sanoTTS: Ultra-lightweight Neural Text-to-Speech Stack for Microcontrollers and Web
Developer ampixa released sanoTTS, an open-source neural text-to-speech (TTS) model family ranging from 294k to 2.2m parameters. The smallest model consumes just 337 KB when quantized to INT8, allowing offline TTS generation on $3 microcontrollers like the ESP32 without requiring an NPU. By drastically shrinking TTS models—up to 1,000x smaller than existing open-weight models like Kokoro—sanoTTS enables high-quality, real-time voice synthesis on resource-constrained IoT devices and web browsers. This breakthrough paves the way for fully offline smart home gadgets, embedded applications, and web applications without relying on cloud APIs. The 1.51m parameter variant (sanoTTS-Amy) achieves a SCOREQ of 4.13 and UTMOS of 4.10, outperforming significantly larger models like KittenTTS (15m params). On an ESP32 microcontroller with 512 KB SRAM, it achieves a Real-Time Factor (RTF) of 0.225 (generating 4 seconds of audio in 1 second) with around 2% Word Error Rate (WER) measured by Whisper.
## BACKGROUND
Text-to-Speech (TTS) models use neural networks to convert written text into natural-sounding speech, but traditional deep learning TTS models typically require high-end GPUs or gigabytes of memory. UTMOS and SCOREQ are objective neural metrics used to assess perceptual speech quality, where higher scores indicate superior audio fidelity. The Real-Time Factor (RTF) measures processing speed relative to audio duration; an RTF below 1.0 means speech generation occurs faster than real-time playback.