New Lightweight TTS Stack Released for Microcontrollers
A new text-to-speech (TTS) software stack, dubbed sanoTTS, has been released by developer ampixa, targeting low-power microcontrollers with limited hardware resources. According to documentation hosted on GitHub, the model family is designed to operate on inexpensive hardware, including $3 microcontrollers with 512kb of SRAM and no dedicated Neural Processing Unit (NPU).
The release includes a range of model sizes, with parameter counts spanning from 294k to 2.2 million. The developer claims these models are significantly smaller than existing industry benchmarks, such as Kokoro and Voxtral TTS. The smallest version, when quantized into int8, occupies approximately 337kb of memory.
Performance Benchmarks and Capabilities
The developer reports that the sanoTTS stack supports 11 voices across six languages. In performance testing shared on Reddit, the 1.51m parameter model, “sanoTTS-Amy,” achieved a SCOREQ of 4.13 and a UTMOS of 4.10. These figures reportedly outperform larger models such as Inflect Nano (4.63m) and KittenTTS (15m) in specific metrics.
On an ESP32 microcontroller, the developer reports a Real-Time Factor (RTF) of 0.225, which translates to the generation of four seconds of audio in approximately one second. The developer also noted a Word Error Rate (WER) of approximately 2% when using Whisper for speech-to-text tasks.
The software is available for web integration via WebAssembly through the npm install sanotts-web package. The project includes a recipe for developers to extend the model to additional languages and voices.
Technical Implementation and Accessibility
The release has drawn interest from the developer community regarding its potential for integration into C++ game engines and embedded systems to provide voice generation without the need for a GPU.
The development approach appears to utilize a distillation method, where an existing, larger “teacher” TTS model is used to train the smaller, more efficient “student” model. This method is intended to simplify the training process and reduce computational costs.
While the technical specifications and benchmarks have been published on GitHub and Hugging Face, these performance claims currently lack independent third-party verification. The project remains open-source, allowing for community inspection and contribution.