voice-tools / src /services /speaker_extraction.py
jcudit's picture
jcudit HF Staff
fix: resolve ZeroGPU pickling errors across all audio processing services
3fb465f
"""
Speaker Extraction Service
Extracts specific speaker from audio using reference clip and cosine similarity matching.
Uses pyannote.audio embedding model for speaker verification.
"""
import logging
import time
from pathlib import Path
from typing import Callable, Dict, List, Optional, Tuple
import numpy as np
import torch
try:
import spaces
except ImportError:
# Create a no-op decorator for environments without spaces package
class spaces:
@staticmethod
def GPU(duration=60):
def decorator(func):
return func
return decorator
# Workaround for PyTorch 2.6+ weights_only security feature
# pyannote models are from trusted source (HuggingFace)
# Monkey-patch torch.load to use weights_only=False for pyannote models
_original_torch_load = torch.load
def _patched_torch_load(*args, **kwargs):
# Force weights_only=False since we trust pyannote models from HuggingFace
kwargs["weights_only"] = False
return _original_torch_load(*args, **kwargs)
torch.load = _patched_torch_load
from pyannote.audio import Pipeline
from src.config.gpu_config import GPUConfig
from src.lib.audio_io import get_audio_duration, read_audio, write_audio
from src.lib.progress import SPEAKER_EXTRACTION_STAGES
from src.models.audio_segment import AudioSegment, SegmentType
from src.models.error_report import ErrorReport
from src.services.audio_concatenation import AudioConcatenationUtility
logger = logging.getLogger(__name__)
# Module-level GPU functions to avoid pickling issues with ZeroGPU
@spaces.GPU(duration=60)
def _extract_embedding_on_gpu(audio_dict: Dict, hf_token: str) -> np.ndarray:
"""
Extract speaker embedding on GPU (or CPU if unavailable).
This is a module-level function to avoid pickling issues with ZeroGPU.
The model is loaded fresh within this GPU context.
Args:
audio_dict: Audio data dict with 'waveform' and 'sample_rate'
hf_token: HuggingFace token for model access
Returns:
Speaker embedding vector
"""
from pyannote.audio import Inference, Model
# Load model fresh in GPU context (avoids pickling)
logger.info("Loading embedding model in GPU context...")
model = Model.from_pretrained("pyannote/wespeaker-voxceleb-resnet34-LM", token=hf_token)
# Move to available device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
logger.info(f"Embedding model loaded on {device}")
# Create inference wrapper
embedding_model = Inference(model, window="whole")
try:
embedding = embedding_model(audio_dict)
# Embedding is already a numpy array from Inference
if isinstance(embedding, torch.Tensor):
embedding = embedding.detach().cpu().numpy()
# Flatten if needed
if len(embedding.shape) > 1:
embedding = embedding.flatten()
logger.info(f"Extracted {len(embedding)}-dimensional embedding")
return embedding
finally:
# Clean up
del embedding_model
del model
if torch.cuda.is_available():
torch.cuda.empty_cache()
@spaces.GPU(duration=60)
def _extract_embeddings_batch_on_gpu(
audio_data: np.ndarray,
sample_rate: int,
segments: List[AudioSegment],
hf_token: str,
progress_callback: Optional[Callable] = None,
) -> List[Tuple[AudioSegment, np.ndarray]]:
"""
Extract embeddings for multiple segments on GPU.
This is a module-level function to avoid pickling issues with ZeroGPU.
The model is loaded fresh within this GPU context.
