File size: 32,897 Bytes
7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 67b64d0 7e5b4d7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 |
"""
ChatterboxTTS Audio Processing & Quality Control Module
======================================================
OVERVIEW:
This module provides comprehensive audio quality validation, enhancement, and
post-processing for TTS-generated audio. It ensures consistent quality across
audiobook chapters by detecting and handling common TTS artifacts.
MAIN COMPONENTS:
1. QUALITY VALIDATION: Detects clipping, silence, flatness, and other artifacts
2. HUM DETECTION: Identifies and flags TTS-generated audio hum using frequency analysis
3. AUDIO ENHANCEMENT: Normalization, trimming, and quality improvements
4. ASR VALIDATION: Optional speech recognition for quality verification
5. SILENCE INSERTION: Adds appropriate pauses based on punctuation boundaries
6. AUDIO HEALTH CHECKS: Comprehensive audio file validation
CRITICAL QUALITY FEATURES:
- TTS hum detection with configurable frequency thresholds
- Audio clipping detection and prevention
- Silence detection at beginning/end of chunks
- Flatness detection (monotone audio identification)
- ASR-based transcription accuracy validation
- Dynamic range and loudness assessment
WORKFLOW:
Raw TTS Audio β Quality Validation β Artifact Detection β
Enhancement Processing β Silence Insertion β Final Audio Output
TECHNICAL DETAILS:
- Supports multiple audio formats (WAV, MP3, FLAC)
- Configurable quality thresholds for different validation types
- Integration with Whisper ASR for transcription validation
- Memory-efficient processing for large audio files
- Detailed logging for quality control debugging
PERFORMANCE IMPACT:
Essential for maintaining consistent audiobook quality and preventing
distribution of low-quality audio with TTS artifacts or technical issues.
"""
import numpy as np
import soundfile as sf
import logging
import shutil
import re
import time
from pathlib import Path
from pydub import AudioSegment, silence
from config.config import *
# Enhanced imports for spectral analysis
try:
import librosa
LIBROSA_AVAILABLE = True
except ImportError:
LIBROSA_AVAILABLE = False
logging.warning("librosa not available - enhanced spectral analysis disabled")
# ============================================================================
# AUDIO QUALITY DETECTION
# ============================================================================
def check_audio_health(wav_path):
"""Enhanced audio health checking"""
data, samplerate = sf.read(str(wav_path))
if len(data.shape) > 1:
data = data[:, 0] # mono only
clipping = np.mean(np.abs(data) > 0.98)
silence_ratio = np.mean(np.abs(data) < 1e-4)
rms = np.sqrt(np.mean(data**2))
mean_abs = np.mean(np.abs(data))
flatness = mean_abs / (rms + 1e-8)
return {
"clipping_ratio": round(clipping, 4),
"silence_ratio": round(silence_ratio, 4),
"flatness": round(flatness, 4),
}
def detect_tts_hum_artifact(wav_path):
"""
Detect low-frequency TTS confusion hum using configurable parameters
"""
if not ENABLE_HUM_DETECTION:
return False, {}
data, sr = sf.read(str(wav_path))
if data.ndim > 1:
data = data[:, 0] # Mono
# FFT analysis for frequency content
fft = np.fft.rfft(data)
freqs = np.fft.rfftfreq(len(data), 1/sr)
# Focus on hum frequency range (configurable at top of file)
hum_mask = (freqs >= HUM_FREQ_MIN) & (freqs <= HUM_FREQ_MAX)
hum_energy = np.sum(np.abs(fft[hum_mask]))
total_energy = np.sum(np.abs(fft))
# Check for sustained low-level amplitude (steady hum characteristic)
segment_size = sr // 4 # 250ms segments
segments = [data[i:i+segment_size] for i in range(0, len(data)-segment_size, segment_size)]
steady_segments = 0
for segment in segments:
rms = np.sqrt(np.mean(segment**2))
if HUM_AMPLITUDE_MIN < rms < HUM_AMPLITUDE_MAX:
steady_segments += 1
# Calculate hum indicators using configurable thresholds
hum_ratio = hum_energy / (total_energy + 1e-10)
