Capstone04's picture
Update asr_diarization/pipeline.py
e4b5544 verified
import os
import json
import re
import torch
import torchaudio
import noisereduce as nr
import numpy as np
from pyannote.audio import Pipeline
from transformers import WhisperProcessor, WhisperForConditionalGeneration, pipeline as hf_pipeline
import tempfile
from pyannote.core import Annotation, Segment
from pyannote.metrics.diarization import DiarizationErrorRate
from jiwer import wer, Compose, ToLowerCase, RemovePunctuation, RemoveMultipleSpaces, Strip
class ASR_Diarization:
def __init__(self, HF_TOKEN,
diar_model="pyannote/speaker-diarization-3.1",
asr_model="Capstone04/TrainedWhisper_Medium",
model_path=None,
use_vad=True,
vad_threshold=0.3,
min_segment_duration=0.5,
snr_threshold=15.0,
min_whisper_duration=0.3):
self.HF_TOKEN = HF_TOKEN
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.use_vad = use_vad
self.vad_threshold = vad_threshold
self.min_segment_duration = min_segment_duration
self.snr_threshold = snr_threshold
self.min_whisper_duration = min_whisper_duration
# Load diarization model
self.diar_pipeline = Pipeline.from_pretrained(diar_model, use_auth_token=HF_TOKEN)
self.diar_pipeline = self.diar_pipeline.to(torch.device(self.device))
# Load WebRTC VAD for post-diarization filtering
if self.use_vad:
try:
import webrtcvad
self.vad = webrtcvad.Vad(2)
print("WebRTC VAD loaded for post-diarization filtering")
except ImportError:
print("WebRTC VAD not available")
self.use_vad = False
# Load ASR model
if model_path and os.path.exists(model_path):
print(f"Loading custom ASR model from: {model_path}")
actual_asr_model = model_path
else:
print(f"Loading default ASR model: {asr_model}")
actual_asr_model = asr_model
processor = WhisperProcessor.from_pretrained(actual_asr_model, token=HF_TOKEN)
model = WhisperForConditionalGeneration.from_pretrained(actual_asr_model, token=HF_TOKEN).to(self.device)
self.asr_pipeline = hf_pipeline(
"automatic-speech-recognition",
model=model,
tokenizer=processor.tokenizer,
feature_extractor=processor.feature_extractor,
device=0 if self.device == "cuda" else -1,
return_timestamps=True
)
def clean_transcription_text(self, text):
"""Clean ASR text for better TTS performance"""
if not text:
return ""
# Basic cleaning
text = text.strip()
# Fix punctuation spacing for TTS
text = re.sub(r'\s+([.,!?;:])', r'\1', text) # Remove space before punctuation
text = re.sub(r'([.,!?;:])(?=\w)', r'\1 ', text) # Add space after punctuation
# Normalize whitespace
text = re.sub(r'\s+', ' ', text)
return text.strip()
def should_keep_segment(self, text, duration, rms_energy):
"""Generalized segment quality assessment"""
# Duration too short
if duration < self.min_whisper_duration:
return False
# Energy too low (likely noise)
if rms_energy < 0.001:
return False
# Text too short or just punctuation
clean_text = text.strip()
if len(clean_text) <= 1:
return False
return True
def calculate_snr(self, audio_path):
"""NEW: Calculate SNR using RMS energy"""
try:
import librosa
y, sr = librosa.load(audio_path, sr=16000, mono=True)
# RMS-based SNR
rms = librosa.feature.rms(y=y)[0]
if len(rms) == 0:
return float('inf')
# Signal = high RMS regions, Noise = low RMS regions
high_rms = rms[rms > np.percentile(rms, 70)]
low_rms = rms[rms <= np.percentile(rms, 30)]
if len(high_rms) == 0 or len(low_rms) == 0:
return float('inf')
signal_power = np.mean(high_rms)
noise_power = np.mean(low_rms)
if noise_power == 0:
return float('inf')
snr = 10 * np.log10(signal_power / noise_power)
return snr
