance_threshold")
if dists and thr is not None and dists[0] > thr:
name = f"UNKNOWN_{unknown_counter}"
unknown_counter += 1
voice_collection.add(
embeddings=[center],
metadatas=[{"name": name}],
ids=[f"unk_{cid}_{unknown_counter}"],
)
else:
if metas and isinstance(metas[0], dict):
name = metas[0].get("nombre") or metas[0].get("name") \
or metas[0].get("speaker") or metas[0].get("identity") or name
except Exception as e:
log.warning(f"Voice KNN query failed: {e}")
cluster_to_name[cid] = name
personas: List[str] = []
vi = 0
for emb in embeddings:
if not emb:
personas.append("UNKNOWN")
else:
label = int(labels[vi])
personas.append(cluster_to_name.get(label, f"SPEAKER_{label:02d}"))
vi += 1
return personas
# ----------------------------------- SRT -------------------------------------
def _fmt_srt_time(seconds: float) -> str:
h = int(seconds // 3600)
m = int((seconds % 3600) // 60)
s = int(seconds % 60)
ms = int(round((seconds - int(seconds)) * 1000))
return f"{h:02}:{m:02}:{s:02},{ms:03}"
def generate_srt_from_diarization(
diarization_segments: List[Dict[str, Any]],
transcriptions: List[str],
speakers_per_segment: List[str],
output_srt_path: str,
cfg: Dict[str, Any],
) -> None:
subs = cfg.get("subtitles", {})
max_cpl = int(subs.get("max_chars_per_line", 42))
max_lines = int(subs.get("max_lines_per_cue", 10))
speaker_display = subs.get("speaker_display", "brackets")
items: List[Dict[str, Any]] = []
n = min(len(diarization_segments), len(transcriptions), len(speakers_per_segment))
for i in range(n):
seg = diarization_segments[i]
text = (transcriptions[i] or "").strip()
spk = speakers_per_segment[i]
items.append({"start": float(seg.get("start", 0.0)), "end": float(seg.get("end", 0.0)), "text": text, "speaker": spk})
out = Path(output_srt_path)
out.parent.mkdir(parents=True, exist_ok=True)
with out.open("w", encoding="utf-8-sig") as f:
for i, it in enumerate(items, 1):
text = it["text"]
spk = it["speaker"]
if speaker_display == "brackets" and spk:
text = f"[{spk}]: {text}"
elif speaker_display == "prefix" and spk:
text = f"{spk}: {text}"
words = text.split()
lines: List[str] = []
cur = ""
for w in words:
if len(cur) + len(w) + (1 if cur else 0) <= max_cpl:
cur = (cur + " " + w) if cur else w
else:
lines.append(cur)
cur = w
if len(lines) >= max_lines - 1:
break
if cur and len(lines) < max_lines:
lines.append(cur)
f.write(f"{i}\n{_fmt_srt_time(it['start'])} --> {_fmt_srt_time(it['end'])}\n")
f.write("\n".join(lines) + "\n\n")
# ------------------------------ Orchestrator ---------------------------------
def process_audio_for_video(
video_path: str,
out_dir: Path,
cfg: Dict[str, Any],
voice_collection=None,
) -> Tuple[List[Dict[str, Any]], Optional[str], str]:
"""
Audio pipeline: FFmpeg -> diarization -> remote ASR (full + clips) -> embeddings -> speaker-ID -> SRT.
Returns (audio_segments, srt_path or None, full_transcription_text).
"""
audio_cfg = cfg.get("audio_processing", {})
sr = int(audio_cfg.get("sample_rate", 16000))
fmt = audio_cfg.get("format", "wav")
wav_path = extract_audio_ffmpeg(video_path, out_dir / f"{Path(video_path).stem}.{fmt}", sr=sr)
log.info("Audio extraído")
diar_cfg = audio_cfg.get("diarization", {})
min_dur = float(diar_cfg.get("min_segment_duration", 20.0))
max_dur = float(diar_cfg.get("max_segment_duration", 50.0))
clip_paths, diar_segs = diarize_audio(wav_path, out_dir, "clips", min_dur, max_dur)
log.info("Clips de audio generados.")
full_transcription = ""
asr_section = cfg.get("asr", {})
if asr_section.get("enable_full_transcription", True):
log.info("Transcripción completa (remota, Space 'asr')...")
full_res = transcribe_audio_remote(wav_path, cfg)
full_transcription = full_res.get("text", "") or ""
log.info("Transcripción completa finalizada.")
log.info("Transcripción por clip (remota, Space 'asr')...")
trans: List[str] = []
for cp in clip_paths:
res = transcribe_audio_remote(cp, cfg)
trans.append(res.get("text", ""))
log.info("Se han transcrito todos los clips.")
embeddings = embed_voice_segments(clip_paths) if audio_cfg.get("enable_voice_embeddings", True) else [[] for _ in clip_paths]
if cfg.get("voice_processing", {}).get("speaker_identification", {}).get("enabled", True):
speakers = identify_speakers(embeddings, voice_collection, cfg)
log.info("Speakers identificados correctamente.")
else:
speakers = [seg.get("speaker", f"SPEAKER_{i:02d}") for i, seg in enumerate(diar_segs)]
audio_segments: List[Dict[str, Any]] = []
for i, seg in enumerate(diar_segs):
audio_segments.append(
{
"segment": i,
"start": float(seg.get("start", 0.0)),
"end": float(seg.get("end", 0.0)),
"speaker": speakers[i] if i < len(speakers) else seg.get("speaker", f"SPEAKER_{i:02d}"),
"text": trans[i] if i < len(trans) else "",
"voice_embedding": embeddings[i],
"clip_path": str(out_dir / "clips" / f"segment_{i:03d}.wav"),
"lang": "ca",
"lang_prob": 1.0,
}
)
srt_base_path = out_dir / f"transcripcion_diarizada_{Path(video_path).stem}"
srt_unmodified_path = str(srt_base_path) + "_unmodified.srt"
try:
generate_srt_from_diarization(
diar_segs,
[a["text"] for a in audio_segments],
[a["speaker"] for a in audio_segments],
srt_unmodified_path,
cfg,
)
except Exception as e:
log.warning(f"SRT generation failed: {e}")
srt_unmodified_path = None
return audio_segments, srt_unmodified_path, full_transcription
|