Speech synthesisPreview
NanoTTS
A CPU inference server for small speech synthesis models, rewritten for multi-threaded serving with an OpenAI-compatible HTTP API.
What it does
Small models, fast on a CPU
Kyutai's Pocket TTS and its fine-tunes, and the Supertonic 3 family: Supertonic 3 and TeraTTS v2.
OpenAI-compatible API
An HTTP API that follows the OpenAI shape, with streaming from the first audio frame.
Tuned for serving
Multi-threaded and NUMA-aware: weights are replicated per NUMA node, and warmed voices start from a memory copy.
Many languages
English, French, German, Italian, Portuguese, Spanish, Russian, Korean and Japanese, depending on the model you pick.
One command
The installer builds the runtime image in Docker and starts the server on 127.0.0.1:8080. A launcher, nanotts, manages it.
Install
curl -fsSL https://raw.githubusercontent.com/usunrise88/NanoTTS/main/install.sh | bashPreview: it installs from source, there is no tagged release yet. Needs Linux on x86-64 (AVX2 preferred), Docker with Compose v2, git and about 12 GiB of disk.