except Exception as e:
print(f"SNR calculation failed: {e}")
return float('inf')
def calculate_rms_energy(self, audio_chunk):
"""Calculate RMS energy for audio chunk"""
return np.sqrt(np.mean(audio_chunk**2))
def run_webrtc_vad_on_segment(self, audio_path, segment_start, segment_end):
"""Run WebRTC VAD on segment to get speech ratio"""
if not self.use_vad:
return 1.0
try:
import wave
# Load audio
with wave.open(audio_path, "rb") as wf:
sample_rate = wf.getframerate()
n_frames = wf.getnframes()
audio_data = wf.readframes(n_frames)
audio_array = np.frombuffer(audio_data, dtype=np.int16)
start_sample = int(segment_start * sample_rate)
end_sample = int(segment_end * sample_rate)
segment_audio = audio_array[start_sample:end_sample]
segment_bytes = segment_audio.tobytes()
# WebRTC VAD processing (30ms frames)
frame_duration = 30
bytes_per_sample = 2
frame_size = int(sample_rate * frame_duration / 1000) * bytes_per_sample
speech_frames = 0
total_frames = 0
for i in range(0, len(segment_bytes) - frame_size + 1, frame_size):
frame = segment_bytes[i:i + frame_size]
if len(frame) == frame_size:
is_speech = self.vad.is_speech(frame, sample_rate)
if is_speech:
speech_frames += 1
total_frames += 1
return speech_frames / total_frames if total_frames > 0 else 0.0
except Exception as e:
print(f"WebRTC VAD failed: {e}")
return 0.0
def run_diarization(self, audio_path):
"""Run diarization with VAD AFTER approach"""
# Step 1: Diarization sees FULL audio first
diarization = self.diar_pipeline(audio_path)
diar_segments = [
{"start": t.start, "end": t.end, "speaker": spk}
for t, _, spk in diarization.itertracks(yield_label=True)
]
raw_speakers = list(set([seg['speaker'] for seg in diar_segments]))
print(f"Diarization detected {len(raw_speakers)} speakers: {sorted(raw_speakers)}")
# Step 2: Calculate SNR for adaptive processing
snr = self.calculate_snr(audio_path)
# Step 3: Apply VAD filtering ONLY if low SNR
if snr < self.snr_threshold and self.use_vad:
print(f"Low SNR ({snr:.1f} dB), applying VAD filtering")
filtered_segments = []
for seg in diar_segments:
# Skip VAD for very short segments
if (seg["end"] - seg["start"]) < 0.2:
continue
speech_ratio = self.run_webrtc_vad_on_segment(
audio_path, seg["start"], seg["end"]
)
if speech_ratio >= self.vad_threshold:
filtered_segments.append(seg)
else:
print(f"Filtered low-speech segment: {seg['start']:.2f}-{seg['end']:.2f} (speech: {speech_ratio:.1%})")
diar_segments = filtered_segments
else:
print(f"Good SNR ({snr:.1f} dB), using all diarization segments")
# Step 4: Duration filtering for Whisper
filtered_segments = [
seg for seg in diar_segments
if (seg["end"] - seg["start"]) >= self.min_whisper_duration
]
print(f"Final: {len(filtered_segments)} segments for Whisper")
return filtered_segments
def merge_consecutive_speaker_segments(self, segments):
"""Merge only consecutive segments from the same speaker while preserving order"""
if not segments:
return []
# Sort by start time to ensure correct order
segments.sort(key=lambda x: x["start"])
merged_segments = []
for seg in segments:
if not merged_segments:
# First segment
merged_segments.append(seg)
else:
last_seg = merged_segments[-1]
# Check if same speaker AND consecutive (small gap < 2 seconds)
if (seg["speaker"] == last_seg["speaker"] and
(seg["start"] - last_seg["end"]) < 2.0):
# Merge with previous segment
last_seg["text"] += " " + seg["text"]
last_seg["end"] = seg["end"]
else:
# Different speaker or large gap - keep as separate segment
merged_segments.append(seg)
