Tiny Audio is a speech-to-text system combining a frozen speech encoder with a pretrained LLM and small trainable projector, achieving 1.8% WER on LibriSpeech test-clean while training only ~80M parameters for $25. The system supports word-level timestamps and speaker diarization, can be deployed locally or on cloud GPUs via an HTTP server with batching, and is available as a live demo or Python package.