scripts/run_whisper_finetuned.py

import argparse
import os
import time
from pathlib import Path
import csv

import torch
import librosa
from transformers import WhisperForConditionalGeneration, WhisperProcessor

def save_csv(file_path, rows):
    with open(file_path, "w", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerows(rows)
        print(f"write csv to {file_path}")


def load_audio(audio_path: str, sr: int = 16000):
    # 读取音频并转成 16k 单声道 numpy float32
    audio, _ = librosa.load(audio_path, sr=sr, mono=True)
    return audio


def transcribe_file(
        audio_path: str,
        model,
        processor,
        language: str = "Chinese",
        task: str = "transcribe",
        timestamps: bool = False,
        max_new_tokens: int = 255,
):
    # 准备特征
    audio = load_audio(audio_path, sr=16000)
    inputs = processor(audio, sampling_rate=16000, return_tensors="pt")

    # 放到设备
    device = next(model.parameters()).device
    input_features = inputs["input_features"].to(device)

    # 生成
    with torch.inference_mode(), torch.autocast(device_type="cuda", enabled=(device.type == "cuda")):
        generated_ids = model.generate(
            input_features=input_features,
            max_new_tokens=max_new_tokens,
            return_timestamps=timestamps,  # 仅部分版本支持；不支持时自动忽略
        )

    # 解码
    text = processor.tokenizer.batch_decode(generated_ids.cpu().numpy(), skip_special_tokens=True)
    return text[0]


def main():
    parser = argparse.ArgumentParser("Simple Whisper Inference")
    parser.add_argument("--model_path", type=str, default="whisper-large-v3-turbo-finetune",
                        help="本地合并模型路径或HF模型名")
    parser.add_argument("--input", type=str, required=True,
                        help="音频文件路径，或目录（将批量处理其中的音频）")
    parser.add_argument("--language", type=str, default="Chinese",
                        help="语言（如 Chinese / English / zh / en）")
    parser.add_argument("--task", type=str, default="transcribe", choices=["transcribe", "translate"],
                        help="任务：转写或翻译")
    parser.add_argument("--timestamps", action="store_true", help="是否返回时间戳（若模型与版本支持）")
    parser.add_argument("--local_files_only", action="store_true", help="仅本地加载，不联网")
    parser.add_argument("--batch_exts", type=str, default=".wav,.mp3,.flac,.m4a",
                        help="当 --input 是目录时，处理这些后缀的文件，逗号分隔")
    args = parser.parse_args()

    # 加载处理器 & 模型
    processor = WhisperProcessor.from_pretrained(
        args.model_path,
        language=args.language,
        task=args.task,
        no_timestamps=not args.timestamps,
        local_files_only=args.local_files_only,
    )
    model = WhisperForConditionalGeneration.from_pretrained(
        args.model_path,
        device_map="auto",
        local_files_only=args.local_files_only,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    )

    model.generation_config.language = args.language.lower()
    model.generation_config.forced_decoder_ids = None
    model.eval()

    path = Path(args.input)
    if path.is_file():
        text = transcribe_file(
            str(path), model, processor,
            language=args.language, task=args.task, timestamps=args.timestamps
        )
        print(f"{path.name} -> {text}")
    else:
        # 目录批量
        exts = {e.strip().lower() for e in args.batch_exts.split(",")}
        files = [p for p in path.rglob("*") if p.suffix.lower() in exts]
        if not files:
            print("目录中未找到可处理的音频文件。")
            return
        for p in sorted(files):
            try:
                t0 = time.time()
                text = transcribe_file(
                    str(p), model, processor,
                    language=args.language, task=args.task, timestamps=args.timestamps
                )
                t1 = time.time()
                print(f"{p.name} -> {text}; time cost: {t1-t0}")
            except Exception as e:
                print(f"{p.name} -> 失败: {e}")

def load_model():
    # model_path = "/Users/jeqin/Downloads/checkpoint-39000-full/whisper-large-v3-turbo-finetune"
    model_path = "/Users/jeqin/Downloads/whisper-large-v3-turbo-finetune_1219"
    lang = "zh"
    t0 = time.time()
    processor = WhisperProcessor.from_pretrained(
        model_path,
        language=lang,
        task="transcribe",
        no_timestamps=True,
        local_files_only=True,
    )
    model = WhisperForConditionalGeneration.from_pretrained(
        model_path,
        device_map="mps",
        local_files_only=True,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    )

    model.generation_config.language = lang.lower()
    model.generation_config.forced_decoder_ids = None
    model.eval()
    print("load model time: ", time.time() - t0)
    return model, processor

def run_test_audios():
    model, processor = load_model()
    audios = Path("../test_data/audio_clips/")
    rows = [["file_name", "inference_time", "inference_result"]]
    for audio in sorted(audios.glob("*en-ac1-16k/*.wav")): # *s/randomforest*.wav"
        try:
            t0 = time.time()
            text = transcribe_file(
                str(audio), model, processor
            )