print(f"Reduced {len(segments)} segments to {len(merged_segments)} while preserving order")
return merged_segments
def run_transcription(self, audio_path, diar_json):
"""Segment-level transcription without word timestamps"""
# Load and standardize audio
audio, sr = torchaudio.load(audio_path)
# Resample to 16kHz for consistency
if sr != 16000:
resampler = torchaudio.transforms.Resample(sr, 16000)
audio = resampler(audio)
sr = 16000
merged_segments = []
speaker_segments = {}
# Calculate SNR for adaptive noise reduction
snr = self.calculate_snr(audio_path)
for seg in diar_json:
start, end, spk = seg["start"], seg["end"], seg["speaker"]
# Skip segments that are too short for Whisper
segment_duration = end - start
if segment_duration < self.min_whisper_duration:
print(f"Skipping short segment for Whisper: {start:.2f}-{end:.2f} ({segment_duration:.2f}s)")
continue
start_sample, end_sample = int(start * sr), int(end * sr)
# Handle both mono and stereo audio
if audio.shape[0] > 1: # Stereo
chunk = torch.mean(audio[:, start_sample:end_sample], dim=0).numpy()
else: # Mono
chunk = audio[0, start_sample:end_sample].numpy()
# Calculate RMS energy for this segment
rms_energy = self.calculate_rms_energy(chunk)
# Adaptive noise reduction based on SNR + RMS
if len(chunk) > int(0.1 * sr):
if snr < 10 or rms_energy < 0.01: # Very noisy or low energy
reduced = nr.reduce_noise(y=chunk, sr=sr, stationary=True, prop_decrease=0.8)
elif snr < 20: # Moderately noisy
reduced = nr.reduce_noise(y=chunk, sr=sr, stationary=True, prop_decrease=0.5)
else: # Clean audio
reduced = chunk
else:
reduced = chunk
try:
# Get text without timestamps
result = self.asr_pipeline(
reduced,
generate_kwargs={
"task": "transcribe",
"language": "en",
"temperature": 0.0 # More accurate transcription
}
)
except Exception as e:
print(f"Whisper failed on segment {start:.2f}-{end:.2f}: {e}")
continue
# Extract just the text (no timestamp processing)
text = result.get("text", "").strip()
# Clean the text for TTS and apply quality filtering
clean_text = self.clean_transcription_text(text)
if clean_text and self.should_keep_segment(clean_text, segment_duration, rms_energy):
seg_dict = {
"speaker": spk,
"start": start, # Keep segment boundaries
"end": end, # Keep segment boundaries
"text": clean_text, # Use cleaned text
"rms_energy": float(rms_energy)
}
merged_segments.append(seg_dict)
if spk not in speaker_segments:
speaker_segments[spk] = []
speaker_segments[spk].append(seg_dict)
return merged_segments, list(speaker_segments.keys())
def run_pipeline(self, audio_path, output_dir=None, base_name=None,
ref_rttm=None, ref_json=None, nse_events=None):
"""Add input validation and proper RTTM format"""
# Validate input audio file
if not os.path.exists(audio_path):
raise FileNotFoundError(f"Audio file not found: {audio_path}")
try:
# Quick validation that it's loadable audio
audio, sr = torchaudio.load(audio_path)
if audio.numel() == 0:
raise ValueError("Audio file is empty")
except Exception as e:
raise ValueError(f"Invalid audio file: {e}")
print(f"Processing with VAD: {'ON' if self.use_vad else 'OFF'}")
# Run diarization and transcription
diar_json = self.run_diarization(audio_path)
merged_segments, speakers = self.run_transcription(audio_path, diar_json)
# Merge consecutive segments by same speaker
merged_segments = self.merge_consecutive_speaker_segments(merged_segments)
# Combine ASR segments with NSE events if provided
if nse_events:
print(f"Combining {len(merged_segments)} ASR segments with {len(nse_events)} NSE events")
all_segments = merged_segments + nse_events
