CosyVoice2-0.5B在Windows下本地完全部署、最小化部署

import os
import time

import torch
import torchaudio
from cosyvoice.cli.cosyvoice import CosyVoice2
from cosyvoice.utils.file_utils import load_wav
from tqdm import tqdm

# 初始化CosyVoice2模型，指定预训练模型路径，不加载jit和trt模型，使用fp32
cosyvoice = CosyVoice2('pretrained_models/CosyVoice2-0.5B',
                       load_jit=True, load_trt=False, fp16=True)

# 设置最大音量
max_val = 1.0
# 设置说话人信息文件的路径
spk2info_path = f'spk2info.pt'


# 如果说话人信息文件存在，则加载
if os.path.exists(spk2info_path):
    spk2info = torch.load(
        spk2info_path, map_location=cosyvoice.frontend.device)
else:
    spk2info = {}


def load_speaker_info(speaker, retrain=False):
    load_start = time.time()
    if retrain:
        if speaker in spk2info:
            del spk2info[speaker]
    if speaker not in spk2info:
        prompt_speech_16k = load_wav(f'prompt/{speaker}.wav', 16000)
        print(f'Extracting speaker info for {speaker}...')
        train_start = time.time()
        # 获取音色embedding
        embedding = cosyvoice.frontend._extract_spk_embedding(
            prompt_speech_16k)
        # 获取语音特征
        prompt_speech_resample = torchaudio.transforms.Resample(
            orig_freq=16000, new_freq=cosyvoice.sample_rate)(prompt_speech_16k)
        speech_feat, speech_feat_len = cosyvoice.frontend._extract_speech_feat(
            prompt_speech_resample)
        # 获取语音token
        speech_token, speech_token_len = cosyvoice.frontend._extract_speech_token(
            prompt_speech_16k)
        # 将音色embedding、语音特征和语音token保存到字典中
        spk2info[speaker] = {'embedding': embedding,
                             'speech_feat': speech_feat, 'speech_token': speech_token}
        # 保存音色embedding
        torch.save(spk2info, spk2info_path)
        print(
            f'Extracting speaker info for {speaker} done!', 'Time:', time.time()-train_start)
    print('Load time:', time.time()-load_start)


# 定义一个文本到语音的函数，参数包括文本内容、是否流式处理、语速和是否使用文本前端处理
def tts_sft(tts_text: str, prompt_text: str, speaker_info: dict, stream=False, speed=1.0, text_frontend=True):
    '''
    参数：
        tts_text：要合成的文本
        prompt_text: 提示文本
        speaker：说话人音频特征
        stream：是否流式处理
        speed：语速
        text_frontend：是否使用文本前端处理

    返回值：
        合成后的音频
    '''

    # 使用tqdm库来显示进度条，对文本进行标准化处理并分割
    for i in tqdm(cosyvoice.frontend.text_normalize(tts_text, split=True, text_frontend=text_frontend)):
        # 提取文本的token和长度
        tts_text_token, tts_text_token_len = cosyvoice.frontend._extract_text_token(
            i)
        # 提取提示文本的token和长度
        prompt_text_token, prompt_text_token_len = cosyvoice.frontend._extract_text_token(
            prompt_text)
        # 获取说话人的语音token长度，并转换为torch张量，移动到指定设备
        speech_token_len = torch.tensor(
            [speaker_info['speech_token'].shape[1]], dtype=torch.int32).to(cosyvoice.frontend.device)
        # 获取说话人的语音特征长度，并转换为torch张量，移动到指定设备
        speech_feat_len = torch.tensor(
            [speaker_info['speech_feat'].shape[1]], dtype=torch.int32).to(cosyvoice.frontend.device)
        # 构建模型输入字典，包括文本、文本长度、提示文本、提示文本长度、LLM提示语音token、LLM提示语音token长度、流提示语音token、流提示语音token长度、提示语音特征、提示语音特征长度、LLM嵌入和流嵌入
        model_input = {'text': tts_text_token, 'text_len': tts_text_token_len,
                       'prompt_text': prompt_text_token, 'prompt_text_len': prompt_text_token_len,
                       'llm_prompt_speech_token': speaker_info['speech_token'], 'llm_prompt_speech_token_len': speech_token_len,
                       'flow_prompt_speech_token': speaker_info['speech_token'], 'flow_prompt_speech_token_len': speech_token_len,
                       'prompt_speech_feat': speaker_info['speech_feat'], 'prompt_speech_feat_len': speech_feat_len,
                       'llm_embedding': speaker_info['embedding'], 'flow_embedding': speaker_info['embedding']}
        # 使用模型进行文本到语音的转换，并迭代输出结果
        for model_output in cosyvoice.model.tts(**model_input, stream=stream, speed=speed):
            yield model_output


