inference.py

import argparse
import numpy as np
import torch
from moshi.models.loaders import CheckpointInfo
from moshi.models.tts import DEFAULT_DSM_TTS_REPO, DEFAULT_DSM_TTS_VOICE_REPO, TTSModel

def load_model():
    checkpoint_info = CheckpointInfo.from_hf_repo(DEFAULT_DSM_TTS_REPO)
    tts_model = TTSModel.from_checkpoint_info(
        checkpoint_info, n_q=32, temp=0.6, device=torch.device("cpu")
    )
    return tts_model

def generate_audio(tts_model, text, voice):
    entries = tts_model.prepare_script([text], padding_between=1)
    voice_path = tts_model.get_voice_path(voice)
    condition_attributes = tts_model.make_condition_attributes(
        [voice_path], cfg_coef=2.0
    )

    pcms = []
    def _on_frame(frame):
        if (frame != -1).all():
            pcm = tts_model.mimi.decode(frame[:, 1:, :]).cpu().numpy()
            pcms.append(np.clip(pcm[0, 0], -1, 1))

    all_entries = [entries]
    all_condition_attributes = [condition_attributes]
    with tts_model.mimi.streaming(len(all_entries)):
        tts_model.generate(all_entries, all_condition_attributes, on_frame=_on_frame)

    audio = np.concatenate(pcms, axis=-1)
    return audio