← Back to Product Feed

GitHub Open Source OpenMOSS/MOSS-TTS-Nano

MOSS-TTS-Nano is an open-source multilingual tiny speech generation model from MOSI.AI and the OpenMOSS team. With only 0.1B parameters, it is designed for realtime speech generation, can run directly on CPU without a GPU, and keeps the deployment stack simple enough for local demos, web serving, and lightweight product integration.

2,049
Traction Score
276
Forks
Apr 10, 2026
Launch Date
View Origin Link

Product Positioning & Context

AI Executive Synthesis
Delivering customizable speech output (rate control) and optimizing real-time streaming performance on CPU-only deployments.
This issue highlights two critical performance and feature gaps for MOSS-TTS-Nano, directly contradicting its "realtime speech generation" and "CPU without a GPU" value proposition. First, the user inquires about speech rate control, a fundamental feature for practical TTS applications. Second, the explicit mention of "CPU streaming mode being sluggish" ("比较卡顿") and the provided performance metrics (4s audio taking 11.6s elapsed time on 4 CPU threads) confirm significant latency issues. This performance deficit on CPU, combined with a lack of basic customization, severely limits the model's utility for real-time interactive applications or scenarios requiring dynamic speech output. The market implication is that the model, despite its small size, fails to meet core performance and feature expectations for its target use cases, hindering adoption in applications where responsiveness and control are paramount.
MOSS-TTS-Nano is an open-source multilingual tiny speech generation model from MOSI.AI and the OpenMOSS team. With only 0.1B parameters, it is designed for realtime speech generation, can run directly on CPU without a GPU, and keeps the deployment stack simple enough for local demos, web serving, and lightweight product integration.
audio-tokenizer chinese english multi-modality multilingual realtime streaming-audio tts

Related Ecosystem & Alternatives

Discover adjacent products, open-source repositories, and developer tools sharing similar technical architecture.

Deep-Dive FAQs

What is OpenMOSS/MOSS-TTS-Nano?
OpenMOSS/MOSS-TTS-Nano is analyzed by our AI as: Delivering customizable speech output (rate control) and optimizing real-time streaming performance on CPU-only deployments.. It focuses on This issue highlights two critical performance and feature gaps for MOSS-TTS-Nano, directly contradicting its "realtime speech generation" and "CPU...
Where did OpenMOSS/MOSS-TTS-Nano originate?
Data for OpenMOSS/MOSS-TTS-Nano was aggregated directly from the GitHub Open Source community ecosystem, representing raw developer and early-adopter sentiment.
When was OpenMOSS/MOSS-TTS-Nano publicly launched?
The initial public indexing or launch date for OpenMOSS/MOSS-TTS-Nano within our tracked developer communities was recorded on April 10, 2026.
How popular is OpenMOSS/MOSS-TTS-Nano?
OpenMOSS/MOSS-TTS-Nano has achieved measurable traction, logging over 2,049 traction score and facilitating 276 recorded discussions or engagements.
Which technical categories define OpenMOSS/MOSS-TTS-Nano?
Based on metadata extraction, OpenMOSS/MOSS-TTS-Nano is categorized under topics such as: audio-tokenizer, chinese, english, multi-modality.
Are there active development issues for OpenMOSS/MOSS-TTS-Nano?
Yes, we are currently tracking open architectural debates and bug reports for this project on GitHub. There are currently 5 active high-priority issues logged recently.
What are some commercial alternatives to OpenMOSS/MOSS-TTS-Nano?
Our semantic intelligence engine identifies potential commercial alternatives in the SaaS space, such as VoxCPM2, which offers overlapping value propositions.
How does the creator describe OpenMOSS/MOSS-TTS-Nano?
The original author or development team describes the product as follows: "MOSS-TTS-Nano is an open-source multilingual tiny speech generation model from MOSI.AI and the OpenMOSS team. With only 0.1B parameters, it is designed for realtime speech generation, can run direc..."

Active Developer Issues (GitHub)

open 可以量化吗
Logged: Apr 17, 2026
open 每次生成需要从 huggingface.co 加载 config.json 等,但是
Logged: Apr 17, 2026
open
Logged: Apr 17, 2026
open 请问是否支持语速设置?CPU流式模式比较卡顿,如何优化呢?
Logged: Apr 17, 2026
open 中文方言支援
Logged: Apr 16, 2026

Community Voice & Feedback

Jandown • May 5, 2026
> onnx在中等性能的cpu上勉勉强强能达到实时,但是效果差(丢词,读错,生硬),一般业务使用不了。 非onnx的模型中等cpu上完全达不到实时,听感是一种折磨。在gpu(3090)上短文本的RTF>1,达不到实时。长文本的RTF在0.35左右。非onnx的效果比onnx的效果好很多,但是达不到实时,所以综合下来,还得优化。

赞同!
ramishi • May 4, 2026
> 我们开源了WenetSpeech粤、川、吴等系列方言数据

@hujingbin1 请问在哪里可以找到?
koxiong • Apr 25, 2026
onnx版本实测下来:
1、Intel(R) Core(TM) i5-9500 CPU @ 3.00GHz--这款设备上,ONNX推理速度没有明显的变化,RTF大约还是在1.13左右
2、在Mac mini上推理,纯CPU:这款速度上比torch版本要快不少,CPU利用率大约在200%,RTF能到0.6左右。
但是都是长文本推理时,都是存在长时间停顿问题。不知道是不是界面做的不好,还是模型本身就会停顿。
以上都是使用4核心推理。
wen0320 • Apr 24, 2026
onnx在中等性能的cpu上勉勉强强能达到实时,但是效果差(丢词,读错,生硬),一般业务使用不了。
非onnx的模型中等cpu上完全达不到实时,听感是一种折磨。在gpu(3090)上短文本的RTF>1,达不到实时。长文本的RTF在0.35左右。非onnx的效果比onnx的效果好很多,但是达不到实时,所以综合下来,还得优化。
hujingbin1 • Apr 19, 2026
> 可提供思路怎样微调方言吗?

