oceansweep's picture
?
ed28876
raw
history blame
7.34 kB
# Diarization_Lib.py
#########################################
# Diarization Library
# This library is used to perform diarization of audio files.
# Currently, uses FIXME for transcription.
#
####################
####################
# Function List
#
# 1. speaker_diarize(video_file_path, segments, embedding_model = "pyannote/embedding", embedding_size=512, num_speakers=0)
#
####################
# Import necessary libraries
import configparser
import json
import logging
import os
from pathlib import Path
import time
# Import Local
from App_Function_Libraries.Audio_Transcription_Lib import speech_to_text
#
# Import 3rd Party
from pyannote.audio import Model
from pyannote.audio.pipelines.speaker_diarization import SpeakerDiarization
import torch
import yaml
#
#######################################################################################################################
# Function Definitions
#
def load_pipeline_from_pretrained(path_to_config: str | Path) -> SpeakerDiarization:
path_to_config = Path(path_to_config).resolve()
print(f"Loading pyannote pipeline from {path_to_config}...")
if not path_to_config.exists():
raise FileNotFoundError(f"Config file not found: {path_to_config}")
# Load the YAML configuration
with open(path_to_config, 'r') as config_file:
config = yaml.safe_load(config_file)
# Store current working directory
cwd = Path.cwd().resolve()
# Change to the directory containing the config file
cd_to = path_to_config.parent.resolve()
print(f"Changing working directory to {cd_to}")
os.chdir(cd_to)
try:
# Create a SpeakerDiarization pipeline
pipeline = SpeakerDiarization()
# Load models explicitly from local paths
embedding_path = Path(config['pipeline']['params']['embedding']).resolve()
segmentation_path = Path(config['pipeline']['params']['segmentation']).resolve()
if not embedding_path.exists():
raise FileNotFoundError(f"Embedding model file not found: {embedding_path}")
if not segmentation_path.exists():
raise FileNotFoundError(f"Segmentation model file not found: {segmentation_path}")
# Load the models from local paths using pyannote's Model class
pipeline.embedding = Model.from_pretrained(str(embedding_path), map_location=torch.device('cpu'))
pipeline.segmentation = Model.from_pretrained(str(segmentation_path), map_location=torch.device('cpu'))
# Set other parameters
pipeline.clustering = config['pipeline']['params']['clustering']
pipeline.embedding_batch_size = config['pipeline']['params']['embedding_batch_size']
pipeline.embedding_exclude_overlap = config['pipeline']['params']['embedding_exclude_overlap']
pipeline.segmentation_batch_size = config['pipeline']['params']['segmentation_batch_size']
# Set additional parameters
pipeline.instantiate(config['params'])
finally:
# Change back to the original working directory
print(f"Changing working directory back to {cwd}")
os.chdir(cwd)
return pipeline
def audio_diarization(audio_file_path):
logging.info('audio-diarization: Loading pyannote pipeline')
config = configparser.ConfigParser()
config.read('config.txt')
processing_choice = config.get('Processing', 'processing_choice', fallback='cpu')
base_dir = Path(__file__).parent.resolve()
config_path = base_dir / 'models' / 'config.yaml'
pipeline = load_pipeline_from_pretrained(config_path)
time_start = time.time()
if audio_file_path is None:
raise ValueError("audio-diarization: No audio file provided")
logging.info("audio-diarization: Audio file path: %s", audio_file_path)
try:
_, file_ending = os.path.splitext(audio_file_path)
out_file = audio_file_path.replace(file_ending, ".diarization.json")
prettified_out_file = audio_file_path.replace(file_ending, ".diarization_pretty.json")
if os.path.exists(out_file):
logging.info("audio-diarization: Diarization file already exists: %s", out_file)
with open(out_file) as f:
global diarization_result
diarization_result = json.load(f)
return diarization_result
logging.info('audio-diarization: Starting diarization...')
diarization_result = pipeline(audio_file_path)
segments = []
for turn, _, speaker in diarization_result.itertracks(yield_label=True):
chunk = {
"Time_Start": turn.start,
"Time_End": turn.end,
"Speaker": speaker
}
logging.debug("Segment: %s", chunk)
segments.append(chunk)
logging.info("audio-diarization: Diarization completed with pyannote")
output_data = {'segments': segments}
logging.info("audio-diarization: Saving prettified JSON to %s", prettified_out_file)
with open(prettified_out_file, 'w') as f:
json.dump(output_data, f, indent=2)
logging.info("audio-diarization: Saving JSON to %s", out_file)
with open(out_file, 'w') as f:
json.dump(output_data, f)
except Exception as e:
logging.error("audio-diarization: Error performing diarization: %s", str(e))
raise RuntimeError("audio-diarization: Error performing diarization")
return segments
def combine_transcription_and_diarization(audio_file_path):
logging.info('combine-transcription-and-diarization: Starting transcription and diarization...')
transcription_result = speech_to_text(audio_file_path)
diarization_result = audio_diarization(audio_file_path)
combined_result = []
for transcription_segment in transcription_result:
for diarization_segment in diarization_result:
if transcription_segment['Time_Start'] >= diarization_segment['Time_Start'] and transcription_segment[
'Time_End'] <= diarization_segment['Time_End']:
combined_segment = {
"Time_Start": transcription_segment['Time_Start'],
"Time_End": transcription_segment['Time_End'],
"Speaker": diarization_segment['Speaker'],
"Text": transcription_segment['Text']
}
combined_result.append(combined_segment)
break
_, file_ending = os.path.splitext(audio_file_path)
out_file = audio_file_path.replace(file_ending, ".combined.json")
prettified_out_file = audio_file_path.replace(file_ending, ".combined_pretty.json")
logging.info("combine-transcription-and-diarization: Saving prettified JSON to %s", prettified_out_file)
with open(prettified_out_file, 'w') as f:
json.dump(combined_result, f, indent=2)
logging.info("combine-transcription-and-diarization: Saving JSON to %s", out_file)
with open(out_file, 'w') as f:
json.dump(combined_result, f)
return combined_result
#
#
#######################################################################################################################