tiny-audio: Open-Source Speech-to-Text with Frozen Encoder and LLM
Summary
Tiny Audio is an open-source speech-to-text system that links a frozen speech encoder to a pretrained LLM via a small trainable projector. The readme details training workflows, word-level timestamps, speaker diarization, a batched HTTP API, and a live demo, plus practical steps to run locally or on RunPod. It emphasizes affordability and hands-on experimentation with WER metrics across multiple datasets.