steady_ratio = steady_segments / len(segments) if segments else 0
# Detection logic using configurable thresholds
has_hum = (hum_ratio > HUM_ENERGY_THRESHOLD) and (steady_ratio > HUM_STEADY_THRESHOLD)
if has_hum:
logging.info(f"π TTS hum detected: {wav_path.name}")
logging.info(f" Frequency range: {HUM_FREQ_MIN}-{HUM_FREQ_MAX}Hz")
logging.info(f" Hum energy ratio: {hum_ratio:.3f} (threshold: {HUM_ENERGY_THRESHOLD})")
logging.info(f" Steady segments: {steady_ratio:.3f} (threshold: {HUM_STEADY_THRESHOLD})")
return has_hum, {
"hum_ratio": hum_ratio,
"steady_ratio": steady_ratio,
"freq_range": f"{HUM_FREQ_MIN}-{HUM_FREQ_MAX}Hz"
}
def smart_audio_validation(wav_path):
"""Comprehensive audio validation with intelligent responses"""
# Standard health check
health = check_audio_health(wav_path)
# TTS hum detection (if enabled)
has_hum, hum_metrics = detect_tts_hum_artifact(wav_path)
# Decision matrix
if health["clipping_ratio"] > 0.05:
return handle_problematic_chunks(wav_path, "clipping", health)
elif health["flatness"] > 0.9:
return handle_problematic_chunks(wav_path, "corrupted", health)
elif has_hum:
return handle_problematic_chunks(wav_path, "tts_hum", hum_metrics)
else:
return wav_path # Passed all checks
def has_mid_energy_drop(wav_tensor, sr, window_ms=250, threshold_ratio=None):
"""Detect mid-chunk energy drops"""
wav = wav_tensor.squeeze().numpy()
win_samples = int(sr * window_ms / 1000)
segments = [wav[i:i+win_samples] for i in range(0, len(wav) - win_samples, win_samples)]
rms_vals = [np.sqrt(np.mean(seg**2)) for seg in segments]
rms_avg = np.mean(rms_vals)
dynamic_thresh = threshold_ratio or max(0.02, 0.1 if rms_avg < 0.01 else 0.2)
drop_sequence = 0
consecutive_required = 2
for i, rms in enumerate(rms_vals):
if i < 3:
continue
if rms < rms_avg * dynamic_thresh:
drop_sequence += 1
if drop_sequence >= consecutive_required:
return True
else:
drop_sequence = 0
return False
def detect_spectral_artifacts(audio_path_or_segment, use_mfcc=True):
"""
Enhanced spectral anomaly detection using MFCC analysis.
Args:
audio_path_or_segment: Path to audio file or AudioSegment object
use_mfcc: Whether to use MFCC-based analysis (requires librosa)
Returns:
float: Quality score (0.0-1.0, higher is better)
"""
try:
# Load audio data
if isinstance(audio_path_or_segment, (str, Path)):
y, sr = sf.read(str(audio_path_or_segment))
elif isinstance(audio_path_or_segment, AudioSegment):
# Convert AudioSegment to numpy array
samples = np.array(audio_path_or_segment.get_array_of_samples())
if audio_path_or_segment.channels == 2:
samples = samples.reshape((-1, 2)).mean(axis=1)
y = samples.astype(np.float32) / audio_path_or_segment.max_possible_amplitude
sr = audio_path_or_segment.frame_rate
else:
return 0.5 # Unknown format, neutral score
# Ensure mono
if len(y.shape) > 1:
y = y[:, 0]
# Basic energy-based anomaly detection (always available)
energy = np.abs(y)
energy_variance = np.var(energy)
# Simple threshold-based scoring
basic_score = 1.0 - min(energy_variance / 0.1, 1.0)
# Enhanced MFCC-based detection if librosa is available
if use_mfcc and LIBROSA_AVAILABLE and ENABLE_MFCC_VALIDATION:
try:
# Compute MFCC features
mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
# Calculate spectral variance across time
mfcc_variance = np.var(mfccs, axis=1)
max_variance_jump = np.max(np.abs(np.diff(mfcc_variance)))
# Normalize and score
mfcc_score = 1.0 - min(max_variance_jump / SPECTRAL_VARIANCE_LIMIT, 1.0)
# Combine scores (weighted average)
final_score = 0.6 * mfcc_score + 0.4 * basic_score
except Exception as e:
logging.debug(f"MFCC analysis failed: {e}")
final_score = basic_score
else:
final_score = basic_score
return max(0.0, min(1.0, final_score))
except Exception as e:
logging.error(f"Spectral artifact detection failed: {e}")
return 0.5 # Neutral score on failure
def evaluate_chunk_quality(audio_path_or_segment, reference_text=None, include_spectral=True, asr_model=None):
"""
Composite quality evaluation for a single audio chunk.
