1038lab

Kokoro-82M-zh-MLX is a fast, high-quality text-to-speech model optimized for Apple Silicon. Built on the official Kokoro architecture with native MLX inference — no GPU required on your Mac.

Features


🎙️ 103 Voices	55 Chinese Female (zf_), 45 Chinese Male (zm_), 3 English (af_*)
🚀 Apple Silicon Native	Runs entirely on Metal GPU via MLX — up to 10× faster than CPU
🔊 High Quality	Based on Kokoro v1.1-zh with 100+ hours of professional training data
📦 Zero Dependencies	Single `model.safetensors` + `voices/*.safetensors`, ready to use

Quick Start

from mlx_audio.tts.generate import load_model

m = load_model("1038lab/Kokoro-82M-zh-MLX")

results = list(m.generate(
    text="今天天气很好，一起去公园散步吧。",
    voice="zf_001",
    speed=1.0,
    lang_code="z"
))
from mlx_audio.audio_io import write as audio_write
audio_write("output.wav", results[0].audio, 24000, format="wav")

Installation

pip install misaki[zh] phonemizer-fork espeakng-loader mlx-audio

Usage

Speech Generation

from mlx_audio.tts.generate import load_model
from mlx_audio.audio_io import write as audio_write

m = load_model("1038lab/Kokoro-82M-zh-MLX")

# Chinese TTS (lang_code="z")
results = list(m.generate(
    text="你好世界，这是一段中文测试。",
    voice="zm_009",       # male voice
    speed=1.0,
    lang_code="z"         # "z"=Chinese, "a"=English
))
audio_write("output.wav", results[0].audio, 24000, format="wav")

# English TTS
results = list(m.generate(
    text="Hello world!",
    voice="af_heart",
    speed=1.0
))
audio_write("output.wav", results[0].audio, 24000, format="wav")

Streaming

results = list(m.generate(text="A long text here...", voice="zf_001", lang_code="z", streaming=True))
for chunk in results:
    if chunk.audio is not None:
        print(f"Chunk {chunk.segment_idx}: {len(chunk.audio)} samples")

Available Voices

Voice Prefix	Type	Count
`zf_`	Chinese Female	55
`zm_`	Chinese Male	45
`af_` / `bf_`	English	3

Parameters

Parameter	Description	Default
`text`	Input text	required
`voice`	Voice identifier	`"zf_001"`
`speed`	Speed multiplier	`1.0`
`lang_code`	Language (`"z"`=Chinese, `"a"`=English)	auto-detect
`streaming`	Enable streaming output	`False`

Performance

Benchmarked on M-series MacBook Pro:

Task	RTF	Latency
zh zf_001, 5s text	~0.07×	~2.4s for 5s audio
en af_heart, 3s text	~0.85×	~2.5s for 3s audio

Lower RTF = faster. RTF < 1.0 means faster than real-time.

Audio Samples

Click any audio below to hear the voice quality:

Voice ID	Demo Text	Audio
zf_001 (Chinese Female)	我现在正在使用最新的语音合成技术...	▶
zm_009 (Chinese Male)	我现在正在使用最新的语音合成技术测试男声效果...	▶
af_heart (English)	Hello! This is Heart speaking...	▶

See samples/index.html for a full interactive voice index with all 15 demo voices.

License

This model is released under the Apache 2.0 License. The underlying model architecture is from hexgrad/Kokoro. The voice embeddings are converted from the official kokoro-v1.1-zh models.

Acknowledgements

Original model by hexgrad
Training data by 龙猫数据 LongMaoData
MLX conversion and voice embedding extraction by 1038lab

1038lab

Kokoro-82M-zh-MLX 是一个专为 Apple Silicon (M1/M2/M3/M4) 优化的开源语音合成模型。基于原生 MLX 架构实现，无需额外 GPU 即可在 Mac 上快速运行。

功能特点


🎙️ 103 个声音	55 个中文女声 (zf_)、45 个中文男声 (zm_)、3 个英文 (af_, bf_)
🚀 Apple Silicon 原生加速	完全支持 Metal GPU — 比 CPU 快 10 倍以上
🔊 高质量语音	基于 Kokoro v1.1-zh，100+ 小时专业训练数据
📦 开箱即用	`model.safetensors` + `voices/*.safetensors`，无需复杂配置

快速开始

from mlx_audio.tts.generate import load_model

m = load_model("1038lab/Kokoro-82M-zh-MLX")

results = list(m.generate(
    text="今天天气很好，一起去公园散步吧。",
    voice="zf_001",
    speed=1.0,
    lang_code="z"
))
from mlx_audio.audio_io import write as audio_write
audio_write("output.wav", results[0].audio, 24000, format="wav")

安装依赖

pip install misaki[zh] phonemizer-fork espeakng-loader mlx-audio

使用说明

语音合成

from mlx_audio.tts.generate import load_model
from mlx_audio.audio_io import write as audio_write

m = load_model("1038lab/Kokoro-82M-zh-MLX")

# 中文发音
results = list(m.generate(
    text="你好世界，这是一段中文测试。",
    voice="zm_009",       # 男声
    speed=1.0,
    lang_code="z"         # 语言代码："z"=中文, "a"=英文
))
audio_write("output.wav", results[0].audio, 24000, format="wav")

# 英文发音
results = list(m.generate(
    text="Hello world!",
    voice="af_heart",
    speed=1.0
))
audio_write("output.wav", results[0].audio, 24000, format="wav")

流式输出

results = list(m.generate(text="这是一段长文本...", voice="zf_001", lang_code="z", streaming=True))
for chunk in results:
    if chunk.audio is not None:
        print(f"音频块 {chunk.segment_idx}: {len(chunk.audio)} samples")

可用声音

声音前缀	类型	数量
`zf_`	中文女声	55
`zm_`	中文男声	45
`af_` / `bf_`	English	3

参数说明

参数	描述	默认值
`text`	要转换的文本	必填
`voice`	声音标识符	`"zf_001"`
`speed`	语速倍率	`1.0`
`lang_code`	语言 (`"z"`=中文, `"a"`=英文)	自动检测
`streaming`	是否启用流式输出	`False`

性能表现

在 M 系列 MacBook Pro 上的实测结果：

任务	RTF	耗时
中文 zf_001，5s 文本	~0.07×	~2.4s 生成 5s 音频
英文 af_heart，3s 文本	~0.85×	~2.5s 生成 3s 音频

RTF 越小越快。RTF < 1.0 表示比实时更快。

示例音频

点击查看各个声音的演示音频：samples/index.html

声音 ID	示范文本	试听
zf_001 (中文女声)	我现在正在使用最新的语音合成技术...	▶
zm_009 (中文男声)	我现在正在使用最新的语音合成技术测试男声效果...	▶
af_heart (English)	Hello! This is Heart speaking...	▶

许可证

本项目基于 Apache 2.0 License 发布。底层模型架构来自 hexgrad/Kokoro。声音嵌入文件由官方 kokoro-v1.1-zh 模型转换而来。

致谢

原始模型由 hexgrad 开发
训练数据由龙猫数据 LongMaoData 提供
MLX 转换及声音提取由 1038lab 完成

Downloads last month: 200

Safetensors

Model size

81.8M params

Tensor type

F32