GoKawiilOpen-source project Tiny Audio trains speech-to-text models for about $25
Tiny Audio is a GitHub project that pairs a frozen pretrained speech encoder with a pretrained LLM through a small trainable projector, requiring only about 80 million trainable parameters. The released model reaches 1.8% word error rate on LibriSpeech test-clean and 7.4% across a pooled set of 12 benchmarks, and the codebase can reportedly be read in an afternoon or trained on a laptop in roughly five minutes. It also supports word-level timestamps and speaker diarization via a zero-shot port of NVIDIA's NeMo masked_asr recipe.
GoKawiil's interpretation of the reporting above, not reported fact.
By freezing large pretrained components and training only a lightweight projector, the project suggests competitive speech recognition no longer requires massive compute budgets, which could lower the barrier for hobbyists and small teams to build custom ASR systems. The inclusion of diarization and timestamping out of the box hints at ambitions beyond basic transcription, toward meeting-notes and multi-speaker use cases, though the project notes overlapping speech is only partially handled. Its small, readable codebase could also make it a reference implementation for others studying encoder-LLM speech architectures.
- Trains with only ~80M trainable parameters by freezing a pretrained speech encoder and LLM
- Achieves 1.8% WER on LibriSpeech test-clean and 7.4% across 12 pooled benchmarks
- Supports word-level timestamps and speaker diarization, though overlapping speech handling is limited
Source: github.com, 2026-10-09
Published there as: “tiny-audio, nanoGPT for speech-to-text”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.