DigiNews

Summary

The article describes a research project that uses ultrasound to capture tongue movements and convert silent speech into text, achieving 15.6% word error rate on open vocabulary speech with about 50 hours of data. It explains the modeling approach, including aligning video embeddings with Whisper's embeddings, data collection practices, and future hardware and accuracy improvements. The piece explores implications for privacy and human-computer interaction, noting current limitations and potential wearable implementations.