app.py

import torch
import gradio as gr
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline



# 加载模型和处理器
def load_model(model_path, use_gpu=True, use_flash_attention_2=False, use_bettertransformer=False):
    device = "cuda:0" if torch.cuda.is_available() and use_gpu else "cpu"
    torch_dtype = torch.float16 if torch.cuda.is_available() and use_gpu else torch.float32

    processor = AutoProcessor.from_pretrained(model_path)
    model = AutoModelForSpeechSeq2Seq.from_pretrained(
        model_path, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True,
        use_flash_attention_2=use_flash_attention_2
    )
    if use_bettertransformer and not use_flash_attention_2:
        model = model.to_bettertransformer()


    model.to(device)

    return processor, model, device, torch_dtype

# 初始化模型
processor, model, device, torch_dtype = load_model(
    model_path=r"panlr/whisper-finetune-teochew",
    use_gpu=True,
    use_flash_attention_2=False,
    use_bettertransformer=False
)

# 创建推理管道
infer_pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=30,
    batch_size=16,
    torch_dtype=torch_dtype,
    device=device
)

# 定义推理函数
def transcribe_audio(audio_path, num_beams=1):
    generate_kwargs = {"num_beams": num_beams}

    forced_decoder_ids = [
        (1, processor.tokenizer.encode("<|startoftranscript|>")[0]),
        (2, processor.tokenizer.encode("<|zh|>")[0]),
        (3, processor.tokenizer.encode("<|transcribe|>")[0]),
    ]
    model.generation_config.forced_decoder_ids = forced_decoder_ids
    # if language is not None:
    #     generate_kwargs["language"] = language

    result = infer_pipe(audio_path, return_timestamps=False, generate_kwargs=generate_kwargs)

    return result['text']

# Gradio 界面
def gradio_interface(audio):
    return transcribe_audio(audio)



# 创建 Gradio 界面
with gr.Blocks() as interface:
    gr.Markdown("## Whisper 潮汕话-正字 语音转录")
    
    audio_input = gr.Audio(
        sources=["microphone", "upload"], 
        type="filepath", 
        label="输入音频",
        value="./example.wav"  # 指定默认音频文件
    )
    
    output_text = gr.Textbox(label="转录结果")

    # 在输入模块的下方添加说明
    gr.Markdown("""
        📢 **使用说明**
        - 本demo部署在CPU上，所以推理速度较慢。对于比较书面的话语，识别效果还不错，对土话、俗话还需要更多的数据。
        - 你可以 **上传音频文件** 或 **使用麦克风** 向模型输入。  
        - 音频文件最好发音清晰、标准。  
        - 默认提供一个示例音频，你可以直接点击“提交”查看转录效果。
        - 示例音频的对应文本： 【状元 林大钦，兵部尚(siên7)书 翁万达，了佮 工部 左侍郎(se6 neng5) 陈一松，拢是 嘉靖 年间 介 进士】 
    """)

    submit_btn = gr.Button("提交")
    submit_btn.click(gradio_interface, inputs=audio_input, outputs=output_text)

# 启动 Gradio 应用
interface.launch()