Acts as a clearinghouse - only runs individual checks when they are specifically enabled.
Args:
audio_path_or_segment: Path to audio file or AudioSegment object
reference_text: Original text for comparison (optional)
include_spectral: Whether to include spectral analysis
asr_model: Pre-loaded ASR model to avoid duplicate loading
Returns:
float: Composite quality score (0.0-1.0)
"""
# Skip all validation if output validation clearinghouse is disabled
if not ENABLE_OUTPUT_VALIDATION:
return 1.0 # Pass all chunks if validation is completely disabled
scores = []
# Spectral anomaly detection (only if MFCC validation is enabled)
if include_spectral and ENABLE_MFCC_VALIDATION:
spectral_score = detect_spectral_artifacts(audio_path_or_segment)
scores.append(spectral_score)
# ASR text validation (only if ASR is enabled AND reference text provided)
if reference_text and ENABLE_ASR:
text_validation_score = validate_output_matches_input(audio_path_or_segment, reference_text, asr_model)
scores.append(text_validation_score)
# Basic audio health (if it's a file path)
if isinstance(audio_path_or_segment, (str, Path)):
try:
health_result = check_audio_health(audio_path_or_segment)
# Convert health result to score (assuming False = good, True = bad)
health_score = 0.2 if health_result else 0.8
scores.append(health_score)
except Exception:
scores.append(0.5) # Neutral score on failure
# Return average of all scores
return sum(scores) / len(scores) if scores else 0.5
def validate_output_matches_input(audio_path_or_segment, reference_text, asr_model=None):
"""
Validate that TTS audio output matches the input text using ASR transcription.
Args:
audio_path_or_segment: Path to audio file or AudioSegment object
reference_text: Original input text that should have been synthesized
asr_model: Optional pre-loaded ASR model (will load whisper if None)
Returns:
float: Validation score (0.0-1.0, higher means better match)
"""
try:
# Convert AudioSegment to temporary file if needed
temp_file = None
if isinstance(audio_path_or_segment, AudioSegment):
import tempfile
temp_file = tempfile.NamedTemporaryFile(suffix='.wav', delete=False)
audio_path_or_segment.export(temp_file.name, format='wav')
audio_path = temp_file.name
else:
audio_path = str(audio_path_or_segment)
# Load ASR model if not provided
if asr_model is None:
try:
from modules.asr_manager import load_asr_model_adaptive
# Use adaptive manager for fallback ASR loading
asr_model, _ = load_asr_model_adaptive()
if asr_model is None:
logging.warning("ASR model loading failed in audio processor")
return 0.8 # Neutral score if ASR unavailable
except ImportError:
logging.warning("Whisper not available for output validation")
return 0.8 # Neutral score if ASR unavailable
# Transcribe the audio
result = asr_model.transcribe(audio_path)
transcribed_text = result.get("text", "").strip()
# Clean up temporary file
if temp_file:
import os
os.unlink(temp_file.name)
# Calculate text similarity using F1 score
similarity_score = calculate_text_similarity(reference_text, transcribed_text)
# Log significant mismatches for debugging
if similarity_score < OUTPUT_VALIDATION_THRESHOLD:
logging.warning(f"π Output validation failed (score: {similarity_score:.3f})")
logging.warning(f" Expected: {reference_text}")
logging.warning(f" Got: {transcribed_text}")
return similarity_score
except Exception as e:
logging.error(f"Output validation failed: {e}")
return 0.8 # Use neutral-good score to avoid regeneration on ASR errors
def calculate_text_similarity(text1, text2):
"""
Calculate similarity between two texts using word-level F1 score.