Args:
audio_data: Full audio array
sample_rate: Sample rate
segments: List of AudioSegment objects to process
hf_token: HuggingFace token for model access
progress_callback: Optional progress callback
Returns:
List of (AudioSegment, embedding) tuples
"""
from pyannote.audio import Inference, Model
# Load model fresh in GPU context (avoids pickling)
logger.info("Loading embedding model in GPU context...")
model = Model.from_pretrained("pyannote/wespeaker-voxceleb-resnet34-LM", token=hf_token)
# Move to available device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
logger.info(f"Embedding model loaded on {device}")
# Create inference wrapper
embedding_model = Inference(model, window="whole")
try:
segments_with_embeddings = []
for i, segment in enumerate(segments):
if progress_callback:
# Progress from 0.15 to 0.40 for embedding computation
embed_progress = 0.15 + (0.25 * (i + 1) / len(segments))
progress_callback(
SPEAKER_EXTRACTION_STAGES[1], embed_progress, 1.0
) # "Computing embeddings"
# Extract segment audio
start_sample = int(segment.start_time * sample_rate)
end_sample = int(segment.end_time * sample_rate)
segment_audio = audio_data[start_sample:end_sample]
# Skip if segment too short
if len(segment_audio) < sample_rate * 0.5: # 0.5 second minimum
continue
# Extract embedding
audio_tensor = torch.from_numpy(segment_audio).unsqueeze(0)
audio_dict = {"waveform": audio_tensor, "sample_rate": sample_rate}
embedding = embedding_model(audio_dict)
# Embedding is already a numpy array from Inference
if isinstance(embedding, torch.Tensor):
embedding = embedding.detach().cpu().numpy()
# Flatten if needed
if len(embedding.shape) > 1:
embedding = embedding.flatten()
segments_with_embeddings.append((segment, embedding))
logger.info(f"Extracted embeddings from {len(segments_with_embeddings)} segments")
return segments_with_embeddings
finally:
# Clean up
del embedding_model
del model
if torch.cuda.is_available():
torch.cuda.empty_cache()
class SpeakerExtractionService:
"""
Service for extracting specific speaker from audio files using reference clips.
Uses speaker embeddings and cosine similarity to match segments.
"""
def __init__(self):
"""Initialize speaker extraction service"""
import os
# Store HF token for GPU functions to use
self.hf_token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_TOKEN")
if not self.hf_token:
raise ValueError(
"HuggingFace token required. Set HF_TOKEN or HUGGINGFACE_TOKEN environment variable."
)
# Initialize audio concatenation utility
self.audio_concatenator = AudioConcatenationUtility()
logger.info("Speaker extraction service initialized")
def extract_reference_embedding(self, reference_clip_path: str) -> np.ndarray:
"""
Extract speaker embedding from reference clip.
Args:
reference_clip_path: Path to reference audio clip
Returns:
Speaker embedding vector (512-dimensional)
Raises:
ValueError: If reference clip is too short or invalid
"""
# Validate reference clip duration
duration = get_audio_duration(reference_clip_path)
if duration < 3.0:
raise ValueError(
f"Reference clip is {duration:.1f}s (minimum 3.0s required for reliable matching)"
)
logger.info(f"Extracting embedding from reference clip ({duration:.1f}s)")
# Read audio
audio_data, sample_rate = read_audio(reference_clip_path, target_sr=16000)
# Check audio quality
rms = np.sqrt(np.mean(audio_data**2))
if rms < 0.01:
logger.warning(
f"Reference clip has low amplitude (RMS={rms:.4f}). "
"Consider using a cleaner sample for better results."
)
# Convert to torch tensor
audio_tensor = torch.from_numpy(audio_data).unsqueeze(0) # Add batch dimension
# Extract embedding using Inference model
audio_dict = {"waveform": audio_tensor, "sample_rate": sample_rate}
# Call module-level GPU function (avoids pickling self)
embedding = _extract_embedding_on_gpu(audio_dict, self.hf_token)
return embedding
def detect_voice_segments(
self, audio_path: str, min_duration: float = 0.5
) -> List[AudioSegment]:
"""
Detect voice activity segments in audio file using simple chunking.
For now, we use fixed-size chunks since VAD requires additional model access.
In production, this should use proper VAD.
