轉寫粵語音頻

transcribe_cantonese.py

import librosa
import torch
from transformers import pipeline
from time import time

# Load the audio file
y, sr = librosa.load('test.mp3', sr=16000)

MODEL_NAME = "alvanlii/whisper-small-cantonese"

# Initialize the pipeline
device = "cuda:0" if torch.cuda.is_available() else "cpu"
pipe = pipeline(
    "automatic-speech-recognition",
    model=MODEL_NAME,
    chunk_length_s=30,
    device=device
)

# Set forced decoder IDs for Cantonese transcription
pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(
    language='zh', task="transcribe"
)

# Define chunk size (in samples) and stride
chunk_size = 30 * sr  # 30 seconds
stride = 5 * sr  # 5 seconds

# Create chunks with overlap
chunks = [y[i:i + chunk_size] for i in range(0, len(y), chunk_size - stride)]

# Process chunks in parallel using the pipeline's batching capability

start_time = time()
results = pipe(chunks, batch_size=8)
end_time = time()
elapsed_time = end_time - start_time

# Combine all transcriptions
full_transcription = '。'.join(result['text'] for result in results)
print(full_transcription)
print(f"Time taken for transcription: {elapsed_time} seconds")