Args:
text1: Reference text
text2: Comparison text
Returns:
float: F1 similarity score (0.0-1.0)
"""
# Normalize texts (lowercase, remove punctuation, split into words)
import re
def normalize_text(text):
# Convert to lowercase and remove punctuation
text = re.sub(r'[^\w\s]', '', text.lower())
# Split into words and filter empty strings
return [word for word in text.split() if word]
words1 = set(normalize_text(text1))
words2 = set(normalize_text(text2))
if not words1 and not words2:
return 1.0 # Both empty
if not words1 or not words2:
return 0.0 # One empty, one not
# Calculate precision, recall, and F1
intersection = words1.intersection(words2)
precision = len(intersection) / len(words2) if words2 else 0
recall = len(intersection) / len(words1) if words1 else 0
if precision + recall == 0:
return 0.0
f1_score = 2 * (precision * recall) / (precision + recall)
return f1_score
def adjust_parameters_for_retry(params, quality_score, attempt_num):
"""
Adjust TTS parameters for regeneration attempts.
Args:
params: Current TTS parameters dictionary
quality_score: Quality score from previous attempt (0.0-1.0)
attempt_num: Current attempt number (0-based)
Returns:
dict: Adjusted parameters
"""
adjusted = params.copy()
# Adjustment strategy based on quality score and attempt number
if quality_score < 0.3:
# Very poor quality - more aggressive adjustments
temp_adj = REGEN_TEMPERATURE_ADJUSTMENT * 2
exag_adj = REGEN_EXAGGERATION_ADJUSTMENT * 2
cfg_adj = REGEN_CFG_ADJUSTMENT * 2
elif quality_score < 0.6:
# Moderate quality issues - standard adjustments
temp_adj = REGEN_TEMPERATURE_ADJUSTMENT
exag_adj = REGEN_EXAGGERATION_ADJUSTMENT
cfg_adj = REGEN_CFG_ADJUSTMENT
else:
# Minor quality issues - gentle adjustments
temp_adj = REGEN_TEMPERATURE_ADJUSTMENT * 0.5
exag_adj = REGEN_EXAGGERATION_ADJUSTMENT * 0.5
cfg_adj = REGEN_CFG_ADJUSTMENT * 0.5
# Apply adjustments based on attempt number
if attempt_num == 1:
# First retry: reduce temperature (less randomness)
adjusted['temperature'] = max(TTS_PARAM_MIN_TEMPERATURE,
adjusted['temperature'] - temp_adj)
elif attempt_num == 2:
# Second retry: adjust exaggeration (less emotion)
adjusted['exaggeration'] = max(TTS_PARAM_MIN_EXAGGERATION,
adjusted['exaggeration'] - exag_adj)
# Also increase cfg_weight (more faithful to text)
adjusted['cfg_weight'] = min(TTS_PARAM_MAX_CFG_WEIGHT,
adjusted['cfg_weight'] + cfg_adj)
return adjusted
# ============================================================================
# PROBLEMATIC CHUNK HANDLING
# ============================================================================
def handle_problematic_chunks(wav_path, issue_type, metrics):
"""Handle chunks with audio issues - quarantine for review"""
quarantine_dir = wav_path.parent / "quarantine"
quarantine_dir.mkdir(exist_ok=True)
# Move to quarantine with descriptive name
quarantine_path = quarantine_dir / f"{wav_path.stem}_{issue_type}.wav"
shutil.move(str(wav_path), str(quarantine_path))
# Log for user review
logging.warning(f"π¨ Quarantined {issue_type}: {wav_path.name} β {quarantine_path.name}")
logging.warning(f" Metrics: {metrics}")
return quarantine_path
def pause_for_chunk_review(quarantine_dir):
"""Pause processing to allow manual chunk review/editing with proper workflow"""
quarantined_files = list(quarantine_dir.glob("*.wav"))
if not quarantined_files:
return # No quarantined files, continue normally
print(f"\nβ οΈ {len(quarantined_files)} chunks quarantined in: {quarantine_dir}")
print("\nQuarantined chunks:")
for qfile in quarantined_files:
print(f" π {qfile.name}")
print("\nπ§ Options:")
print("1. Continue processing (use quarantined chunks as-is)")
print("2. Pause to manually review/edit chunks")
while True:
choice = input("\nEnter choice [1/2]: ").strip()
if choice in ['1', '2']:
break
print("β Invalid choice. Please enter 1 or 2.")
