Application of Deep Learning Models for Bone-Conducted Speech Signals Extracted in the Form of Bone Conduction Headphones

In this study, we used deep learning to align bone-conducted speech signals with air-conducted speech signals, aiming to replace traditional air conduction microphones in voice-based services capturing surrounding sounds. We fabricated headphones, placing bone conduction microphones on the rami (the branches of a bone in the jaw area), in line with traditional bone conduction headphone configurations. Using LSTM, CNN, and CRNN models, we created databases that aligned bone-conducted speech signals with their air-conducted counterparts and tested them with bone-conducted speech signals captured via our custom-made headphones. The CNN model demonstrated superior performance in accurately distinguishing three English words (“apple,” “hello,” and “pass”), including their voiceless pronunciations. In conclusion, our study shows that deep learning models can effectively use bone-conducted speech signals extracted from the rami for automatic speech recognition (ASR), paving the way for future ASR technology that precisely recognizes only the speaker’s voice.