Can I load the Kokoro-82M model (python) from a local directory that isn't the basic transformers cache directory?

[ThoBoul]

[Fire-Cream]

no problem

This is my directory structure.

Here is the script.

`
from kokoro import KModel, KPipeline
from pathlib import Path
import numpy as np
import soundfile as sf
import torch
import tqdm

加载语音

voice = 'zf_001' if True else 'zm_001'
voice_tensor = torch.load(f'voices/{voice}.pt', weights_only=True)

# 检测是 cpu 运行还是 gpu 运行

device = 'cuda' if torch.cuda.is_available() else 'cpu'

# 加载模型

repo_id = 'hexgrad/Kokoro-82M-v1.1-zh'
model_path = 'models/kokoro-v1_1-zh.pth'
config_path = 'models/config.json'
model = KModel(model=model_path, config=config_path, repo_id=repo_id).to(device).eval()

基础设置

采样率

SAMPLE_RATE = 24000

段落空白间隔

N_ZEROS = 5000

测试文本

texts = [
("早上好",)
]

加载英文模型

en_pipeline = KPipeline(lang_code='a', repo_id=repo_id, model=False)

def en_callable(text):
if text == 'Kokoro':
return 'kˈOkəɹO'
elif text == 'Sol':
return 'sˈOl'
return next(en_pipeline(text)).phonemes

def speed_callable(len_ps):
speed = 0.8
if len_ps <= 83:
speed = 1
elif len_ps < 183:
speed = 1 - (len_ps - 83) / 500
return speed * 1.1

加载中文模型

zh_pipeline = KPipeline(lang_code='z', repo_id=repo_id, model=model, en_callable=en_callable)

path = Path(file).parent

wavs = []
for paragraph in tqdm.tqdm(texts):
for i, sentence in enumerate(paragraph):
generator = zh_pipeline(sentence, voice=voice, speed=speed_callable)
f = path / f'zh{len(wavs):02}.wav'
result = next(generator)
wav = result.audio
sf.write(f, wav, SAMPLE_RATE)
if i == 0 and wavs and N_ZEROS > 0:
wav = np.concatenate([np.zeros(N_ZEROS), wav])
wavs.append(wav)

sf.write(path / f'HEARME_{voice}.wav', np.concatenate(wavs), SAMPLE_RATE)
`

[DrugLover]

1. download the files (only needs config.json and weight file)
2. load model by yourself, you need to set model_path and config_path by yourself

from kokoro.model import KModel
local_model = KModel(model=model_path, config=config_path)
pipeline = KPipeline(lang_code='z', model=local_model, device='cuda')