SAVRN
Search Contact SAVRN

Open-weight model · Text to speech

PocketTTS-ZH-6L

by Hustfyb hustfyb/PocketTTS-ZH-6L

中文语音合成(TTS)模型,支持语音克隆与数字/中英混读,CPU 优先设计,权重原生 int8 量化。 基于 Kyutai Pocket TTS 架构:6 层 FlowLM(dmodel 1024)+ Mimi 神经音频编解码器(24 kHz),端到端流式推理。本版本(R21)相对上一代的核心变化: - 原生 int8 交付:FlowLM 注意力与 FFN 权重已离线量化打包,无需运行时转换,体积 223 MB,CPU 速度提升约 1 倍 - 9174 token 词表:官方英文 BPE-4000 基底 +…

Parameters
Context
Weights232.8 MB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads

Model Card

By Hustfyb, published under apache-2.0, revision bf7a6c1557bc.

中文语音合成(TTS)模型,支持语音克隆与数字/中英混读,CPU 优先设计,权重原生 int8 量化。 基于 Kyutai Pocket TTS 架构:6 层 FlowLM(dmodel 1024)+ Mimi 神经音频编解码器(24 kHz),端到端流式推理。本版本(R21)相对上一代的核心变化: - 原生 int8 交付:FlowLM 注意力与 FFN 权重已离线量化打包,无需运行时转换,体积 223 MB,CPU 速度提升约 1 倍 - 9174 token 词表:官方英文 BPE-4000 基底 + 追加单字汉字/全角标点,bytefallback 兜底生僻字符 - 中英混读:APP、VIP、5G、WiFi 等英文术语按中文习惯读法 - 模型为预量化 int8 权重(modelint8.pt),加载即用,无需 quantize=True 或有界面演示(需 pip install gradio): - 采样温度 temp=0.3(默认);文本越长对参考音频质量越敏感 文本准确性用 whisper-small 对合成音频转写后计算字错误率(CER),每个场景 10 次独立采样取均值,生产配置(单步采样)。数值越低越好。 测试环境:Intel i7-12700KF,生成 52 字客服长句(约 8.5 秒音频),每档 3 次取均值。 参考音频:客服目标音色(kefulong / digitsorder / codeswitch / shortunseen)与 AISHELL-3 说话人 A/B(speakerAlong /…

Read Hustfyb's full model card

PocketTTS-ZH-6L(R21)中文语音克隆模型

中文语音合成(TTS)模型,支持语音克隆与数字/中英混读,CPU 优先设计,权重原生 int8 量化。

基于 Kyutai Pocket TTS 架构:6 层 FlowLM(d_model 1024)+ Mimi 神经音频编解码器(24 kHz),端到端流式推理。本版本(R21)相对上一代的核心变化:

  • 原生 int8 交付:FlowLM 注意力与 FFN 权重已离线量化打包,无需运行时转换,体积 223 MB,CPU 速度提升约 1 倍
  • 9174 token 词表:官方英文 BPE-4000 基底 + 追加单字汉字/全角标点,byte_fallback 兜底生僻字符
  • 内置文本规范化(TN):阿拉伯数字、金额、日期、订单号自动转中文读法(368.5三百六十八点五,16 位订单号逐位读出)
  • 中英混读:APP、VIP、5G、WiFi 等英文术语按中文习惯读法

English summary: A Mandarin Chinese TTS model with voice cloning, built on the Pocket TTS architecture (6-layer flow transformer + Mimi codec). Weights ship pre-quantized int8 (223 MB, no runtime conversion needed): 0.18× real-time single-threaded on an i7-12700KF (≈5.5× faster than real-time audio production), 0.16× with 4 threads. Built-in text normalization reads Arabic digits, amounts, dates and order numbers in Mandarin (digit-by-digit for serial numbers). 9174-token SentencePiece vocabulary. Intended for customer-service dialogue and short-form Mandarin speech. See Performance and Limitations below.

快速开始

pip install pocket-tts soundfile wetext
cd <本目录>
python demo.py --prompt your_voice.wav --text "您的订单编号是20260918123456,退款368.5元预计三个工作日内到账。" --out out.wav
  • TN 默认开启:数字、金额、日期、订单号自动转中文读法;--no-tn 可关闭
  • 模型为预量化 int8 权重(model_int8.pt),加载即用,无需 quantize=True

或有界面演示(需 pip install gradio):

python demo.py --gradio

界面中可看到 TN 规范化后的文本,并可选关闭 TN。

Python API:

import sys
sys.path.insert(0, "<本目录>")
from demo import load  # int8 加载 + TN
from tn import normalize_tn

model = load()
state = model.get_state_for_audio_prompt("your_voice.wav")  # 3~10 秒干净普通话
text = normalize_tn("您的订单编号是20260918123456")  # TN 显式调用(可选,demo.synthesize 默认已含)
audio = model.generate_audio(state, text)
  • 参考音频:3~10 秒、干净、单说话人普通话,效果最佳
  • 采样温度 temp=0.3(默认);文本越长对参考音频质量越敏感