Args:
audio_path: Path to audio file
min_duration: Minimum segment duration in seconds
Returns:
List of AudioSegment objects for voice activity
"""
logger.info(f"Detecting voice segments in {Path(audio_path).name}...")
# Simple approach: split into fixed chunks (can be improved with VAD)
duration = get_audio_duration(audio_path)
# Create 5-second chunks (good balance for embedding extraction)
chunk_duration = 5.0
segments = []
current_time = 0.0
while current_time < duration:
end_time = min(current_time + chunk_duration, duration)
if end_time - current_time >= min_duration:
audio_segment = AudioSegment(
start_time=current_time,
end_time=end_time,
speaker_id="UNKNOWN",
confidence=1.0,
segment_type=SegmentType.SPEECH,
)
segments.append(audio_segment)
current_time = end_time
logger.info(f"Created {len(segments)} segments ({chunk_duration}s chunks)")
return segments
def extract_target_embeddings(
self, target_audio_path: str, progress_callback: Optional[Callable] = None
) -> List[Tuple[AudioSegment, np.ndarray]]:
"""
Extract embeddings from all voice segments in target audio.
Args:
target_audio_path: Path to target audio file
progress_callback: Optional callback for progress updates (stage, current, total)
Returns:
List of tuples (AudioSegment, embedding)
"""
# Detect voice segments
segments = self.detect_voice_segments(target_audio_path)
if len(segments) == 0:
logger.warning("No voice segments detected in target audio")
return []
# Load full audio
audio_data, sample_rate = read_audio(target_audio_path, target_sr=16000)
# Call module-level GPU function (avoids pickling self)
segments_with_embeddings = _extract_embeddings_batch_on_gpu(
audio_data=audio_data,
sample_rate=sample_rate,
segments=segments,
hf_token=self.hf_token,
progress_callback=progress_callback,
)
return segments_with_embeddings
def compute_similarity(self, embedding1: np.ndarray, embedding2: np.ndarray) -> float:
"""
Compute cosine similarity between two embeddings.
Args:
embedding1: First embedding vector
embedding2: Second embedding vector
Returns:
Cosine similarity score (-1 to 1, higher is more similar)
"""
# Normalize embeddings
norm1 = np.linalg.norm(embedding1)
norm2 = np.linalg.norm(embedding2)
if norm1 == 0 or norm2 == 0:
return 0.0
# Compute cosine similarity
similarity = np.dot(embedding1, embedding2) / (norm1 * norm2)
return float(similarity)
def match_segments(
self,
reference_embedding: np.ndarray,
segments_with_embeddings: List[Tuple[AudioSegment, np.ndarray]],
threshold: float = 0.40,
min_confidence: float = 0.30,
) -> List[Tuple[AudioSegment, float]]:
"""
Match segments against reference embedding using similarity threshold.
Args:
reference_embedding: Reference speaker embedding
segments_with_embeddings: List of (segment, embedding) tuples
threshold: Similarity threshold (lower is stricter, 0.0-1.0)
min_confidence: Minimum segment confidence to include
Returns:
List of (segment, similarity_score) tuples for matched segments
"""
matched = []
for segment, embedding in segments_with_embeddings:
# Filter by segment confidence
if segment.confidence < min_confidence:
continue
# Compute similarity
similarity = self.compute_similarity(reference_embedding, embedding)
# Match if similarity exceeds threshold
# Note: threshold is inverted - lower threshold = stricter matching
# We use (1 - threshold) as the actual similarity threshold
similarity_threshold = 1.0 - threshold
if similarity >= similarity_threshold:
matched.append((segment, similarity))
logger.info(
f"Matched {len(matched)}/{len(segments_with_embeddings)} segments "
f"(threshold={threshold:.2f}, min_confidence={min_confidence:.2f})"
)
return matched
def validate_reference_clip(self, reference_clip_path: str) -> Tuple[bool, str]:
"""
Validate reference clip quality and duration.
