voice-tools / src /services /batch_processor.py
jcudit's picture
jcudit HF Staff
feat: complete audio speaker separation feature with 3 workflows
cb39c05
"""
BatchProcessor orchestrates the complete voice extraction pipeline.
This service coordinates VAD filtering, voice identification, speech/nonverbal
classification, quality filtering, and segment extraction to implement the
complete voice profiling workflow.
"""
import logging
from pathlib import Path
from typing import List, Optional, Tuple
import numpy as np
from ..lib.audio_io import extract_segment, read_audio, write_audio
from ..lib.format_converter import m4a_to_wav, wav_to_m4a
from ..models.audio_segment import AudioSegment, SegmentCollection, SegmentType
from ..models.processing_job import ExtractionMode, JobStatus, ProcessingJob
from ..models.voice_profile import VoiceProfile
from .speech_extractor import SpeechExtractor
from .vad_filter import VADFilter
from .voice_identifier import VoiceIdentifier
logger = logging.getLogger(__name__)
class BatchProcessor:
"""
Orchestrates the complete voice extraction pipeline.
Pipeline stages:
1. Format conversion (m4a → wav if needed)
2. VAD pre-filtering (identify voice activity regions)
3. Voice identification (match target speaker)
4. Speech/nonverbal classification
5. Quality filtering
6. Segment extraction and output generation
"""
def __init__(
self,
vad_threshold: float = 0.5,
voice_similarity_threshold: float = 0.7,
speech_confidence_threshold: float = 0.6,
enable_vad: bool = True,
):
"""
Initialize the batch processor.
Args:
vad_threshold: VAD confidence threshold (0-1)
voice_similarity_threshold: Voice matching threshold (0-1)
speech_confidence_threshold: Speech classification threshold (0-1)
enable_vad: Whether to use VAD pre-filtering
"""
self.vad_threshold = vad_threshold
self.voice_similarity_threshold = voice_similarity_threshold
self.speech_confidence_threshold = speech_confidence_threshold
self.enable_vad = enable_vad
# Initialize services
self.vad_filter = VADFilter()
self.voice_identifier = VoiceIdentifier()
self.speech_extractor = SpeechExtractor()
logger.info("BatchProcessor initialized")
def process_file(
self,
input_file: Path,
voice_profile: VoiceProfile,
output_dir: Path,
extraction_mode: ExtractionMode = ExtractionMode.SPEECH,
apply_quality_filter: bool = True,
) -> Tuple[List[AudioSegment], dict]:
"""
Process a single audio file through the complete pipeline.
Args:
input_file: Path to input audio file (m4a or wav)
voice_profile: Reference voice profile to match
output_dir: Directory for output files
extraction_mode: What to extract (SPEECH, NONVERBAL, or BOTH)
apply_quality_filter: Whether to filter by quality thresholds
Returns:
Tuple of (extracted_segments, statistics)
"""
logger.info(f"Processing file: {input_file}")
# Ensure output directory exists
output_dir.mkdir(parents=True, exist_ok=True)
# Convert to wav if needed
if input_file.suffix.lower() == ".m4a":
logger.info("Converting m4a to wav")
wav_path = output_dir / f"{input_file.stem}_temp.wav"
m4a_to_wav(str(input_file), str(wav_path))
working_file = wav_path
else:
working_file = input_file
# Load audio
audio, sample_rate = read_audio(str(working_file))
logger.info(f"Loaded audio: {len(audio) / sample_rate:.2f}s at {sample_rate}Hz")
# Stage 1: VAD pre-filtering (optional but recommended)
if self.enable_vad:
logger.info("Stage 1: VAD pre-filtering")
vad_stats = self.vad_filter.get_voice_activity_stats(
audio, sample_rate, self.vad_threshold
)
logger.info(
f"VAD: {vad_stats['voice_percentage']:.1f}% voice activity "
f"({vad_stats['voice_duration']:.1f}s of {vad_stats['total_duration']:.1f}s)"
)
if not vad_stats["worth_processing"]:
logger.warning("Insufficient voice activity, skipping file")
return [], {"error": "Insufficient voice activity"}
# Get voice-only segments for processing
vad_segments = self.vad_filter.detect_voice_activity(
audio, sample_rate, self.vad_threshold
)
else:
# Process entire file
vad_segments = [(0.0, len(audio) / sample_rate)]
vad_stats = {}
# Stage 2: Voice identification
logger.info("Stage 2: Voice identification")
matched_segments = self.voice_identifier.match_voice_profile(
str(working_file), voice_profile, similarity_threshold=self.voice_similarity_threshold
)
logger.info(f"Found {len(matched_segments)} segments matching voice profile")
if not matched_segments:
logger.warning("No matching voice segments found")
return [], {"error": "No matching voice segments"}
# Stage 3: Speech/nonverbal classification
logger.info(f"Stage 3: Speech/nonverbal classification (mode: {extraction_mode.value})")
if extraction_mode == ExtractionMode.SPEECH:
classified_segments = self.speech_extractor.extract_speech_segments(
audio, sample_rate, matched_segments, self.speech_confidence_threshold
)
elif extraction_mode == ExtractionMode.NONVERBAL:
classified_segments = self.speech_extractor.extract_nonverbal_segments(
audio, sample_rate, matched_segments, self.speech_confidence_threshold
