Closed-System Speech POC
Reconstructing canonical speech waveforms for a closed phrase set with lightweight LiteRT-friendly models.
About
This repo is a proof-of-concept for reconstructing canonical speech waveforms for a closed set of phrases using lightweight models and LiteRT-friendly ops. It provides a CLI (main.py) and scripts to generate training assets from macOS `say`, train a small banded Keras reconstruction model, evaluate MAE/MSE against targets, and convert the SavedModel to TFLite with an optional parity check. A local dashboard supports adding words via TTS or microphone, recording mic audio to replace targets (with denoise/filter options), populating phoneme entries, training banded models, running validation/postulate modes, and generating text-only postulates. The demo surfaces a LiteRT web runtime and HTML pages for speech and TFLite demos.
Summary
An experimental proof-of-concept that reconstructs canonical speech waveforms for a closed set of phrases using lightweight Keras models and LiteRT-friendly ops, with a local dashboard for training and validation.
Tech
- Python
- TensorFlow
- Keras
- ai_edge_litert
- TFLite
- LiteRT.js
- numpy
- pronouncing
- HTML
- JavaScript
- WASM
Tags
- litert
- proof-of-concept
- signal-processing
- speech-synthesis
- tensorflow