roatienza/efficientspeech

Implements a shallow 2-block pyramid transformer with depth-wise separable convolutions for mel-spectrogram generation, achieving 104× real-time factor on Raspberry Pi 4 with just 266k parameters in its tiny variant. Supports multiple model scales (266k to 4M parameters), CPU/GPU inference, PyTorch 2.0 compilation, and ONNX export for edge deployment. Trained on LJSpeech with forced alignment data, compatible with both eager and compiled execution modes across embedded and high-performance hardware.

180 stars. No commits in the last 6 months.