Args:
reference_clip_path: Path to reference clip
Returns:
Tuple of (is_valid, message)
"""
try:
# Check duration
duration = get_audio_duration(reference_clip_path)
if duration < 3.0:
return False, f"Reference clip is {duration:.1f}s (minimum 3.0s required)"
# Check audio quality
audio_data, sample_rate = read_audio(reference_clip_path)
rms = np.sqrt(np.mean(audio_data**2))
if rms < 0.01:
return (
True,
f"Warning: Low audio quality (RMS={rms:.4f}). Consider using cleaner sample.",
)
return True, "Reference clip is valid"
except Exception as e:
return False, f"Error validating reference clip: {str(e)}"
def extract_and_export(
self,
reference_clip: str,
target_file: str,
output_path: str,
threshold: float = 0.40,
min_confidence: float = 0.30,
concatenate: bool = True,
silence_duration_ms: int = 150,
crossfade_duration_ms: int = 75,
sample_rate: int = 44100,
bitrate: str = "192k",
progress_callback: Optional[Callable] = None,
) -> Dict:
"""
Extract speaker from target file and export to audio file.
Args:
reference_clip: Path to reference clip of target speaker
target_file: Path to target audio file
output_path: Path for output file(s)
threshold: Speaker matching threshold (0.0-1.0, lower is stricter)
min_confidence: Minimum confidence for including segments
concatenate: If True, concatenate segments; if False, export separately
silence_duration_ms: Silence duration between concatenated segments
crossfade_duration_ms: Crossfade duration for smooth transitions
sample_rate: Output sample rate
bitrate: Output bitrate
progress_callback: Optional callback for progress updates
Returns:
Extraction report dictionary or ErrorReport on failure
"""
start_time = time.time()
logger.info(f"Extracting speaker from {Path(target_file).name}")
logger.info(f"Reference clip: {Path(reference_clip).name}")
logger.info(f"Threshold: {threshold:.2f}, Min confidence: {min_confidence:.2f}")
try:
if progress_callback:
progress_callback(
SPEAKER_EXTRACTION_STAGES[0], 0.0, 1.0
) # "Loading reference audio"
# Extract reference embedding
reference_embedding = self.extract_reference_embedding(reference_clip)
except Exception as e:
logger.error(f"Failed to extract reference embedding: {e}")
error_report: ErrorReport = {
"status": "failed",
"error": f"Failed to extract reference embedding: {e}",
"error_type": "audio_io",
}
return error_report
try:
if progress_callback:
progress_callback(SPEAKER_EXTRACTION_STAGES[1], 0.15, 1.0) # "Computing embeddings"
# Extract target embeddings
# Note: progress_callback cannot be passed due to ZeroGPU pickling constraints
segments_with_embeddings = self.extract_target_embeddings(
target_file,
progress_callback=None, # Cannot pass callback to avoid pickling errors
)
if progress_callback:
progress_callback(
SPEAKER_EXTRACTION_STAGES[2], 0.4, 1.0
) # "Matching voice segments"
# Match segments
matched_segments = self.match_segments(
reference_embedding,
segments_with_embeddings,
threshold=threshold,
min_confidence=min_confidence,
)
except Exception as e:
logger.error(f"Failed to process and match segments: {e}")
error_report: ErrorReport = {
"status": "failed",
"error": f"Failed to process and match segments: {e}",
"error_type": "processing",
}
return error_report
if len(matched_segments) == 0:
logger.warning("No matching segments found")
report = self.generate_extraction_report(
reference_clip=reference_clip,
target_file=target_file,
threshold=threshold,
matched_segments=[],
processing_time=time.time() - start_time,
output_file=None,
)
return report
try:
if progress_callback:
progress_callback(SPEAKER_EXTRACTION_STAGES[3], 0.75, 1.0) # "Extracting segments"
# Load target audio
target_audio, target_sr = read_audio(target_file)
except Exception as e:
