Voice AI / Space
    Submit
    Back

    sanoTTS

    View on GitHub
    Git Repo
    Ampixa

    Implements small neural text-to-speech models that run inference locally on low-power microcontrollers and inside web browsers.

    About sanoTTS

    sanoTTS is an ultra-lightweight neural text-to-speech engine designed to run entirely on low-cost microcontrollers or directly within a web browser without relying on cloud servers or dedicated NPUs.

    For the Non-Technical Reader

    Imagine giving physical objects—like a $3 smart toy, an offline medical device, or an interactive greeting card—their own distinct human voice without needing an internet connection or paying monthly cloud API fees. Because the entire AI voice stack fits in a tiny footprint under 4 MB, audio synthesis happens instantly and completely offline right on the device or browser.

    For the Technical Reader

    sanoTTS offers a model family ranging from 294k to 2.3M parameters, leading quality benchmarks (SCOREQ/UTMOS) in the sub-15M parameter class. Key technical specs include:

    • Hardware Support: Runs real-time on a ~$3 ESP32-S3 microcontroller (audio output via GPIO into an LM386 amp) and live in-browser via WebAssembly (WASM).
    • Inference & Dependencies: Pure NumPy inference for Python without Torch or ONNX Runtime dependencies; ships with an embedded espeak-ng WASM phonemizer.
    • Quantization & Footprint: Features heart-nano, an int8 quantized voice packing a complete TTS engine into just 337 KB of weights.
    • Multilingual Capabilities: Offers 11 voices across 6 languages (English, Nepali, Hindi, Vietnamese, Indonesian, Chinese).
    • Licensing: GPL-3.0 open source.

    Why It Matters

    By drastically cutting model size without sacrificing speech clarity, sanoTTS eliminates the financial and privacy constraints of cloud voice APIs. It enables edge-native, zero-latency speech generation on low-power hardware, drastically dropping infrastructure costs for mass-market hardware deployments and web applications.

    The "Voice AI Space Lab" Idea

    Offline Smart Companion Badge: Build an offline, interactive event badge or toy using an ESP32-S3 chip and a small speaker. When users press buttons or interact with sensors, the device dynamically synthesizes speech localized to six languages entirely offline without needing Wi-Fi or cellular connectivity.

    Explore the project further on GitHub, test the live client-side synthesis via the Interactive Browser Demo, or inspect model packages on Hugging Face.