ItoTTS Brings High-Quality Text-to-Speech to $5 ESP32-S3 Microcontrollers
The Lokutor team released ItoTTS, a compact streaming text-to-speech (TTS) engine capable of synthesizing 24 kHz natural English audio on a $5 ESP32-S3 microcontroller. It offers two distinct voices requiring only 4.89 MB of model weights per voice while scoring a 4.46 UTMOS naturalness score, closely approaching its teacher model, StyleTTS 2. By bringing high-fidelity speech synthesis to ultra-low-cost, resource-constrained microcontrollers, ItoTTS enables fully local, private voice interfaces for smart home hardware and embedded AI gadgets. It lowers the hardware barrier for voice-enabled local LLMs without relying on cloud APIs. The codebase is licensed under GPLv3 while model weights are released under CC BY-NC-SA 4.0 for non-commercial use to prevent large corporations from commercializing the work without authorization. Current limitations include relying on a host machine for text-to-phoneme conversion and unmeasured processing speed on physical hardware.
## BACKGROUND
Text-to-speech (TTS) on embedded microcontrollers like the ESP32-S3 traditionally suffers from robotic audio quality due to strict memory and compute limitations. UTMOS is a neural network-based metric that predicts speech naturalness scores, while StyleTTS 2 is a state-of-the-art text-to-speech model that serves as the teacher network for training smaller models like ItoTTS.