Tiny Audio Achieves 1.8% WER via Small Trainable Projectors
October 9, 2026
Tiny Audio achieves 1.8% WER on LibriSpeech test-clean by connecting a frozen speech encoder to a pretrained LLM using an 80M parameter trainable projector. The system can be trained for approximately $25 and requires only 6 GB of memory for inference.
HOW THIS AFFECTS YOU
●
builderYou can deploy highly efficient speech-to-text capabilities on consumer hardware or edge devices with minimal training costs.