if choice == "2":
print(f"\nπ Processing paused for manual review.")
print(f"π Quarantined chunks are in: {quarantine_dir}")
print("\nπ Instructions:")
print(" 1. Edit the audio files in the quarantine folder")
print(" 2. Keep the original filenames (chunk numbering intact)")
print(" 3. Leave edited files IN the quarantine folder")
print(" 4. Press Enter below to continue processing")
input("\nβΈοΈ Press Enter when you've finished editing...")
# Verify files still exist after user editing
edited_files = list(quarantine_dir.glob("*.wav"))
if not edited_files:
print("β οΈ No files found in quarantine folder after editing!")
return
print(f"β
Found {len(edited_files)} edited files, continuing...")
# Move all chunks back to main audio folder (whether edited or not)
moved_count = 0
for qfile in quarantine_dir.glob("*.wav"):
# Extract original chunk name from quarantine filename - FIXED LINE:
original_name = re.sub(r'_(clipping|corrupted|tts_hum)$', '', qfile.stem) + ".wav"
main_path = qfile.parent.parent / original_name
try:
shutil.move(str(qfile), str(main_path))
moved_count += 1
print(f"β©οΈ Restored: {original_name}")
except Exception as e:
logging.error(f"β Failed to restore {qfile.name}: {e}")
print(f"\nβ
Restored {moved_count} chunks to main audio folder")
# Clean up empty quarantine directory
if not any(quarantine_dir.iterdir()):
quarantine_dir.rmdir()
return moved_count
# ============================================================================
# AUDIO EFFECTS AND PROCESSING
# ============================================================================
def detect_end_artifact(wav_path, window_ms=100):
"""Enhanced artifact detection"""
data, sr = sf.read(str(wav_path))
if data.ndim > 1:
data = data[:, 0]
win_samples = int(window_ms / 1000 * sr)
if len(data) < win_samples * 2:
return False
end = data[-win_samples:]
middle = data[len(data)//2 : len(data)//2 + win_samples]
rms_end = np.sqrt(np.mean(end**2))
rms_mid = np.sqrt(np.mean(middle**2)) + 1e-10
rms_ratio = rms_end / rms_mid
zcr = np.mean(np.diff(np.sign(end)) != 0)
fft = np.fft.rfft(end)
freqs = np.fft.rfftfreq(len(end), 1/sr)
low_band = fft[freqs < 150]
low_energy = np.sum(np.abs(low_band)) / (np.sum(np.abs(fft)) + 1e-10)
logging.info(f"{GREEN}[DEBUG]{RESET} Artifact metrics - {YELLOW}RMS ratio: {rms_ratio:.3f}{RESET}, "
f"{GREEN}ZCR: {zcr:.3f}{RESET}, {CYAN}LowEnergy: {low_energy:.3f}{RESET}")
return rms_ratio > 0.6 or zcr > 0.2 or low_energy > 0.4
def find_end_of_speech(wav_path, sr=16000):
"""Find end of speech using Silero VAD"""
import torch
import os
# Set environment variables to suppress PyTorch Hub verbosity
old_vars = {}
suppress_vars = {
'TORCH_HUB_VERBOSE': '0',
'PYTHONWARNINGS': 'ignore',
'TF_CPP_MIN_LOG_LEVEL': '3'
}
# Save old values and set new ones
for key, value in suppress_vars.items():
old_vars[key] = os.environ.get(key)
os.environ[key] = value
# Temporarily disable logging for this operation
old_level = logging.getLogger().level
logging.getLogger().setLevel(logging.ERROR)
try:
model, utils = torch.hub.load(
repo_or_dir='snakers4/silero-vad',
model='silero_vad',
force_reload=False,
verbose=False
)
(get_speech_timestamps, _, read_audio, _, _) = utils
wav = read_audio(str(wav_path), sampling_rate=sr)
speech_segments = get_speech_timestamps(wav, model, sampling_rate=sr)
if not speech_segments:
return None
last_seg_end = speech_segments[-1]['end']
return int(last_seg_end * 1000 / sr)
