Moss Tts Nano

moss-tts-nano

Source: OpenMOSS/MOSS-TTS-Nano (Apache 2.0) Model: 0.1B params · 48kHz stereo · CPU-friendly · multilingual

When to use

• Add TTS / voice output to any agent or CLI tool
• Voice clone from a reference audio sample
• Realtime streaming speech generation without GPU
• Long-text narration with automatic chunking

Do NOT use for

• High-fidelity studio audio (use a larger model)
• Languages outside Chinese / English / supported list

Setup

# ONNX version — no PyTorch needed at inference time (recommended)
pip install onnxruntime numpy soundfile
git clone https://github.com/OpenMOSS/MOSS-TTS-Nano
cd MOSS-TTS-Nano
# Download ONNX weights from HuggingFace
huggingface-cli download OpenMOSS-Team/MOSS-TTS-Nano-100M-ONNX --local-dir ./onnx_weights
huggingface-cli download OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano-ONNX --local-dir ./onnx_tokenizer

# PyTorch version — full features + finetuning
conda create -n moss-tts python=3.10
conda activate moss-tts
pip install -r requirements.txt
huggingface-cli download OpenMOSS-Team/MOSS-TTS-Nano --local-dir ./checkpoints

Basic inference

# ONNX CPU (fastest, no GPU needed)
from onnx_tts_runtime import OnnxTTSRuntime

tts = OnnxTTSRuntime(
    tts_model_dir="./onnx_weights",
    tokenizer_dir="./onnx_tokenizer",
)

# Simple generation
tts.generate(
    text="Hello, this is MOSS TTS Nano.",
    output_path="output.wav",
)

# Voice clone from reference audio
tts.generate(
    text="Hello, this is MOSS TTS Nano.",
    reference_audio="path/to/speaker.wav",
    output_path="cloned_output.wav",
)

# CLI — generate
moss-tts-nano generate \
  --text "Xin chào, đây là MOSS TTS Nano." \
  --reference-audio speaker.wav \
  --output output.wav

# CLI — serve FastAPI
moss-tts-nano serve --host 0.0.0.0 --port 8000

Streaming inference

from moss_tts_nano_runtime import MossTTSNanoRuntime

runtime = MossTTSNanoRuntime(checkpoint_dir="./checkpoints")

# Stream audio chunks as they're generated
for chunk in runtime.stream(
    text="Long text here...",
    reference_audio="speaker.wav",
):
    play_audio_chunk(chunk)  # feed to speaker / buffer

FastAPI server

# Start server
python app_onnx.py  # ONNX version (recommended for prod)
python app.py       # PyTorch version

# POST /tts
import httpx, base64, soundfile as sf, io

resp = httpx.post("http://localhost:8000/tts", json={
    "text": "Hello from YAMTAM.",
    "reference_audio": None,   # omit for default voice
})
audio_bytes = base64.b64decode(resp.json()["audio_base64"])
sf.write("out.wav", *sf.read(io.BytesIO(audio_bytes)))

Finetuning

cd finetuning/

# Prepare dataset: each line = {"text": "...", "audio": "path/to/file.wav"}
python prepare_data.py --input your_data/ --output dataset/

# Finetune (single GPU or CPU)
python train.py \
  --base-checkpoint ../checkpoints \
  --dataset dataset/ \
  --output-dir finetuned/ \
  --epochs 10 \
  --batch-size 4

# Inference with finetuned checkpoint
python ../infer.py --checkpoint finetuned/ --text "..." --output out.wav

YAMTAM agent integration pattern

import httpx

def yamtam_speak(text: str, voice: str | None = None) -> bytes:
    """Call local MOSS-TTS-Nano server, return WAV bytes."""
    payload = {"text": text}
    if voice:
        payload["reference_audio"] = voice
    r = httpx.post("http://localhost:8000/tts", json=payload, timeout=30)
    r.raise_for_status()
    import base64
    return base64.b64decode(r.json()["audio_base64"])

Model variants

HuggingFace: OpenMOSS-Team/MOSS-TTS-Nano ModelScope: openmoss/MOSS-TTS-Nano

Supported languages

Chinese (Mandarin) · English · (expanding via finetuning)

License

Apache 2.0 — free for commercial use.

References

• Repo: https://github.com/OpenMOSS/MOSS-TTS-Nano
• Paper: https://arxiv.org/abs/2603.18090
• Demo: https://openmoss.github.io/MOSS-TTS-Nano-Demo/
• HF Space: https://huggingface.co/spaces/OpenMOSS-Team/MOSS-TTS-Nano