app.py

import sys
import os

# 相当于 export PYTHONPATH=$PWD/:$PYTHONPATH
sys.path.append(os.getcwd())

# 如果还需要 PATH
os.environ["PATH"] = os.path.join(os.getcwd(), "fireredasr") + ":" + \
                     os.path.join(os.getcwd(), "fireredasr", "utils") + ":" + \
                     os.environ["PATH"]

import gradio as gr
import librosa
import soundfile as sf
from fireredasr.models.fireredasr import FireRedAsr

import os
from huggingface_hub import snapshot_download

model_dir = "pretrained_models/FireRedASR-AED-L"
if not os.path.exists(model_dir):
    print("Downloading FireRedASR-AED-L from Hugging Face...")
    snapshot_download(
        repo_id="FireRedTeam/FireRedASR-AED-L",
        local_dir=model_dir,
        local_dir_use_symlinks=False
    )

# 1. 加载模型（只加载一次）
model = FireRedAsr.from_pretrained(
    "aed",
    "pretrained_models/FireRedASR-AED-L"
)


def preprocess_audio(input_path):
    """
    将音频重采样到 16kHz 单声道并保存到临时文件
    """
    audio, sr = librosa.load(input_path, sr=16000, mono=True)
    tmp_path = input_path + "_16k.wav"
    sf.write(tmp_path, audio, 16000)
    return tmp_path


def transcribe_fn(audio_file):
    """
    audio_file: Gradio 会返回音频的临时文件路径
    """
    if audio_file is None:
        return "请先录音或上传音频"

    # 预处理（重采样）
    processed_path = preprocess_audio(audio_file)

    batch_uttid = ["audio_input"]
    batch_wav_path = [processed_path]

    results = model.transcribe(
        batch_uttid,
        batch_wav_path,
        {
            "use_gpu": 0,  # Spaces CPU = 0, GPU Space 可改成 1
            "beam_size": 3,
            "nbest": 1,
            "decode_max_len": 0,
            "softmax_smoothing": 1.0,
            "aed_length_penalty": 0.0,
            "eos_penalty": 1.0
        }
    )

    if isinstance(results, list):
        return results[0].get("text", str(results[0]))
    else:
        return str(results)


# 2. Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# 🔊 FireRedASR 实时语音识别 Demo\n支持麦克风实时识别 + 文件上传")

    with gr.Tab("📁 文件上传识别或者实时麦克风录音"):
        gr.Markdown("上传音频文件或者实时录音，点击按钮进行识别。")
        file_audio = gr.Audio(sources=["upload"], type="filepath", label="上传音频文件（支持 WAV/MP3 等）")
        file_output = gr.Textbox(label="识别结果")
        file_btn = gr.Button("开始识别")
        file_btn.click(fn=transcribe_fn, inputs=file_audio, outputs=file_output)

if __name__ == "__main__":
    demo.launch()