def synthesis(tts_text: str, speaker: str):  # 生成语音
    # 获取说话人信息
    prompt_text = open(f'prompt/{speaker}.txt', 'r', encoding='utf-8').read()
    # 遍历每个文本的生成结果
    for i, j in enumerate(tts_sft(tts_text, prompt_text, speaker_info=spk2info[speaker], stream=False, speed=1.0, text_frontend=True)):
        # 保存生成的语音到文件，文件名包含文本的前四个字符
        torchaudio.save(f"output/{speaker}_{tts_text[0:4]}_{time.strftime('%Y%m%d%H%M%S', time.localtime())}.wav", j['tts_speech'], cosyvoice.sample_rate)
    print("生成完成！")


speaker = ""

while True:
    os.system("cls")
    spk2info_list = "，".join(list(spk2info.keys()))
    print(f"当前音色：{speaker}") if speaker else print("当前未选择音色")
    print("""
1. 查看可用音色
2. 加载音色
3. 生成语音
4. 删除音色
5. 退出
""")
    try:
        choice = int(input("请输入选项："))
        if choice == 1:  # 查看可用音色
            if spk2info_list:
                print("可用音色：", spk2info_list)
                time.sleep(3)
            else:
                print("当前没有可用的音色")
                time.sleep(2)
            continue

        elif choice == 2:  # 加载音色
            os.system("cls")
            speaker_name = input("请输入要加载的音色名称：")
            if speaker_name not in spk2info.keys():
                retrain = input("音色不存在，是否训练？[Y/N]")
                if "y" in retrain or "Y" in retrain:
                    load_speaker_info(speaker_name)
                    speaker = speaker_name
                elif "n" in retrain or "N" in retrain:
                    continue
                else:
                    print("输入错误，请重新输入")
                    time.sleep(1)
                    continue
            else:
                speaker = speaker_name
                print("音色加载成功！")
                time.sleep(1)

        elif choice == 3:  # 生成语音
            if not speaker:
                print("请先选择音色")
                time.sleep(1)
                continue
            os.system("cls")
            print("当前音色：", speaker)
            text = input("请输入要合成的文本：")
            synthesis(text, speaker)

        elif choice == 4:  # 删除音色
            os.system("cls")
            delete_speaker = input("请输入要删除的音色名称：")
            if delete_speaker not in spk2info.keys():
                print("音色不存在")
                time.sleep(1)
                continue
            else:
                spk2info.pop(delete_speaker)
                torch.save(spk2info, "spk2info.pt")
                print("音色删除成功！")
                time.sleep(1)

        elif choice == 5:
            break
        else:
            print("输入错误，请重新输入")
            time.sleep(1)
    except Exception as e:
        print("输入错误，请重新输入")
        print(e)
        time.sleep(1)
print("程序即将结束..")

from pydub import AudioSegment
import os

# 扫描指定目录下的所有mp3文件
convert_list = []
for root, dirs, files in os.walk("prompt"):
    for file in files:
        if file.endswith(".mp3"):
            convert_list.append(os.path.join(root, file))

# 是否删除源文件？
delete_mp3 = True

mp3_file:str
for mp3_file in convert_list:
    sound = AudioSegment.from_mp3(mp3_file)
    sound.export(mp3_file.lower().replace(".mp3", ".wav"),format="wav")
    if delete_mp3:
        os.remove(mp3_file)
print("转换完成")