⬢github Python · 168 ★ +14 since we first saw it · pushed 2 d ago · MIT
alexkroman/tiny-audio
A speech-to-text system you can train for $25. A frozen speech encoder and a Qwen LLM joined by a small trained projector, with word timestamps, speaker diarization, a batched HTTP server, a live demo, and a free course.
Tiny Audio is an open-source speech-to-text model built by connecting a frozen pretrained speech encoder to a Qwen LLM through a small trainable projector (~80M trained parameters). It reports 1.8% WER on LibriSpeech test-clean, adds word timestamps, speaker diarization, and streaming, and can be served as a batched HTTP API on RunPod or locally. Training reportedly costs about $25.
Why now: Recently featured on Hacker News as 'nanoGPT for speech-to-text', thanks to its unusually cheap trainable training recipe, strong benchmarks, live demo, and free course.
Who it is for: Developers and ML learners who want a small, hackable, and cheap speech-to-text stack with timestamps, diarization, and self-hostable serving.
asraudio-llmforced-alignmenthuggingfacellmloramultimodalpipecatpytorchqwen
Stars over our 135 snapshots: 154 to 168, since 1 d ago.
Where people talked about it
- Mmastodon tiny-audio, nanoGPT for speech-to-text https://github.com/alexkroman/tiny-audio # HackerNews # Tech # SpeechRecognition 1 d ago
- Yhn tiny-audio, nanoGPT for speech-to-text 1 d ago
API: https://socialmediatrends-api.osmike.com/v1/repos/alexkroman/tiny-audio