Ampixa/sanoTTS

sanoTTS (सानो = 'small' in Nepali): a ~1.4M-param neural TTS that runs on a $3 chip or in the browser. Leads SCOREQ/UTMOS in the sub-15M class.

What it solves

sanoTTS provides a family of extremely small neural text-to-speech (TTS) voices that can run locally on low-power hardware without requiring a cloud connection, a dedicated Neural Processing Unit (NPU), or a server. It enables real-time speech synthesis on devices as simple as a $3 ESP32-S3 microcontroller or directly within a web browser via WebAssembly (WASM).

How it works

The system uses a multi-stage pipeline: an espeak-ng phonemizer converts text to phoneme IDs, a duration model predicts timing, an acoustic model predicts generator latents, and a decoder renders the final audio. Depending on the target platform, it uses different decoders: a compact time-domain decoder for browsers (fp32 WASM) and a quantized int8 iSTFT decoder for microcontrollers to ensure real-time performance.

Who it’s for

  • Embedded developers building voice-enabled devices on cheap microcontrollers.
  • Web developers wanting to add client-side, serverless TTS to their websites.
  • Python developers looking for a lightweight, dependency-free TTS library.

Highlights

  • Ultra-compact size: Models range from 294k to 2.3M parameters, with the smallest (heart-nano) occupying only 337 KB.
  • Broad language support: Includes 30 voices across 16 languages, including English, Spanish, French, Arabic, and Nepali.
  • High efficiency: Achieves a Real-Time Factor (RTF) of 0.383 on an ESP32-S3, meaning it synthesizes speech 2.6x faster than real-time.
  • Zero dependencies: The Python package uses pure numpy for inference, and the browser version runs entirely client-side.
  • Customizable: Includes a full training recipe for users to create and distill their own voices.

Related

  • Project
    MobiusQuant/OpenMobius-skill

    OpenMobius‑skill is a plug‑in for AI coding agents (Claude Code, Codex, OpenClaw, Hermes, Cursor, WorkBuddy) that adds a curated trading‑knowledge base (≈2 000 cards from ICT, SMC, ChanLun, etc.), real‑time market data, 60+ technical indicators, and chart‑generation. It routes user queries through strict/augment/compare modes, supports multiple schools of analysis, and installs self‑contained with its own Python virtual environment, model weights, and Playwright Chromium.

  • Project
    zh-plus/openlrc

    A Python library that transcribes audio using faster-whisper and uses LLMs to translate and polish the text into synchronized .lrc subtitles.

  • Project
    FireRedTeam/FireRedASR2S

    An industrial-grade all-in-one ASR system that combines speech recognition, voice activity detection, language identification, and punctuation prediction, with specialized support for Chinese dialects.

  • Project
    shinpr/claude-code-workflows

    Claude Code Development Workflows is a set of Claude Code plugins that turn the Claude AI assistant into a structured development pipeline. The plugins provide “recipes” (e.g. `/recipe‑implement`, `/recipe‑design`, `/recipe‑review`) that enforce outcome agreement, design documentation, work‑plan generation, task‑level implementation with automated verification, and an independent security review. Three plugin families cover backend, React frontend, and full‑stack projects, and they can be installed via Claude’s marketplace. The system is useful when you need repeatable safety gates and a formal hand‑off between design and code generation; for quick prototypes you can skip it.