Megadose The AI curve, compressed.

FETTUCCINE: Fast and efficient brain-to-text decoding on mobile devices

· ArXiv · AI/CL/LG ·
FETTUCCINE reached 4.39% word error rate, with mobile decoding throughput up to 370 times real-time.

The framework uses automatic speech recognition models to turn neural activity during attempted speech into text. Evaluated on publicly available recordings from one intracortical speech neuroprosthesis user, it outperformed previous end-to-end methods. The authors report running the models on a commercially available mobile device, calling it the first demonstration of mobile brain-to-text decoding. Fine-tuning also extended usable performance to recordings from later days.

ArXiv · AI/CL/LG's note

score 5

Categories: Research