Inflect-Nano offers a tiny local voice model for agents
Inflect-Nano-v1 is a very small TTS model that creates one English male voice. It has 4.63 million total for inference, with 3.46 million in the acoustic model and 1.17 million in the vocoder.
It produces 24 kHz audio and can run locally with a simple PyTorch inference script. Its size is about 17 times smaller than Kokoro, about 108 times smaller than , and almost 1000 times smaller than Fish Audio S2 Pro.
The quality is limited: the voice can sound robotic, difficult new text can cause mistakes, and the vocoder is a major weak point. Even so, it is a useful baseline for extremely small local , offline , embedded devices, browser or WASM projects, and local .
Key points
- Inflect-Nano-v1 is an English-only TTS model with one male voice.
- It uses 4.63 million total for inference.
- It creates 24 kHz audio and through a PyTorch script.
- It is much smaller than Kokoro, , and Fish Audio S2 Pro by the stated comparisons.
- Its limits include robotic sound, mistakes on hard new text, and a weak vocoder.