MOSS-TTS-Nano

1. Project Overview

official concept map and architecture diagram:

! MOSS-TTS-Nano Concept

! MOSS-TTS-Nano Architecture

2. What does it mostly do?

3. Language Support

README currently lists 20 languages: Chinese, English, German, Spanish, French, Japanese, Italian, Hungarian, Korean, Russian, Persian, Arabic, Polish, Portuguese, Czech, Danish, Swedish, Greek, Turkish, etc.

This means that it is not only suitable for Chinese reading, but also can enter cross-border e-commerce, overseas education, overseas customer service, international content broadcast and other scenes. However, the actual naturalness and accent performance of each language still need to be verified with customer samples.

4. Applicable Scenario

5. Not suitable for the scene

6. Architecture Understanding

MOSS-TTS-Nano a pure autoregressive architecture using Audio Tokenizer LLM. This can be explained to the customer:

1. Audio is first converted into discrete audio token through MOSS-Audio-Tokenizer-Nano.
2. The TTS model generates audio tokens just like the language model generates text tokens.
3. The audio tokenizer then decodes the token into 48 kHz, two-channel audio.

The official also provides the MOSS-Audio-Tokenizer-Nano architecture and evaluation map:

! MOSS-Audio-Tokenizer-Nano Architecture

! MOSS-Audio-Tokenizer Evaluation

The pre-sale selling point of this architecture is the unified audio token representation, which can be extended to voice, dialogue, sound effects and other models in the MOSS-TTS family. However, for the current customer landing, the most practical is Nano's lightweight deployment.

How to use #7.

Environment:

conda create -n moss-tts-nano python=3.12 -y
conda activate moss-tts-nano

git clone https://github.com/OpenMOSS/MOSS-TTS-Nano.git
cd MOSS-TTS-Nano

pip install -r requirements.txt
pip install -e .

Voice Cloning:

python infer.py \
  --prompt-audio-path assets/audio/zh_1.wav \
  --text "欢迎关注模思智能、上海创智学院与复旦大学自然语言处理实验室。"

Local Web Demo:

python app.py

Open:

http://127.0.0.1:18083

ONNX CPU reasoning:

python infer_onnx.py \
  --prompt-audio-path assets/audio/zh_1.wav \
  --text "Welcome to the ONNX Runtime CPU demo."

CLI:

moss-tts-nano generate \
  --backend onnx \
  --prompt-speech assets/audio/zh_1.wav \
  --text "欢迎关注模思智能、上海创智学院与复旦大学自然语言处理实验室。"

Service Model:

moss-tts-nano serve --backend onnx

8. What can I say before sales

One-sentence positioning:

"MOSS-TTS-Nano is a lightweight TTS model that can run locally, is CPU-friendly, supports voice cloning and multi-language, and is suitable for turning the text output of AI applications into low-latency voice output."

Customer Value Mapping:

9. PoC Recommendations

It is recommended to prepare three types of audio samples before sales: ordinary reading, business speech and interactive short sentences. Don't just test a short text, you must test long text, numbers, English abbreviations, names, professional terms and mixed reading in Chinese and English.

10. Frequently Asked Customer Questions

11. Risks and Considerations

1. Voice cloning compliance: must have clear authorization, especially employees, anchors, teachers, customer service and other real timbre.
2. Sound quality needs to be measured: small models pursue deployment efficiency and cannot reach top commercial dubbing quality by default.
3. Language coverage does not equal quality: 20 languages are subject to separate acceptance by target market.
4. Dependent installation: README mentions that WeTextProcessing / pynini may require additional processing.
5. Production still needs service layer: authentication, concurrency, cache, log, audit, text cleaning, sensitive word control all need to be supplemented.

12. My Pre-Sales Judgment

The advantages of MOSS-TTS-Nano are clear: lightweight, open source, local, ONNX, and voice cloning. It is not used to confront the strongest commercial TTS in emotional performance, but is very suitable for customers who "need to put voice capabilities in edge, local, browser or lightweight services.

pre-sales recommendations are used for education, reading, AI assistant, enterprise knowledge base voice broadcast, and terminal-side voice demo. When pushing forward, PoC should be done with the customer's real text and target hardware: if CPU delay, sound quality and compliance can be passed, it can become a very cost-effective TTS scheme component.

13. REFERENCE

-GitHub: https://github.com/OpenMOSS/MOSS-TTS-Nano

-Demo: https://openmoss.github.io/MOSS-TTS-Nano-Demo/

-Hugging Face: https://huggingface.co/OpenMOSS-Team/MOSS-TTS-Nano

-ModelScope: https://modelscope.cn/models/openmoss/MOSS-TTS-Nano

-Thesis: https://arxiv.org/abs/2603.18090

-ONNX Weights: https://huggingface.co/OpenMOSS-Team/MOSS-TTS-Nano-100M-ONNX