asr.py

import torch
import csv
from transformers import pipeline
import os
import librosa
from itertools import islice

OVER_SIZE_LIMIT = 200_000_000

csv.field_size_limit(OVER_SIZE_LIMIT)
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

generate_kwargs = {
    "language": "Japanese",
    "no_repeat_ngram_size": 0,
    "repetition_penalty": 1.0,
}

pipe = pipeline(
    "automatic-speech-recognition",
    model="litagin/anime-whisper",
    device="cuda",
    torch_dtype=torch.float16,
    chunk_length_s=30.0,
    batch_size=64,
)

# CSVファイルの準備
csv_file = r"C:\Users\user\Pictures\speech_recognition_results.csv"

# transcript.csvからデータを読み込む
transcripts = {}
with open(r"C:\Users\user\Pictures\transcript.csv", mode="r", encoding="utf-8") as file:
    reader = csv.DictReader(file)
    for row in reader:
        transcripts[row["filename"]] = row["transcript"]

audio_dir = r"C:\Users\user\Pictures\dataset_converted"
audio_files = os.listdir(audio_dir)

# CSV から既に処理されたファイル名を取得
processed_files = set()
if os.path.exists(csv_file):
    with open(csv_file, mode="r", encoding="utf-8") as file:
        reader = csv.DictReader(file)
        for row in reader:
            processed_files.add(row["Filename"])

# 処理を開始
batch_size = 256  # バッチサイズ
with open(csv_file, mode="a", newline="", encoding="utf-8") as file:
    writer = csv.writer(file)

    # ヘッダーがない場合は書き込む
    if not processed_files:
        writer.writerow(["Filename", "True", "ASR"])

    # 未処理ファイルをフィルタリング
    unprocessed_files = [
        f for f in audio_files if f in transcripts and f not in processed_files
    ]

    # バッチ処理
    for i in range(0, len(unprocessed_files), batch_size):
        batch_files = unprocessed_files[i : i + batch_size]
        audio_paths = [os.path.join(audio_dir, f) for f in batch_files]

        # 音声ファイルを読み込み
        audios = []
        for audio_path in audio_paths:
            y, sr = librosa.load(audio_path, sr=16000)
            audios.append(y)

        # ASR パイプラインを使ってバッチ処理
        results = pipe(audios, generate_kwargs=generate_kwargs)

        # CSVに結果を保存
        for audio_file, result in zip(batch_files, results):
            asr_result = result["text"]
            true_text = transcripts[audio_file]
            writer.writerow([audio_file, true_text, asr_result])

            print("Filename:", audio_file)
            print("True:", true_text)
            print("ASR :", asr_result)
            print()

print(f"結果は {csv_file} に保存されました。")