# Sort by start time for proper timeline
all_segments.sort(key=lambda x: x["start"])
else:
all_segments = merged_segments
if output_dir and base_name:
os.makedirs(output_dir, exist_ok=True)
# Save RTTM with standard format and precision
rttm_path = os.path.join(output_dir, f"{base_name}.rttm")
with open(rttm_path, "w") as f:
for seg in diar_json:
f.write(
f"SPEAKER {base_name} 1 {seg['start']:.3f} "
f"{seg['end']-seg['start']:.3f} <NA> <NA> "
f"{seg['speaker']} <NA> <NA>\n"
)
# Save transcription (with NSE events if available)
merged_path = os.path.join(output_dir, f"{base_name}_merged_transcription.json")
with open(merged_path, "w") as f:
json.dump(all_segments, f, indent=2)
# Evaluation if refs are provided
eval_results = None
if ref_rttm or ref_json:
eval_results = self.evaluate(output_dir, base_name,
ref_rttm=ref_rttm, ref_json=ref_json)
return {
"speakers": speakers,
"segments": all_segments, # Return combined segments
"evaluation": eval_results
}
def evaluate(self, output_dir, base_name, ref_rttm=None, ref_json=None):
# Add output_dir validation
if not output_dir or not base_name:
return None
results = {}
hyp_rttm = os.path.join(output_dir, f"{base_name}.rttm")
hyp_json = os.path.join(output_dir, f"{base_name}_merged_transcription.json")
if ref_rttm and os.path.exists(hyp_rttm):
def load_rttm(path):
ann = Annotation()
for line in open(path):
if line.startswith("SPEAKER"):
p = line.split()
start, dur, spk = float(p[3]), float(p[4]), p[7]
ann[Segment(start, start+dur)] = spk
return ann
der_score = DiarizationErrorRate()(load_rttm(ref_rttm), load_rttm(hyp_rttm))
results["DER"] = round(der_score * 100, 2)
if ref_json and os.path.exists(hyp_json):
def load_words_from_hypothesis(path):
"""Load text from YOUR pipeline output (has 'text' field)"""
data = json.load(open(path))
# Filter out NSE events for WER calculation (only use speech)
speech_segments = [seg for seg in data if seg.get("speaker") != "NSE"]
# Directly use segment text instead of tokens
return " ".join([seg["text"] for seg in speech_segments])
def load_words_from_reference(path):
"""Load text from REFERENCE file (has 'tokens' field)"""
data = json.load(open(path))
# Filter out NSE events for WER calculation (only use speech)
speech_segments = [seg for seg in data if seg.get("speaker") != "NSE"]
# Reference format has tokens, not direct text
return " ".join([tok["text"] for seg in speech_segments for tok in seg["tokens"]])
# Use appropriate loader for each file
ref_text = load_words_from_reference(ref_json)
hyp_text = load_words_from_hypothesis(hyp_json)
transform = Compose([ToLowerCase(), RemovePunctuation(),
RemoveMultipleSpaces(), Strip()])
results["WER_raw"] = round(wer(ref_text, hyp_text), 4)
results["WER_normalized"] = round(wer(transform(ref_text), transform(hyp_text)), 4)
return results if results else None
def __call__(self, inputs, nse_events=None):
"""FIXED: Add proper temporary file cleanup"""
if isinstance(inputs, dict):
if "audio_bytes" in inputs:
audio_bytes = inputs["audio_bytes"]
elif "audio" in inputs:
audio_bytes = inputs["audio"]
else:
raise ValueError("No audio found in inputs")
else:
audio_bytes = inputs
tmp_path = None
try:
# Create temporary file for processing
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
tmp.write(audio_bytes)
tmp_path = tmp.name
# Run pipeline with NSE events
result = self.run_pipeline(tmp_path, nse_events=nse_events)
return result
finally:
# Always clean up temporary file
if tmp_path and os.path.exists(tmp_path):
os.unlink(tmp_path)