app.py

import os
import sys
import gradio as gr
import numpy as np
import torch
import librosa
import torchaudio
from scipy.signal import resample
import time
import requests

from huggingface_hub import snapshot_download

ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(f'{ROOT_DIR}/third_party/Matcha-TTS')

from cosyvoice.cli.cosyvoice import CosyVoice
from cosyvoice.utils.file_utils import load_wav

def download_audio_from_drive(file_id, save_path="temp_prompt.wav"):
    url = f"https://drive.google.com/uc?export=download&id={file_id}"
    response = requests.get(url)
    if not response.ok or b"<html" in response.content[:100]:
        raise RuntimeError("Google Drive 音檔下載失敗")
    with open(save_path, "wb") as f:
        f.write(response.content)
    return save_path

preset_speakers = {
    "6歲": {
        "file_id": "1PomN05sTjxXBR3q5kOgB7CUEpig2HJmx",
        "transcription": "名字是微笑號，只是呢你們看，這一輛微笑號它這裡有寫八百型的喔，它是八百山出來"
    },
    "8歲": {
        "file_id": "1U1DKzN_BsU-DL02HZIRlx-opUJLhaQ41",
        "transcription": "出來了出來了，你知道這個餐具是可以挖的，這個餐具可以用窩課魅一起挖嗎?"
    }
}

def apply_preset(speaker_key):
    if speaker_key in preset_speakers:
        file_id = preset_speakers[speaker_key]["file_id"]
        transcription = preset_speakers[speaker_key]["transcription"]
        local_path = download_audio_from_drive(file_id)
        return local_path, transcription
    return None, ""

max_val = 0.8
target_sr = 22050
prompt_sr = 16000

def download_with_retry(repo_id, max_retries=10, wait_sec=5):
    for attempt in range(1, max_retries + 1):
        try:
            print(f"正在下載模型（第 {attempt} 次）")
            model_dir = snapshot_download(repo_id)
            print("模型下載成功")
            return model_dir
        except Exception as e:
            print(f"第 {attempt} 次下載失敗：{e}")
            if attempt < max_retries:
                print(f"等待 {wait_sec} 秒後重試")
                time.sleep(wait_sec)
            else:
                print("已達最大重試次數")
                raise e

model_dir = download_with_retry("MediaTek-Research/BreezyVoice")
cosyvoice = CosyVoice(model_dir)

def postprocess(speech, top_db=60, hop_length=220, win_length=440):
    speech, _ = librosa.effects.trim(
        speech, top_db=top_db,
        frame_length=win_length,
        hop_length=hop_length
    )
    if speech.abs().max() > max_val:
        speech = speech / speech.abs().max() * max_val
    speech = torch.concat([speech, torch.zeros(1, int(target_sr * 0.2))], dim=1)
    return speech

def synthesize_speech(speaker_audio_path, content_text, speaker_text, speaker_key):
    if speaker_key in preset_speakers and speaker_key is not None:
        prompt_wav_path = speaker_audio_path
        if not speaker_text.strip():
            speaker_text = preset_speakers[speaker_key]["transcription"]
    else:
        prompt_wav_path = speaker_audio_path

    if not prompt_wav_path:
        raise gr.Error("請上傳或選擇語音樣本")
    if not speaker_text or len(speaker_text.strip()) < 5:
        raise gr.Error("語音樣本的轉寫內容太短，至少輸入5字以上")

    prompt_speech_16k = postprocess(load_wav(prompt_wav_path, prompt_sr))
    output = cosyvoice.inference_zero_shot(content_text, speaker_text, prompt_speech_16k)

    speed_factor = 1.0
    if speed_factor != 1.0:
        new_length = int(len(output['tts_speech']) / speed_factor)
        audio_data = resample(output['tts_speech'], new_length)
    else:
        audio_data = output['tts_speech'].numpy().flatten()

    os.makedirs("results", exist_ok=True)
    out_path = "results/output.wav"
    torchaudio.save(out_path, torch.tensor(audio_data).unsqueeze(0), sample_rate=target_sr)
    return out_path

with gr.Blocks() as demo:
    gr.Markdown("""
    # 小睿語音合成
    選擇預設語音 (6歲、8歲) 或上傳5~15秒語音樣本及該則語音樣本的轉寫，並輸入要合成的句子。
    視該時段運算資源和語句長度而定約需70至2000秒，若等候時間過長，可改用較快但較不穩的版本：
    https://colab.research.google.com/drive/1aQSLwzbK9QnNMH4A5eVfqFYfPFXHwlXx?usp=sharing
    """)

    speaker_selector = gr.Dropdown(
        label="選擇語音樣本（可選）",
        choices=list(preset_speakers.keys()),
        value=None,
        interactive=True
    )

    with gr.Row():
        audio_input = gr.Audio(label="上傳或套用語音樣本", type="filepath", interactive=True)
        content_input = gr.Textbox(label="要合成的文字句子", placeholder="例如：今天天氣真好")
        transcription_input = gr.Textbox(
            label="語音樣本的轉寫（必填）",
            placeholder="例如：只是呢你們看，這一輛微笑號它這裡有寫八百型的喔...",
            lines=3
        )

    output_audio = gr.Audio(label="合成結果", type="filepath")
    btn = gr.Button("開始語音合成")

    speaker_selector.change(
        fn=apply_preset,
        inputs=speaker_selector,
        outputs=[audio_input, transcription_input]
    )

    btn.click(
        fn=synthesize_speech,
        inputs=[audio_input, content_input, transcription_input, speaker_selector],
        outputs=output_audio
    )

    gr.Markdown("""
    為了加速，已關閉自動語音辨識，務必正確輸入語音樣本的文字轉寫。
    """)

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