app.py

import gradio as gr
import os
import tempfile
import subprocess
from transformers import pipeline
import torch
from zipfile import ZipFile

from fastapi import FastAPI

app = FastAPI()

device = "cuda:0" if torch.cuda.is_available() else "cpu"

if torch.cuda.is_available():
    model_id = "openai/whisper-small.en"
else:
    model_id = "openai/whisper-tiny.en"

pipe = pipeline(
    "automatic-speech-recognition",
    model=model_id,
    chunk_length_s=30,
    device=device,
)


def support_gbk(zip_file: ZipFile):
    name_to_info = zip_file.NameToInfo
    # copy map first
    for name, info in name_to_info.copy().items():
        real_name = name.encode("cp437").decode("gbk")
        if real_name != name:
            info.filename = real_name
            del name_to_info[name]
            name_to_info[real_name] = info
    return zip_file


def handel(f):
    if not f:
        raise gr.Error("请上传文件")
    if f.name.endswith(".zip"):
        with support_gbk(ZipFile(f.name, "r")) as z:
            dir = tempfile.TemporaryDirectory()
            z.extractall(path=dir.name)
            return handel_files(
                [
                    os.path.join(filepath, filename)
                    for filepath, _, filenames in os.walk(dir.name)
                    for filename in filenames
                ]
            )
    else:
        return handel_files([f.name])


def ffmpeg_convert(file_input, file_output):
    if subprocess.run(["ffmpeg", "-y", "-i", file_input, file_output]).returncode:
        raise gr.Error("ffmpeg_convert 失败, 请检查文件格式是否正确")


def handel_files(f_ls):
    files = []
    for file in f_ls:
        
        file_output=None
        if file.endswith(".m4a"):
            file_output = file.replace(".m4a", ".wav")
            ffmpeg_convert(file, file_output)
        elif file.endswith(".mp3"):
            file_output = file.replace(".mp3", ".wav")
            ffmpeg_convert(file, file_output)
        elif file.endswith(".wav"):
            # check wav file is valid or not
            file_output = file+".wav"
            ffmpeg_convert(file, file_output)
            
        if file_output:
            files.append(file_output)
        else:
            gr.Warning(f"存在不合法文件{os.path.basename(file)},已跳过处理")
            
    ret = []
    for file in files:
        ret.append(whisper_handler(file))

    return "\n\n".join(ret)


def whisper_handler(file):
    file_name = os.path.basename(file)
    gr.Info(f"处理文件 - {file_name.split('.')[0]}")
    return pipe(file)["text"]


with gr.Blocks() as blocks:
    f = gr.File(file_types=[".zip", ".mp3", ".wav", ".m4a"])
    b = gr.Button(value="提交")
    t = gr.Textbox(label="结果")

    b.click(handel, inputs=f, outputs=t)

blocks.queue(max_size=3)

app = gr.mount_gradio_app(app, blocks, path="/")