性能

文本准确性用 whisper-small 对合成音频转写后计算字错误率(CER),每个场景 10 次独立采样取均值,生产配置(单步采样)数值越低越好。

主指标:客服业务长句(52 字),语音克隆(客服参考音色)

指标 CER (n=10)
客服长句(主指标) 8.1%(上一代 13.65%)
数字订单句(16 位订单号 + 金额) 0%(ASR 逐字核对通过)
中英混读句(APP/VIP/官网) ASR 全对

分文本长度(客服参考音色,n=10)

文本 CER (n=10)
短句未见(9 字) 17.1%
中句未见(15 字) 6.2%
长句未见(31 字) 6.4%

速度与资源(int8 权重)

测试环境:Intel i7-12700KF,生成 52 字客服长句(约 8.5 秒音频),每档 3 次取均值。

项目 数值
CPU 推理速度(单线程,库默认) 0.18× 实时(生成 1 秒音频约需 0.18 秒)
CPU 推理速度(2 线程) 0.17× 实时
CPU 推理速度(4 线程) 0.16× 实时
权重文件 223 MB(model_int8.pt,FlowLM int8 + Mimi fp32)
FlowLM 参数量 约 89M(注意力 + FFN 已 int8)
采样率 24 kHz

示例音频

参考音频:客服目标音色(kefu_long / digits_order / codeswitch / short_unseen)与 AISHELL-3 说话人 A/B(speakerA_long / speakerB_long,仅用于演示音色克隆,不在交付范围内)。均由交付的 int8 模型 + TN 默认配置生成。

场景 文本 试听
客服长句 您好,非常抱歉给您带来了不好的使用体验,麻烦您先提供一下订单编号,我这边马上帮您查询订单信息和售后政策。
数字订单 您的订单编号是20260918123456,退款368.5元预计三个工作日内到账。
中英混读 请打开APP查看您的VIP会员状态,或者登录官网查询订单详情。
短句 今天天气很好,适合外出散步。
说话人 A 请问您反映的问题我们已经记录下来,稍后会有专员与您联系,请保持电话畅通。
说话人 B 人工智能正在改变我们的生活方式。

输入文本说明

  • TN 默认开启(推荐):阿拉伯数字、金额、日期、电话、订单号/运单号/航班号自动转中文口语读法;订单号类编号逐位读出(如 20260918123456二零二六零九一八一二三四五六)
  • 关闭 TN(--no-tn)时,请自行使用中文数字写法;词表虽有数字 token,但训练数据全部经过 TN,关闭后数字读法不受控
  • 网址(URL)、邮箱等复合符号朗读不稳定,建议预处理或避免
  • 纯英文句子不受支持(可能输出静音),中英混读仅限常见术语词汇

局限性与已知问题

  • 极短句 + 语音克隆可能不稳定:约 9 字且无标点的文本,在嘈杂/电话信道参考音频下可能输出静音或复述参考音频内容(建议参考此类文本时实测);demo 已内置静音检测自动换 seed 重试
  • 未见说话人泛化有限:对全新音色克隆长业务句 CER 明显高于已适应音色,建议参考音频清晰、时长 5 秒以上
  • 必须提供参考音频(语音克隆是本模型的设计形态);无参考音频的直接合成质量不可用
  • 方言、外语、唱歌不适用;情绪化/高表现力朗读能力有限
  • 音色相似度未做定量评测,建议实际试听

伦理与使用限制

  • 本模型仅用于已获得合法授权的语音合成。未经声音主体明确同意,禁止克隆其声音
  • 禁止用于冒充、欺诈、制造虚假信息或任何违法用途
  • 合成内容应明确标识为 AI 生成

文件清单

文件 说明
model_int8.pt 预量化 int8 权重(FlowLM int8 + Mimi fp32),223 MB
config.yaml 模型架构配置(自包含,不含权重路径依赖)
tokenizer.model SentencePiece 分词器(9174 词,含 byte_fallback)
tn.py 文本规范化(TN)模块,wetext + 业务规则
demo.py 推理演示(int8 加载 + TN,CLI + Gradio)
samples/ 示例音频(见上表)
README.md 本文件

Identity and Version

Repository
hustfyb/PocketTTS-ZH-6L
Publisher
Hustfyb
Task
Text to speech
Modality
Audio
Library
pocket-tts
Parameters
Not stated by the source
Languages
zh
Revision
bf7a6c1557bc6ce00abbd433465661dd07e505c2
First published
2026-09-16
Last updated
2026-09-18

Files and Weights

13 files, 234.8 MB in total. The weights are 1 file totalling 232.8 MB in pt.

