Babytalk

Babytalk is an offline speech recognition and synthesis system that runs on ESP32 microcontrollers.

BabyTalk is an offline speech-to-text and text-to-speech library designed for ESP32-S3 and ESP32-P4 microcontrollers. It uses compressed NVIDIA Citrinet-256 models for speech recognition and sanoTTS for text synthesis, allowing devices to transcribe audio and speak replies locally. The system operates entirely without cloud services or internet access, keeping data private and functional during network outages. It supports C, MicroPython, and Erlang or Elixir via AtomVM, making it suitable for headless hardware like LoRa nodes and voice-controlled appliances. The custom MMRT inference engine executes models directly from flash memory while using dual-core processing to maintain efficient execution.

Babytalk at a glance

Best for: Software engineer
Runs on: ESP32-S3, ESP32-P4, MicroPython, AtomVM
Pricing: Open source under the MIT license
Released: October 2026
Category: AI audio tool
Website: github.com/tlack/babytalk

Key features

Fully offline speech-to-text transcription
On-device text-to-speech synthesis
Custom MMRT inference engine for flash execution
Support for custom wake phrases without training
Dual-core weight processing and layer fusion
Configurable 4-bit and 8-bit model quantization

Use it for

Voice input for keyboards-free LoRa radio nodes
Reading text messages aloud on screen-free speakers
Offline voice control for embedded hardware

Alternatives to Babytalk

Moonshine — Runs entirely in your browser with no audio leaving your device.
Esp32 Pie ASR — Faster than real time streaming speech recognition on a microcontroller alone
Speech-to-Speech — Fully modular, swappable components with local or hosted execution options.
sherpa-onnx — Runs complex audio AI tasks offline on local devices and embedded hardware.
Piper — Runs locally and offers fast neural text-to-speech synthesis.
Superwhisper — Transcribes speech offline and online with custom AI modes and templates.
Whistle Esp32 — Transcribes speech and executes commands locally on a ten-dollar microcontroller.
WeNet — Achieves state-of-the-art results on public speech datasets.
FluidAudio — Runs local audio AI models on Apple Neural Engine.
ESPnet — Provides reproducible recipes and hundreds of pretrained models for various speech tasks.
Willow — Fast, private, open source voice control using local hardware.
kokoro-onnx — Runs local TTS inference with efficient ONNX runtime performance.

All Babytalk alternatives