)
else: # BOTH
classified_segments = matched_segments
logger.info(f"Classified {len(classified_segments)} segments as {extraction_mode.value}")
# Stage 4: Quality filtering
if apply_quality_filter:
logger.info("Stage 4: Quality filtering")
filtered_segments = self.speech_extractor.filter_by_quality(
audio, sample_rate, classified_segments
)
logger.info(
f"Quality filter: {len(filtered_segments)}/{len(classified_segments)} segments passed"
)
else:
filtered_segments = classified_segments
# Stage 5: Extract and save segments
logger.info("Stage 5: Extracting segments")
extracted_segments = []
for i, segment in enumerate(filtered_segments):
# Extract audio segment
segment_audio = extract_segment(audio, sample_rate, segment["start"], segment["end"])
# Create output filename
segment_type = segment.get("segment_type", SegmentType.SPEECH)
output_filename = (
f"{input_file.stem}_segment_{i + 1:03d}_"
f"{segment_type.value}_{segment['start']:.2f}s-{segment['end']:.2f}s.m4a"
)
output_path = output_dir / output_filename
# Save as m4a
temp_wav = output_dir / f"temp_segment_{i}.wav"
write_audio(str(temp_wav), segment_audio, sample_rate)
wav_to_m4a(str(temp_wav), str(output_path))
temp_wav.unlink() # Clean up temp file
# Create AudioSegment record
audio_segment = AudioSegment(
start_time=segment["start"],
end_time=segment["end"],
duration=segment["end"] - segment["start"],
segment_type=segment_type,
confidence=segment.get("confidence", 0.0),
voice_similarity=segment.get("similarity", 0.0),
snr=segment.get("snr"),
stoi=segment.get("stoi"),
pesq=segment.get("pesq"),
output_file=str(output_path),
)
extracted_segments.append(audio_segment)
# Generate statistics
collection = SegmentCollection(extracted_segments)
statistics = {
"input_file": str(input_file),
"total_duration": len(audio) / sample_rate,
"segments_extracted": len(extracted_segments),
"total_extracted_duration": collection.total_duration,
"extraction_percentage": collection.total_duration / (len(audio) / sample_rate) * 100,
"average_segment_duration": collection.average_duration,
"average_confidence": collection.average_confidence,
"average_quality_snr": collection.average_quality["snr"],
"vad_stats": vad_stats,
}
logger.info(
f"Extraction complete: {len(extracted_segments)} segments, "
f"{statistics['total_extracted_duration']:.2f}s "
f"({statistics['extraction_percentage']:.1f}%)"
)
# Clean up temp wav if created
if input_file.suffix.lower() == ".m4a":
working_file.unlink()
return extracted_segments, statistics
def process_batch(self, job: ProcessingJob) -> ProcessingJob:
"""
Process multiple files in batch.
Args:
job: ProcessingJob with configuration and file list
Returns:
Updated ProcessingJob with results
"""
logger.info(f"Starting batch job: {job.job_id}")
job.start()
# Create voice profile from reference file
try:
logger.info(f"Extracting voice profile from: {job.reference_file}")
voice_profile = self.voice_identifier.extract_voice_profile(str(job.reference_file))
logger.info(
f"Voice profile extracted: quality={voice_profile.embedding_quality:.2f}, "
f"duration={voice_profile.reference_duration:.2f}s"
)
except Exception as e:
logger.error(f"Failed to extract voice profile: {e}")
job.fail(f"Voice profile extraction failed: {e}")
return job
# Process each input file
for input_file in job.input_files:
try:
logger.info(
f"Processing file {job.files_processed + 1}/{len(job.input_files)}: {input_file}"
)
segments, stats = self.process_file(
Path(input_file),
voice_profile,
Path(job.output_dir),
extraction_mode=job.extraction_mode,
apply_quality_filter=True,
)
# Update job statistics
job.add_success(
input_duration=stats.get("total_duration", 0),
extracted_duration=stats.get("total_extracted_duration", 0),
)
logger.info(f"File processed successfully: {len(segments)} segments extracted")
except Exception as e:
logger.error(f"Failed to process {input_file}: {e}")
job.add_failure(str(input_file), str(e))
# Complete job
job.complete()
logger.info(
f"Batch job complete: {job.files_processed} files processed, {job.files_failed} failed"
)
return job
def estimate_processing_time(self, audio_file: Path, enable_vad: bool = True) -> dict:
"""
Estimate processing time for an audio file.
Args:
audio_file: Path to audio file
enable_vad: Whether VAD will be used
Returns:
Dictionary with time estimates
"""
# Load audio to get duration
audio, sample_rate = read_audio(str(audio_file))
total_duration = len(audio) / sample_rate
if enable_vad:
# Quick VAD scan
stats = self.vad_filter.get_voice_activity_stats(audio, sample_rate, self.vad_threshold)
voice_duration = stats["voice_duration"]
# Estimate: ~0.4x realtime with VAD
estimated_time = voice_duration * 0.4
else:
# Estimate: ~0.8x realtime without VAD
estimated_time = total_duration * 0.8
return {
"total_duration": total_duration,
"voice_duration": voice_duration if enable_vad else total_duration,
"estimated_processing_time": estimated_time,
"estimated_minutes": estimated_time / 60,
}