tiny-audio, nanoGPT for speech-to-text (github.com)

The launch of "Tiny Audio," a cost-effective speech-to-text system, has made waves in the AI/ML community by enabling users to train a high-performance model for just $25. This innovative framework connects a frozen, pretrained speech encoder with a small, trainable projector and a pretrained language model, achieving a remarkable 1.8% word error rate (WER) on the LibriSpeech test-clean benchmark and 7.4% across a diverse pool of benchmarks while utilizing only about 80 million parameters. Users can quickly run training loops on personal laptops and deploy live demos without complex installations. The significance of Tiny Audio lies in its accessibility and efficiency. By offering a straightforward implementation through a live demo, users can upload audio files or record directly to obtain transcripts with minimal setup. It supports features like word-level timestamps and speaker diarization in a zero-shot manner, enhancing its utility in diverse applications. Additionally, the minimal hardware requirements and straightforward Python integration make it an appealing choice for developers and researchers looking to experiment with speech recognition technology. The codebase's portability and simplicity allow for extensive customization, fostering innovation within the AI community.