Weights1 file · 232.8 MB
Configuration3 files · 10.0 KB
Tokenizer1 file · 131.8 KB
Documentation1 file · 7.2 KB
Other6 files · 1.9 MB
Repository1 file · 938 B
Every file
FileTypeSizeSHA-256
model_int8.ptWeights232.8 MB 7e3c46d1a54e
config.yamlConfiguration1.4 KB
demo.pyConfiguration5.6 KB
tn.pyConfiguration3.0 KB
README.mdDocumentation7.2 KB
samples/codeswitch.wavOther311.1 KB 340fc674bc41
samples/digits_order.wavOther376.4 KB 56cae9ae831a
samples/kefu_long.wavOther445.5 KB 66f5c08f4f26
samples/short_unseen.wavOther122.9 KB 692ca2ded006
samples/speakerA_long.wavOther441.6 KB 088c02a7a49d
samples/speakerB_long.wavOther180.5 KB c061eaac65a0
.gitattributesRepository938 B
tokenizer.modelTokenizer131.8 KB 039a38b5ee40

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
232.8 MB
Download from Hustfyb

Released by Hustfyb through its official repository on Hugging Face. Read the license.

Memory Requirements

PrecisionWeights in memory
As published232.8 MB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About PocketTTS-ZH-6L

Can I use PocketTTS-ZH-6L commercially?

Yes. PocketTTS-ZH-6L is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Text to speech

Kokoro-82M

Hexgrad

Kokoro is an open-weight TTS model with 82 million parameters. Despite its lightweight architecture, it delivers comparable quality to larger models while being significantly faster and more cost-efficient. With Apache-licensed weights, Kokoro can be deployed anywhere from production environments to personal projects. You can run this basic cell on Google Colab. Listen to samples. For more languages and details, see Advanced Usage. Under the hood, kokoro uses misaki, a G2P library at https://github.com/hexgrad/misaki Model SHA256 Hash: 496dba118d1a58f5f3db2efc88dbdc216e0483fc89fe6e47ee1f2c53f18ad1e4 Data: Kokoro was trained exclusively on permissive/non-copyrighted audio data and IPA…

Open weights apache-2.0

Model · Text to speech

XTTS-v2

Coqui.ai

ⓍTTS is a Voice generation model that lets you clone voices into different languages by using just a quick 6-second audio clip. There is no need for an excessive amount of training data that spans countless hours. This is the same or similar model to what powers Coqui Studio and Coqui API. - Supports 17 languages. - Voice cloning with just a 6-second audio clip. - Emotion and style transfer by cloning. - Cross-language voice cloning. - Multi-lingual speech generation. - 24khz sampling rate. - 2 new languages; Hungarian and Korean - Architectural improvements for speaker conditioning. - Enables the use of multiple speaker references and interpolation between speakers. - Stability…

Open weights other coqui

Model · Text to speech

audio.cpp-gguf

Audio.cpp

This directory contains audio.cpp-native GGUF conversions of multiple speech models. These files are intended for use with audio.cpp. If you enjoy the project, please star audio.cpp on GitHub and this Hugging Face repository. For conversion details, supported layouts, direct-file loading, sidecar embedding, and the latest compatibility notes, see the audio.cpp GGUF guide: - https://github.com/0xShug0/audio.cpp/blob/main/docs/gguf.md!!! Converted and quantized packages are checked with automated metrics, but perceived quality can still differ for human listeners. Please validate the exact package, backend, and route to confirm the output is acceptable for your use case. The table lists the…

Open weights other audio.cpp

Model · Text to speech

chatterbox

Resemble AI

Chatterbox Multilingual V3 is the latest general-purpose multilingual TTS model in the Chatterbox family. It keeps the same 0.5B model size while improving speaker similarity, reducing hallucinations, and producing more natural, conversational speech across languages. V3 is designed for broad language coverage like V2, but with stronger stability and more expressive generation. It is the recommended multilingual model for users who want one voice cloning model that works across many languages. Try it in the Chatterbox Multilingual TTS V3 Space. Alongside V3, we are releasing the Single Language Pack: dedicated finetunes for priority languages where tighter quality control, stronger…

Open weights mit chatterbox

Model · Text to speech

F5-TTS

Yushen CHEN

Download F5-TTS or E2 TTS and place under ckpts/ Paper: F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching

Open weights cc-by-nc-4.0 f5-tts

Model · Text to speech

Kokoro-82M-v1.0-ONNX

ONNX Community

Kokoro is a frontier TTS model for its size of 82 million parameters (text in/audio out). First, install the kokoro-js library from NPM using: You can then generate speech as follows: Optionally, save the audio to a file: The model is resilient to quantization, enabling efficient high-quality speech synthesis at a fraction of the original model size.

Open weights apache-2.0 transformers.js