Post-processing

To ensure high-quality and synchronized data, several post-processing steps were applied after recording:

1. High-Pass Filtering

A 5th-order Butterworth high-pass filter with a 50 Hz cutoff was applied to the throat microphone (accelerometer) data to remove low-frequency noise, including gravitational acceleration components.

2. Timing Alignment

To correct mismatches caused by timing differences between throat and acoustic microphone signals, synchronization adjustments were applied. This alignment step is crucial for training robust deep learning models.

Alignment of throat and acoustic microphone signals using cross-correlation

3. Noise Reduction

Minor background noise in acoustic microphone recordings was removed using the pretrained causal version of the Demucs speech enhancement model.

Signal waveforms before and after noise enhancement

4. Silence Trimming & Manual Review

Silent segments at the beginning and end of each utterance were manually trimmed. Every recording was reviewed to ensure accurate pronunciation and sentence alignment.

5. Upsampling

The throat microphone recordings were upsampled from 8 to 16 kHz to match the sampling rate of the acoustic recordings.