moss-tts-nano-speech — Claude Code skill: local CPU text-to-speech

What moss-tts-nano-speech does

MOSS-TTS-Nano comes from MOSI.AI and the OpenMOSS team: 0.1B parameters, an audio tokenizer paired with an autoregressive LLM, stereo output at 48 kHz. Twenty languages are covered, and the token stream runs at 12.5 Hz. The skill hands the agent working recipes — the moss-tts-nano generate CLI, infer.py, the MossTTSNanoInference Python class, and the HTTP routes of the local server.

Voice cloning takes a reference WAV through --prompt-speech or --prompt-audio-path, and the same timbre carries across the whole text. Long documents are split automatically while the speaker stays the same. Streaming via infer_stream returns chunks as they are produced, which cuts the wait before the first sound.

Setup runs through conda with Python 3.12: conda create -n moss-tts-nano python=3.12 -y, then clone the repo, pip install -r requirements.txt and pip install -e .. If WeTextProcessing fails to build, the instructions include a workaround with pynini and a GitHub fork. Weights download from Hugging Face on the first run, so that run needs a network connection.

The local web demo starts with moss-tts-nano serve and answers on http://127.0.0.1:18083; the model stays in memory, so later requests skip the load. Language is picked up from the text and the reference audio, no code parameter needed.