DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Silent speech with ultrasound

Quality: 8/10 Relevance: 9/10

Summary

Researchers train a model to decode silent speech from ultrasound tongue imaging, achieving 15.6% WER on open-vocabulary cross-speaker tests with about 50 hours of data. The approach combines ResNet-18 2+1d video encoding with Whisper-based decoding, highlighting a path toward private, voice-free interaction while noting hardware challenges and the early prototype status.

🚀 Service construit par Johan Denoyer