app.py

# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Liu Yue)
import os
import sys
import argparse
import gradio as gr
import numpy as np
import torch
import torchaudio
import random
import librosa
from funasr import AutoModel
from funasr.utils.postprocess_utils import rich_transcription_postprocess

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

from modelscope import snapshot_download, HubApi
from huggingface_hub import snapshot_download as hf_snapshot_download

# 模型下载部分保持不变
hf_snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')
snapshot_download('iic/SenseVoiceSmall', local_dir='pretrained_models/SenseVoiceSmall')
hf_snapshot_download('FunAudioLLM/CosyVoice-ttsfrd', local_dir='pretrained_models/CosyVoice-ttsfrd')

# 注意：ttsfrd 在某些 CPU 环境下可能需要特定编译，这里保留原样
os.system(
    "cd pretrained_models/CosyVoice-ttsfrd/ && "
    "pip install ttsfrd_dependency-0.1-py3-none-any.whl && "
    "pip install ttsfrd-0.4.2-cp310-cp310-linux_x86_64.whl && "
    "apt install -y unzip && "
    "rm -rf resource && "
    "unzip resource.zip -d ."
)

from cosyvoice.cli.cosyvoice import AutoModel as CosyVoiceAutoModel
from cosyvoice.utils.file_utils import logging, load_wav
from cosyvoice.utils.common import set_all_random_seed, instruct_list

# -----------------------------
# i18n & UI 配置 (省略部分重复文本以节省篇幅，逻辑保持不变)
# -----------------------------
LANG_EN, LANG_ZH = "En", "Zh"
MODE_ZERO_SHOT, MODE_INSTRUCT = "zero_shot", "instruct"
# ... [此处省略 UI_TEXT 定义，与原代码一致] ...

def t(lang: str, key: str) -> str:
    lang = lang if lang in UI_TEXT else LANG_ZH
    return UI_TEXT[lang][key]

def mode_choices(lang: str):
    return [(t(lang, "mode_zero_shot"), MODE_ZERO_SHOT), (t(lang, "mode_instruct"), MODE_INSTRUCT)]

def steps_for(lang: str, mode_value: str) -> str:
    return t(lang, "steps_instruct") if mode_value == MODE_INSTRUCT else t(lang, "steps_zero_shot")

# -----------------------------
# 音频处理
# -----------------------------
max_val, top_db, hop_length, win_length = 0.8, 60, 220, 440

def generate_seed():
    return {"__type__": "update", "value": random.randint(1, 100000000)}

def postprocess(wav):
    speech = load_wav(wav, target_sr=target_sr, min_sr=16000)
    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)
    torchaudio.save(wav, speech, target_sr)
    return wav

# --- 修改点 1: 移除 @spaces.GPU 装饰器 ---
def prompt_wav_recognition(prompt_wav):
    if prompt_wav is None: return ""
    res = asr_model.generate(
        input=prompt_wav,
        language="auto",
        use_itn=True,
    )
    text = res[0]["text"].split("|>")[-1]
    return text

# --- 修改点 2: 移除 @spaces.GPU 装饰器 ---
def generate_audio(tts_text, mode_value, prompt_text, prompt_wav_upload, prompt_wav_record, instruct_text, seed, stream, ui_lang):
    stream = False # CPU 模式下流式传输可能较慢，建议维持 False
    if len(tts_text) > 200:
        gr.Warning(t(ui_lang, "warn_too_long"))
        return (target_sr, default_data)

    speed = 1.0
    prompt_wav = prompt_wav_upload if prompt_wav_upload is not None else prompt_wav_record

    # 校验逻辑保持不变...
    if mode_value == MODE_INSTRUCT:
        if instruct_text == "" or prompt_wav is None:
            return (target_sr, default_data)
    elif mode_value == MODE_ZERO_SHOT:
        if prompt_wav is None or prompt_text == "":
            return (target_sr, default_data)

    if mode_value == MODE_ZERO_SHOT:
        set_all_random_seed(seed)
        speech_list = []
        # CosyVoice 内部会自动处理 torch.device
        for i in cosyvoice.inference_zero_shot(
            tts_text, "You are a helpful assistant.<|endofprompt|>" + prompt_text,
            postprocess(prompt_wav), stream=stream, speed=speed
        ):
            speech_list.append(i["tts_speech"])
        return (target_sr, torch.concat(speech_list, dim=1).numpy().flatten())

    if mode_value == MODE_INSTRUCT:
        set_all_random_seed(seed)
        speech_list = []
        for i in cosyvoice.inference_instruct2(
            tts_text, instruct_text, postprocess(prompt_wav), stream=stream, speed=speed
        ):
            speech_list.append(i["tts_speech"])
        return (target_sr, torch.concat(speech_list, dim=1).numpy().flatten())

    return (target_sr, default_data)

# ... [此处省略 UI 回调函数 on_mode_change, on_language_change 和 main() 定义，与原代码一致] ...

if __name__ == "__main__":
    # --- 修改点 3: 确保 CosyVoice 加载在 CPU ---
    cosyvoice = CosyVoiceAutoModel(
        model_dir="pretrained_models/Fun-CosyVoice3-0.5B",
        load_trt=False, # CPU 不支持 TensorRT
        fp16=False,     # CPU 必须关闭 FP16
    )
    
    prompt_sr = 16000
    target_sr = 24000
    default_data = np.zeros(target_sr)

    # --- 修改点 4: SenseVoice ASR 模型强制使用 CPU ---
    model_dir = "pretrained_models/SenseVoiceSmall"
    asr_model = AutoModel(
        model=model_dir,
        disable_update=True,
        log_level="DEBUG",
        device="cpu", # 明确指定为 cpu
    )

    # 启动 Gradio
    # main() 函数包含 demo 定义
    from main_logic import main # 假设 main 在当前脚本或调整后的结构中
    main()