1hr_demo / segment.py
tuklu's picture
Override with Parquet version
a0984c9 verified
import json
from pydub import AudioSegment
import os
import sys
def segment_audio_from_alignment(audio_path, alignment_path):
"""
Segments a long audio file into chunks based on a word-level
timestamped alignment file.
"""
output_dir = "wavs"
metadata_path = "metadata.csv"
os.makedirs(output_dir, exist_ok=True)
print(f"Loading full audio from: {audio_path}")
try:
full_audio = AudioSegment.from_file(audio_path)
except Exception as e:
print(f"--- ERROR: Could not load the audio file. ---")
print(f"Ensure ffmpeg is correctly installed and the path is correct.")
print(f"Details: {e}")
sys.exit(1)
print(f"Loading alignment data from: {alignment_path}")
with open(alignment_path, 'r', encoding='utf-8') as f:
alignment_data = json.load(f)
with open(metadata_path, 'w', encoding='utf-8') as metadata_file:
print(f"Writing chunks to '{output_dir}' and metadata to '{metadata_path}'...")
chunk_counter = 0
for segment in alignment_data.get('segments', []):
for word_data in segment.get('words', []):
start_time = word_data.get('start')
end_time = word_data.get('end')
text = word_data.get('word', '').strip()
if start_time is None or end_time is None or not text:
continue
start_ms = int(start_time * 1000)
end_ms = int(end_time * 1000)
audio_chunk = full_audio[start_ms:end_ms]
chunk_counter += 1
chunk_filename = os.path.join(output_dir, f"chunk_{chunk_counter:05d}.wav")
audio_chunk.export(chunk_filename, format="wav")
metadata_file.write(f"{os.path.basename(output_dir)}/{os.path.basename(chunk_filename)}|{text}\n")
print(f"Processed {chunk_counter} audio chunks.")
def main():
audio_file = "master.mp3"
alignment_file = "aligned_transcription.json"
if not os.path.exists(audio_file) or not os.path.exists(alignment_file):
print("--- ERROR: Make sure 'master.mp3' and 'aligned_transcription.json' are in the project directory. ---")
sys.exit(1)
segment_audio_from_alignment(audio_file, alignment_file)
if __name__ == "__main__":
main()