app.py

import json
import requests

from datetime import datetime

import time
import traceback

API_URL = "https://api-inference.huggingface.co/models/"


def date_now():
    return datetime.now().strftime("%Y-%m-%d %H:%M:%S")

def record_opt(msg):
    return f"{date_now()} {msg}\n"


def speech_recognize(audio, model_name, hf_token, opt):
    opt += record_opt("转录开始 ...")
    yield "转录中，请稍等...", opt
    start = time.monotonic()

    with open(audio, "rb") as f:
        data = f.read()
    try:
        url = API_URL + model_name
        print(f">>> url is {url}")
        headers = {"Authorization": f"Bearer {hf_token}"}
        response = requests.request("POST", url, headers=headers, data=data)
        text = json.loads(response.content.decode("utf-8"))
        print(f">>> text is {text}")
        text = text['text']
    except:
        text = f"转录失败：\n{traceback.format_exc()}"

    cost = time.monotonic() - start
    opt += record_opt(f"转录结束，耗时{cost:.3f}s")
    yield text, opt

import gradio as gr

with gr.Blocks() as demo:
    gr.HTML("""<h2 align="center">Automatic Speech Recognition (OpenAI Whisper with Inference API)</h2>""")
    with gr.Row():
        gr.Markdown(
            """🤗 调用 huggingface API，使用 OpenAI Whisper 模型进行语音识别，也可以称为语音转文本（Speech to Text, STT）

            👉 目的是练习使用 Gradio Audio 组件和探索使用 Huggingface Inference API

            > 💡提示：需要填写 Huggingface token 来调用 Huggingface Inference API
            """
        )
    with gr.Row():
        with gr.Column():
            audio = gr.Audio(source="microphone", type="filepath")
            model_name = gr.Dropdown(
                label="选择模型",
                choices=[
                    "openai/whisper-large-v3",
                    "openai/whisper-large-v2",
                    "openai/whisper-large",
                    "openai/whisper-medium",
                    "openai/whisper-small",
                    "openai/whisper-base",
                    "openai/whisper-tiny",
                ],
                value="openai/whisper-large-v2",
            )
            hf_token = gr.Textbox(label="Huggingface token")
        with gr.Column():
            output = gr.Textbox(label="转录结果")
            operation = gr.Textbox(label="组件操作历史")
    audio.start_recording(
        lambda x: x + record_opt("开始录音 ..."),
        inputs=operation, outputs=operation
    )
    audio.play(
        lambda x: x + record_opt("播放录音"),
        inputs=operation, outputs=operation
    )
    audio.pause(
        lambda x: x + record_opt("暂停播放"),
        inputs=operation, outputs=operation
    )
    audio.stop(
        lambda x: x + record_opt("停止播放"),
        inputs=operation, outputs=operation
    )
    audio.end(
        lambda x: x + record_opt("播放完毕"),
        inputs=operation, outputs=operation
    )
    audio.stop_recording(speech_recognize, inputs=[audio, model_name, hf_token, operation], outputs=[output, operation])

demo.queue(max_size=4, concurrency_count=4)
demo.launch()