finally:
# Restore everything
logging.getLogger().setLevel(old_level)
for key, old_value in old_vars.items():
if old_value is None:
os.environ.pop(key, None)
else:
os.environ[key] = old_value
def fade_out_wav(wav_path, output_path=None, fade_ms=20):
"""Apply fade-out to audio"""
data, sr = sf.read(str(wav_path))
if data.ndim > 1:
data = data[:, 0]
fade_samples = int(sr * fade_ms / 1000)
if len(data) < fade_samples:
return
debug_path = wav_path.parent / f"{wav_path.stem}_pre_fade.wav"
sf.write(str(debug_path), data, sr)
fade_curve = np.linspace(1.0, 0.0, fade_samples)
data[-fade_samples:] *= fade_curve
sf.write(str(output_path or wav_path), data, sr)
def apply_smart_fade(wav_path):
"""Apply smart fade with artifact detection"""
eos_ms = find_end_of_speech(wav_path)
if detect_end_artifact(wav_path):
fade_out_wav(wav_path)
def apply_smart_fade_memory(audio_segment):
"""Apply smart fade with artifact detection - in memory version"""
# For now, apply a gentle fade to all audio to prevent clicks
# TODO: Add proper artifact detection for memory processing
return audio_segment.fade_out(50) # 50ms fade out
def smart_audio_validation_memory(audio_segment, sample_rate):
"""Enhanced audio validation in memory - returns (audio, is_quarantined)"""
# Basic validation - can be enhanced with hum detection later
# For now, just return the audio as-is
is_quarantined = False
# Could add memory-based hum detection here
# is_quarantined = detect_hum_memory(audio_segment, sample_rate)
return audio_segment, is_quarantined
def add_contextual_silence_memory(audio_segment, boundary_type):
"""Add appropriate silence based on content boundary type - in memory"""
from pydub import AudioSegment
from config.config import (
SILENCE_CHAPTER_START, SILENCE_CHAPTER_END, SILENCE_SECTION_BREAK, SILENCE_PARAGRAPH_END,
SILENCE_COMMA, SILENCE_SEMICOLON, SILENCE_COLON, SILENCE_PERIOD, SILENCE_QUESTION_MARK,
SILENCE_EXCLAMATION, SILENCE_DASH, SILENCE_ELLIPSIS, SILENCE_QUOTE_END
)
silence_durations = {
# Structural boundaries
"chapter_start": SILENCE_CHAPTER_START,
"chapter_end": SILENCE_CHAPTER_END,
"section_break": SILENCE_SECTION_BREAK,
"paragraph_end": SILENCE_PARAGRAPH_END,
# Punctuation boundaries
"comma": SILENCE_COMMA,
"semicolon": SILENCE_SEMICOLON,
"colon": SILENCE_COLON,
"period": SILENCE_PERIOD,
"question_mark": SILENCE_QUESTION_MARK,
"exclamation": SILENCE_EXCLAMATION,
"dash": SILENCE_DASH,
"ellipsis": SILENCE_ELLIPSIS,
"quote_end": SILENCE_QUOTE_END,
}
if boundary_type in silence_durations:
duration = silence_durations[boundary_type]
silence_segment = AudioSegment.silent(duration=duration)
return audio_segment + silence_segment
return audio_segment
def smart_fade_out(wav_path, silence_thresh_db=-40, min_silence_len=300):
"""Smart fade-out for natural audio endings"""
audio = AudioSegment.from_wav(wav_path)
tail_window_ms = 2000
if len(audio) < tail_window_ms:
logging.info(f"β οΈ {YELLOW}Skipping fade: {wav_path.name} too short ({len(audio)}ms < {tail_window_ms}ms){RESET}")
return
tail = audio[-tail_window_ms:]
silent_ranges = silence.detect_silence(tail, min_silence_len=min_silence_len, silence_thresh=silence_thresh_db)
min_tail_energy = max(tail.get_array_of_samples())
if not silent_ranges or min_tail_energy > audio.max_possible_amplitude * 0.1:
logging.info(f"β
{GREEN}No fade needed for {wav_path.name} (no valid trailing silence){RESET}")
return
fade_start_ms = silent_ranges[0][0]
