example.py

import sys
sys.path.append('third_party/Matcha-TTS')
from cosyvoice.cli.cosyvoice import AutoModel
import torchaudio
from cosyvoice.utils.file_utils import load_wav


def inference_contextspeech_onesample_test(cosymodel, tts_text, prompt_speech, llm_prompt_speech, stream=False, speed=1.0, text_frontend=True):
        tts_text = cosymodel.frontend.text_normalize(tts_text, split=False, text_frontend=text_frontend)
        tts_text_token, tts_text_token_len = cosymodel.frontend._extract_text_token(tts_text)
        flow_embedding = cosymodel.frontend._extract_spk_embedding(prompt_speech)
        embedding = cosymodel.frontend._extract_spk_embedding(llm_prompt_speech)
        model_input = {'text': tts_text_token, 'text_len': tts_text_token_len, 'llm_embedding': embedding, 'flow_embedding': flow_embedding}
        print('synthesis text {}'.format(tts_text))
        for model_output in cosymodel.model.tts(**model_input, stream=stream, speed=speed):
            speech_len = model_output['tts_speech'].shape[1] / cosymodel.sample_rate
            yield model_output

import shutil

def cosyvoice2_example():
    """ CosyVoice2 Usage, check https://funaudiollm.github.io/cosyvoice2/ for more details
    """

    transcription = "到哪都是坐，一下车被人打断双腿，你觉得值得吗？"
    context_description = "他正被一个陌生人以暴力威胁要求换座位，对方意图不轨。"
    personal_experience = "他过去多次被亲近的人以类似方式戏弄和考验，习惯了在这种局面下保持镇定。"
    emotions = ["讽刺", "冷静"]
    paralinguistic_description = "用慢悠悠的语调带着嘲弄意味地说，中间有多次停顿。"

    text = ''

    # shutil.copy2("person_context_para_emotion_llm.pt", "pretrained_models/CosyVoice2-0.5B_cetts/llm.pt")
    # shutil.copy2("person_context_dpsk_para_emotion_llm.pt", "pretrained_models/CosyVoice2-0.5B_cetts/llm.pt")
    # text += f"角色之前经历过:{personal_experience}"
    # text += f"角色现在所处场景:{context_description}"
    # text += f"{paralinguistic_description}"
    # text += f"请你模仿这个角色,用{','.join(emotions)}的语气说话。<|endofprompt|>"
    # text += transcription

    # shutil.copy2("person_context_emotion_llm.pt", "pretrained_models/CosyVoice2-0.5B_cetts/llm.pt")
    # text += f"角色之前经历过:{personal_experience}"
    # text += f"角色现在所处场景:{context_description}"
    # text += f"请你模仿这个角色,用{','.join(emotions)}的语气说话。<|endofprompt|>"
    # text += transcription

    shutil.copy2("emotion_llm.pt", "pretrained_models/CosyVoice2-0.5B_cetts/llm.pt")
    text += f"请你模仿这个角色,用{','.join(emotions)}的语气说话。<|endofprompt|>"
    text += transcription

    prompt_wav_path = './asset/zero_shot_prompt.wav'


    cosyvoice = AutoModel(model_dir='pretrained_models/CosyVoice2-0.5B_cetts')
    for model_output in inference_contextspeech_onesample_test(
        cosyvoice,
        tts_text=text,
        prompt_speech=prompt_wav_path,
        llm_prompt_speech=prompt_wav_path,
    ):
        torchaudio.save(f'test.wav', model_output['tts_speech'], cosyvoice.sample_rate)

def main():
    cosyvoice2_example()


if __name__ == '__main__':
    main()