可以尝试扩充词表,加special accent tags,然后训练。或者尝试instruct训练,给出对应方言的描述,如“请用粤语说”,然后加粤语数据微调。我们开源了WenetSpeech粤、川、吴等系列方言数据,可以关注下。
alpacaking • Apr 17, 2026
您好,我们发布了推理速度更快的onnx版,欢迎试用。
gyt1145028706 • Apr 17, 2026
您好, 可以试一下我们 ONNX 实现, 实测下来比朴素 pytorch 版本快不少
https://github.com/OpenMOSS/MOSS-TTS-Nano/commit/7928ec16500378de31d17bc86ce719cc9fd7b84f
wen0320 • Apr 17, 2026
我的也非常慢,实时播放的音频一卡一卡的:
Done | mode=voice_clone | prompt=zh_1 | attn=eager | tts_batch=1 | codec_batch=1 | exec=cpu | cpu_threads=12 | audio=4.80s | elapsed=13.22s
state=done | emitted=4.80s | lead=-6.89s | first_audio=1.41s
CPU:Intel(R) Core(TM) i5-10400 CPU @ 2.90GHz
gyt1145028706 • Apr 17, 2026
您好, 我们马上会出一个 ONNX 版本, 敬请期待
surfincanoy • Apr 17, 2026
[pynini](https://github.com/SystemPanic/pynini-windows/releases/tag/v2.1.6.post1)

If you are using Windows without Conda, you can download the package from the link and install it. However, after installation, you need to modify the version number of the pynini package in your virtual environment (for example, in my case: .venv\Lib\site-packages\pynini\__init__.py): change `__version__ = "2.1.6.post1"` to `__version__ = "2.1.6"`. Then, install both pynini and WeTextProcessing, and it should work.
CloudRipple • Apr 17, 2026
可以尝试把hf的来源替换为镜像源
```bash
export HF_ENDPOINT=https://hf-mirror.com
```
alpacaking • Apr 17, 2026
可以尝试使用 modelscope 下载好模型,然后指定模型路径启动服务
```bash
modelscope download --model openmoss/MOSS-TTS-Nano-100M
modelscope download --model openmoss/MOSS-Audio-Tokenizer-Nano
```
xiami2019 • Apr 17, 2026
你好,感谢关注!也可以从 https://modelscope.cn/models/openmoss/MOSS-TTS-Nano 上下载模型。
alpacaking • Apr 17, 2026
或者您也可以参考 [#10](https://github.com/OpenMOSS/MOSS-TTS-Nano/issues/10#issuecomment-4265340787) 的解决方案
alpacaking • Apr 17, 2026
您好,不需要先改回环境变量。更可能是这些环境变量没有在当前 Python 进程中完全生效,或者仍有某个 Hugging Face 缓存路径没有被缩短。

从类似报错来看,关键是要确认这些环境变量是否真的在当前启动 `python app.py` 的同一个终端会话里生效。如果没有生效,Hugging Face 仍然可能继续使用默认的长路径缓存目录。

能否请您补充以下信息,方便我们进一步确认:

1. 在运行 `python app.py` 之前,同一个终端里执行下面命令,并把输出贴出来:

echo %HF_HOME%
echo %HF_HUB_CACHE%
echo %HUGGINGFACE_HUB_CACHE%
echo %TRANSFORMERS_CACHE%
echo %HF_MODULES_CACHE%
echo %TORCH_HOME%

2. 请把完整报错日志贴出来,尤其是包含 `FileNotFoundError: [WinError 206]` 的那一段完整路径。

3. 也请执行下面命令,并贴一下输出:

python -c "import os; print('HF_HOME=', os.environ.get('HF_HOME')); print('HF_HUB_CACHE=', os.environ.get('HF_HUB_CACHE')); print('HUGGINGFACE_HUB_CACHE=', os.environ.get('HUGGINGFACE_HUB_CACHE')); print('TRANSFORMERS_CACHE=', os.environ.get('TRANSFORMERS_CACHE')); print('HF_MODULES_CACHE=', os.environ.get('HF_MODULES_CACHE')); print('TORCH_HOME=', os.environ.get('TORCH_HOME'))"

如果方便的话,也建议您额外设置一个更短的动态模块缓存目录后再试一次:

set HF_HOME=C:\hf
set HF_HUB_CACHE=C:\hf\hub
set HF_MODULES_CACHE=C:\hf\modules
set TRANSFORMERS_CACHE=C:\hf\hu...

Discovery Source

GitHub Open Source GitHub Open Source

Aggregated via automated community intelligence tracking.

Tech Stack Dependencies

No direct open-source NPM package mentions detected in the product documentation.

Media Tractions & Mentions

No mainstream media stories specifically mentioning this product name have been intercepted yet.

Deep Research & Science

No direct peer-reviewed scientific literature matched with this product's architecture.