fade_length_ms = tail_window_ms - fade_start_ms
if fade_length_ms < 100:
logging.info(f"β
{GREEN}No fade needed for {wav_path.name} (fade too short: {fade_length_ms}ms){RESET}")
return
fade_start_point = silent_ranges[0][0]
logging.info(f"β οΈ {RED}Fading tail of {wav_path.name} from {fade_start_point}ms to end{RESET}")
faded = audio[:fade_start_point] + audio[fade_start_point:].fade_out(duration=fade_length_ms)
faded.export(wav_path, format="wav")
# ============================================================================
# AUDIO TRIMMING
# ============================================================================
def trim_audio_endpoint(audio_segment, threshold=None, buffer_ms=None):
"""
Trim audio to the detected end of speech using RMS energy analysis.
Args:
audio_segment: pydub AudioSegment object
threshold: RMS threshold for speech detection (from config if None)
buffer_ms: Buffer to add after detected endpoint (from config if None)
Returns:
Trimmed AudioSegment
"""
if threshold is None:
threshold = SPEECH_ENDPOINT_THRESHOLD
if buffer_ms is None:
buffer_ms = TRIMMING_BUFFER_MS
# Convert to numpy array for analysis
samples = np.array(audio_segment.get_array_of_samples())
if audio_segment.channels == 2:
samples = samples.reshape((-1, 2)).mean(axis=1)
# Normalize samples
samples = samples.astype(np.float32) / audio_segment.max_possible_amplitude
# Calculate RMS in sliding windows (50ms windows)
window_size = int(0.05 * audio_segment.frame_rate) # 50ms
rms_values = []
for i in range(0, len(samples) - window_size, window_size // 2):
window = samples[i:i + window_size]
rms = np.sqrt(np.mean(window ** 2))
rms_values.append(rms)
# Find actual end of speech using energy decay detection
speech_end_idx = 0 # Default to beginning if no speech found
# Look for a significant and sustained drop in energy
# Scan backwards to find where energy consistently stays above a higher threshold
strong_speech_threshold = threshold * 3 # 3x threshold for "real" speech
for i in range(len(rms_values) - 1, -1, -1):
if rms_values[i] > strong_speech_threshold:
# Found strong speech, check if it's sustained
# Look forward to see if energy drops and stays low
sustained_speech = True
windows_ahead = min(10, len(rms_values) - i) # Look ahead up to 10 windows (250ms)
# Check if most of the next windows have reasonable speech levels
speech_count = 0
for j in range(i, min(i + windows_ahead, len(rms_values))):
if rms_values[j] > threshold:
speech_count += 1
# If this looks like the end of sustained speech content
if speech_count >= max(1, windows_ahead * 0.3): # At least 30% speech in next windows
speech_end_idx = i
break
# If no strong speech found, fall back to simple threshold method but be conservative
if speech_end_idx == 0:
for i in range(len(rms_values) - 1, -1, -1):
if rms_values[i] > threshold * 2: # Use 2x threshold for fallback
speech_end_idx = i
break
# Convert back to milliseconds and add buffer
# Convert window index to sample position, then to milliseconds
sample_position = speech_end_idx * (window_size // 2)
speech_end_ms = int(sample_position * 1000 / audio_segment.frame_rate)
trim_point_ms = min(speech_end_ms + buffer_ms, len(audio_segment))
return audio_segment[:trim_point_ms]
def process_audio_with_trimming_and_silence(audio_segment, boundary_type, enable_trimming=None):
"""
Complete audio processing: trim to speech endpoint + add punctuation-based silence.
