Releases·Model Release
Dev releases sanoTTS: 294k-param (337 KB) TTS stack runs on $3 ESP32; 1.5m model beats Inflect Nano and KittenTTS on SCOREQ
A developer released sanoTTS, a text-to-speech stack built while attempting to fit TTS onto a $3 chip with 512 KB of SRAM and no NPU. The release spans 294k to 2.2m parameters with 11 voices and 6 languages, with the 337 KB smallest model running fully on the microcontroller and a 1.46m-parameter model claimed to beat models 3x and 10x its size.
546 engagement1 sourceThu, Sep 3, 2026, 10:01 PM
AI Overview
- sanoTTS spans 294k–2.2m params with 11 voices and 6 languages, 1000x smaller than Kokoro and 9000x smaller than Voxtral TTS
- The 1.5m model scores SCOREQ 4.13 and UTMOS 4.10, beating Inflect Nano (4.63m, 3.81) and KittenTTS (15m, 3.02)
- 294k model quantizes to int8 at 337 KB; author claims smallest neural TTS ever with ~2% WER on Whisper
- On ESP32 (512 KB SRAM, no NPU) sanoTTS hits RTF 0.225, generating 4 seconds of audio in 1 second
- Runs in-browser via WebAssembly (npm install sanotts-web); recipe provided to extend languages and voices; code on GitHub and HF
Entities
Related
Products·launched
sanoTTS — tiny neural text-to-speech running on microcontrollers, 30 voices in 16 languages
sanoTTS is a family of very small neural TTS models (from 337 KB to 2.27M parameters) with 30 voices across 16 languages. The smallest voice runs offline on a ~$3 ESP32-S3 dev board with no server, API key, or internet. It targets developers who need speech synthesis on-device in embedded or browser environments.
sanoTTS is a family of very small neural TTS models (from 337 KB to 2.27M parameters) with 30 voices across 16 languages. The smallest voice runs offline on a ~$3 ESP32-S3 dev board with no server, API key, or internet. It targets developers who need speech synthesis on-device in embedded or browser environments.
1 engagement1 sourceredditWed, Sep 9, 2026, 08:43 PM
Products·launched
sanoTTS anatomy — interactive visualization of a 294,279-parameter TTS model's internals
A website that visualizes the internals of sanoTTS, a tiny text-to-speech system with 294,279 parameters. It was vibe coded to show every real intermediate tensor captured from the shipped int8 model while synthesizing an actual sentence, helping people learn how a small TTS model works.
A website that visualizes the internals of sanoTTS, a tiny text-to-speech system with 294,279 parameters. It was vibe coded to show every real intermediate tensor captured from the shipped int8 model while synthesizing an actual sentence, helping people learn how a small TTS model works.
2 engagement1 sourceredditSun, Sep 20, 2026, 08:30 AM
Developments·Paper
GrainSpeech: 264.8K-Parameter Model Hits 17.9x Real-Time Speech Synthesis on Microcontrollers
A new arXiv paper introduces GrainSpeech, a compact acoustic model for Mel-spectrogram generation that uses a fixed-receptive-field convolutional encoder and a Mel-specific gradient-variance supervision scheme. At only 264.8K parameters, it matches UTMOS scores of far larger models and runs 17.9x real-time on a microcontroller, making on-device speech synthesis practical for embedded developers.
A new arXiv paper introduces GrainSpeech, a compact acoustic model for Mel-spectrogram generation that uses a fixed-receptive-field convolutional encoder and a Mel-specific gradient-variance supervision scheme. At only 264.8K parameters, it matches UTMOS scores of far larger models and runs 17.9x real-time on a microcontroller, making on-device speech synthesis practical for embedded developers.
0 engagement1 sourcearxivWed, Sep 16, 2026, 03:59 PM
Products·building
ItoTTS: streaming on-device TTS engine for ESP32-S3 with two English voices
ItoTTS is a small, streaming text-to-speech engine that runs on a $5 ESP32-S3 microcontroller, providing two natural English voices at 24 kHz with 4.89 MB of weights per voice. It is part of the Lokutor team's plan (following Oído) to give local LLMs a voice entirely on-device, without cloud services. It targets hobbyists and developers building offline voice assistants on cheap hardware.
ItoTTS is a small, streaming text-to-speech engine that runs on a $5 ESP32-S3 microcontroller, providing two natural English voices at 24 kHz with 4.89 MB of weights per voice. It is part of the Lokutor team's plan (following Oído) to give local LLMs a voice entirely on-device, without cloud services. It targets hobbyists and developers building offline voice assistants on cheap hardware.
3 engagement1 sourceredditMon, Oct 5, 2026, 01:38 PM
Products·launched
Sopro V2: open-source CPU-friendly TTS with voice cloning
Sopro V2 is an open-source text-to-speech model that performs zero-shot voice cloning from 5-20 seconds of reference audio. It supports English, European Portuguese, French, and German, and is designed to run locally on CPU with a simple web UI. It solves the problem of accessible, high-quality voice cloning for developers and hobbyists.
Sopro V2 is an open-source text-to-speech model that performs zero-shot voice cloning from 5-20 seconds of reference audio. It supports English, European Portuguese, French, and German, and is designed to run locally on CPU with a simple web UI. It solves the problem of accessible, high-quality voice cloning for developers and hobbyists.
38 engagement1 sourceredditThu, Aug 27, 2026, 02:38 PM