logger.error(f"Failed to load target audio: {e}")
error_report: ErrorReport = {
"status": "failed",
"error": f"Failed to load target audio: {e}",
"error_type": "audio_io",
}
return error_report
try:
# Export matched segments
output_path_obj = Path(output_path)
if concatenate:
# Concatenate all matched segments
segment_audio_list = []
for segment, similarity in matched_segments:
start_sample = int(segment.start_time * target_sr)
end_sample = int(segment.end_time * target_sr)
segment_audio = target_audio[start_sample:end_sample]
segment_audio_list.append(segment_audio)
# Concatenate with crossfade
concatenated = self.audio_concatenator.concatenate_segments(
segment_audio_list,
sample_rate=target_sr,
silence_duration_ms=silence_duration_ms,
crossfade_duration_ms=crossfade_duration_ms,
)
# Resample if needed
if sample_rate != target_sr:
from src.lib.audio_io import resample_audio
concatenated = resample_audio(concatenated, target_sr, sample_rate)
output_sr = sample_rate
else:
output_sr = target_sr
# Write output
write_audio(str(output_path), concatenated, output_sr)
logger.info(f"Exported concatenated audio to {output_path}")
output_file = str(output_path)
else:
# Export segments separately
output_path_obj.mkdir(parents=True, exist_ok=True)
for i, (segment, similarity) in enumerate(matched_segments, start=1):
start_sample = int(segment.start_time * target_sr)
end_sample = int(segment.end_time * target_sr)
segment_audio = target_audio[start_sample:end_sample]
# Resample if needed
if sample_rate != target_sr:
from src.lib.audio_io import resample_audio
segment_audio = resample_audio(segment_audio, target_sr, sample_rate)
output_sr = sample_rate
else:
output_sr = target_sr
segment_file = output_path_obj / f"segment_{i:03d}.m4a"
write_audio(str(segment_file), segment_audio, output_sr)
logger.info(f"Exported {len(matched_segments)} segments to {output_path_obj}")
output_file = str(output_path_obj)
if progress_callback:
progress_callback("Complete", 1.0, 1.0)
# Generate report
report = self.generate_extraction_report(
reference_clip=reference_clip,
target_file=target_file,
threshold=threshold,
matched_segments=matched_segments,
processing_time=time.time() - start_time,
output_file=output_file,
)
return report
except Exception as e:
logger.error(f"Failed to export speaker segments: {e}")
error_report: ErrorReport = {
"status": "failed",
"error": f"Failed to export speaker segments: {e}",
"error_type": "processing",
}
return error_report
def generate_extraction_report(
self,
reference_clip: str,
target_file: str,
threshold: float,
matched_segments: List[Tuple[AudioSegment, float]],
processing_time: float,
output_file: Optional[str],
) -> Dict:
"""
Generate extraction report JSON.
Args:
reference_clip: Reference clip path
target_file: Target file path
threshold: Matching threshold used
matched_segments: List of matched (segment, similarity) tuples
processing_time: Processing time in seconds
output_file: Output file path
Returns:
Report dictionary
"""
total_duration = sum(seg.duration for seg, _ in matched_segments)
avg_confidence = (
sum(similarity for _, similarity in matched_segments) / len(matched_segments)
if matched_segments
else 0.0
)
low_confidence = sum(
1
for _, similarity in matched_segments
if similarity < (1.0 - threshold + 0.1) # Within 0.1 of threshold
)
report = {
"reference_clip": str(reference_clip),
"target_file": str(target_file),
"threshold": threshold,
"segments_found": len(matched_segments),
"segments_included": len(matched_segments),
"total_duration_seconds": round(total_duration, 2),
"average_confidence": round(avg_confidence, 3),
"low_confidence_segments": low_confidence,
"processing_time_seconds": round(processing_time, 1),
"output_file": output_file,
}
return report