Args:
audio_segment: pydub AudioSegment object
boundary_type: Boundary type from text processing
enable_trimming: Whether to trim audio (from config if None)
Returns:
Processed AudioSegment with trimming and appropriate silence
"""
if enable_trimming is None:
enable_trimming = ENABLE_AUDIO_TRIMMING
processed_audio = audio_segment
# Step 1: Trim to speech endpoint if enabled
if enable_trimming:
processed_audio = trim_audio_endpoint(processed_audio)
# Step 2: Add punctuation-appropriate silence
processed_audio = add_contextual_silence_memory(processed_audio, boundary_type)
return processed_audio
# ============================================================================
# SILENCE AND CONTEXTUAL AUDIO
# ============================================================================
def add_contextual_silence(wav_path, boundary_type):
"""Add appropriate silence based on content boundary type"""
silence_durations = {
# Structural boundaries
"chapter_start": SILENCE_CHAPTER_START,
"chapter_end": SILENCE_CHAPTER_END,
"section_break": SILENCE_SECTION_BREAK,
"paragraph_end": SILENCE_PARAGRAPH_END,
# Punctuation boundaries
"comma": SILENCE_COMMA,
"semicolon": SILENCE_SEMICOLON,
"colon": SILENCE_COLON,
"period": SILENCE_PERIOD,
"question_mark": SILENCE_QUESTION_MARK,
"exclamation": SILENCE_EXCLAMATION,
"dash": SILENCE_DASH,
"ellipsis": SILENCE_ELLIPSIS,
"quote_end": SILENCE_QUOTE_END,
}
if boundary_type in silence_durations:
duration = silence_durations[boundary_type]
audio = AudioSegment.from_wav(wav_path)
silence_segment = AudioSegment.silent(duration=duration)
extended_audio = audio + silence_segment
extended_audio.export(wav_path, format="wav")
logging.info(f"π Added {duration}ms silence for {boundary_type}: {wav_path.name}")
def add_chunk_end_silence(wav_path):
"""Add configurable silence to end of chunk if enabled"""
if not ENABLE_CHUNK_END_SILENCE or CHUNK_END_SILENCE_MS <= 0:
return
try:
audio = AudioSegment.from_wav(wav_path)
silence_segment = AudioSegment.silent(duration=CHUNK_END_SILENCE_MS)
audio_with_silence = audio + silence_segment
audio_with_silence.export(wav_path, format="wav")
logging.info(f"β Added {CHUNK_END_SILENCE_MS}ms end silence to {wav_path.name}")
except Exception as e:
logging.warning(f"β οΈ Failed to add end silence to {wav_path.name}: {e}")
# ============================================================================
# AUDIO UTILITY FUNCTIONS
# ============================================================================
def get_wav_duration(wav_path):
"""Get WAV file duration"""
import wave
with wave.open(str(wav_path), 'rb') as wf:
frames = wf.getnframes()
rate = wf.getframerate()
return frames / float(rate)
def get_chunk_audio_duration(wav_path):
"""Get actual audio duration from WAV file"""
try:
data, sr = sf.read(str(wav_path))
return len(data) / sr
except:
# Fallback to wave module
return get_wav_duration(wav_path)
|