            t = time.time()-t0
            print(f"{audio.name} -> {text}; time cost: {t}")
            rows.append([f"{audio.parent.name}/{audio.name}", t, text])
        except Exception as e:
            print(f"{audio.name} -> 失败: {e}")
    save_csv("csv/fine-tune_whisper-0901.csv", rows)

def run_recordings():
    from scripts.asr_utils import get_origin_text_dict, get_text_distance
    model, processor = load_model()
    audios = Path("../test_data/recordings/")
    rows = [["file_name", "time", "inference_result"]]
    original = get_origin_text_dict()
    for audio in sorted(audios.glob("*.wav"), key=lambda x: int(x.stem)):
        print(audio)
        try:
            t0 = time.time()
            text = transcribe_file(
                str(audio), model, processor
            )
            t = time.time()-t0
            print(text)
            print("inference time:", t)
            d, nd, diff = get_text_distance(original[audio.stem], text)
            rows.append([audio.name, round(t, 3), text, d, round(nd,3), diff])
        except Exception as e:
            print(f"{audio.name} -> 失败: {e}")
    save_csv("csv/fine-tune_whisper.csv", rows)


def run_test_dataset():
    from test_data.audios import read_dataset
    model, processor = load_model()
    test_data = Path("../test_data/dataset/dataset.txt")
    audio_parent = Path("../test_data/")
    rows = [["file_name", "time", "inference_result"]]
    result_list = []
    count = 0
    try:
        for audio_path, sentence, duration in read_dataset(test_data):
            count += 1
            print(f"processing {count}: {audio_path}")

            t1 = time.time()
            text = transcribe_file(
                str(audio_parent/audio_path), model, processor
            )
            t = time.time() - t1
            print("inference time:", t)
            print(text)
            result_list.append({
                "index": count,
                "audio_path": audio_path,
                "reference": sentence,
                "duration": duration,
                "inference_time": round(t, 3),
                "inference_result": text
            })
    except Exception as e:
        print(e)
    except KeyboardInterrupt as e:
        print(e)
    import json
    with open("csv/whisper_finetuned_dataset_results.json", "w", encoding="utf-8") as f:
        json.dump(result_list, f, ensure_ascii=False, indent=2)

def run_test_emilia():
    from test_data.audios import read_emilia
    model, processor = load_model()
    parent = Path("../test_data/ZH-B000008")
    result_list = []
    count = 0
    try:
        for audio_path, sentence, duration in read_emilia(parent, count_limit=5000):
            count += 1
            print(f"processing {count}: {audio_path}")

            t1 = time.time()
            text = transcribe_file(
                str(audio_path), model, processor
            )
            t = time.time() - t1
            print("inference time:", t)
            print(text)
            result_list.append({
                "index": count,
                "audio_path": audio_path.name,
                "reference": sentence,
                "duration": duration,
                "inference_time": round(t, 3),
                "inference_result": text
            })
    except Exception as e:
        print(e)
    except KeyboardInterrupt as e:
        print(e)
    import json
    with open("csv/whisper_finetune_emilia_results.json", "w", encoding="utf-8") as f:
        json.dump(result_list, f, ensure_ascii=False, indent=2)


def run_test_st():
    from test_data.audios import read_st
    model, processor = load_model()
    # parent = Path("../test_data/ST-CMDS-20170001_1-OS")
    result_list = []
    count = 0
    try:
        for audio_path, sentence in read_st(count_limit=5000):
            count += 1
            print(f"processing {count}: {audio_path}")

            t1 = time.time()
            text = transcribe_file(
                str(audio_path), model, processor
            )
            t = time.time() - t1
            print("inference time:", t)
            print(text)
            result_list.append({
                "index": count,
                "audio_path": audio_path.name,
                "reference": sentence,
                # "duration": duration,
                "inference_time": round(t, 3),
                "inference_result": text
            })
    except Exception as e:
        print(e)
    except KeyboardInterrupt as e:
        print(e)
    import json
    with open("csv/whisper_finetune_st_results.json", "w", encoding="utf-8") as f:
        json.dump(result_list, f, ensure_ascii=False, indent=2)

def run_test_wenet():
    from test_data.audios import read_wenet
    model, processor = load_model()
    result_list = []
    count = 0
    try:
        for audio_path, sentence in read_wenet(count_limit=5000):
            count += 1
            print(f"processing {count}: {audio_path}")

            t1 = time.time()
            text = transcribe_file(
                str(audio_path), model, processor
            )
            t = time.time() - t1
            print("inference time:", t)
            print(text)
            result_list.append({
                "index": count,
                "audio_path": audio_path.name,
                "reference": sentence,
                # "duration": duration,
                "inference_time": round(t, 3),
                "inference_result": text
            })
    except Exception as e:
        print(e)
    except KeyboardInterrupt as e:
        print(e)
    import json
    with open("csv/whisper_finetune_wenet_results.json", "w", encoding="utf-8") as f:
        json.dump(result_list, f, ensure_ascii=False, indent=2)

if __name__ == "__main__":
    # main()
    # run_recordings()
    # run_test_dataset()
    # run_test_emilia()
    run_test_wenet()