Improved evaluation of waveform reconstruction in speech decoding based on invasive brain-computer interfaces.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) reconstruct speech from neural signals, but lack standardized objective metrics for evaluating waveform quality.
• Existing metrics like correlation coefficient (CC) and mel cepstral distortion (MCD) are inconsistently applied and have limitations.

Purpose of the Study:

• To address the critical need for a robust and validated method to evaluate reconstructed speech waveform quality in BCIs.
• To identify a standardized objective evaluation metric for the speech BCI field that accurately predicts subjective listener ratings.

Main Methods:

• Reviewed literature on waveform reconstruction from intracranial signals and identified issues with current evaluation methods.
• Collected Mean Opinion Scores (MOS) from human raters on reconstructed audio from 10 published speech BCI studies.
• Systematically evaluated combinations of objective metrics (STOI, MCD) using leave-one-dataset-out cross-validation and compared linear/non-linear regression models.

Main Results:

• A non-linear Random Forest regressor model demonstrated the highest accuracy (R² = 0.892) in predicting subjective MOS ratings.
• The proposed model accurately maps STOI and MCD objective metrics to predicted MOS scores.
• The Random Forest model significantly outperforms existing methods in predicting perceptual quality.

Conclusions:

• The study identifies a critical lack of standardized evaluation methods in speech BCI research, hindering cross-study comparisons.
• A cross-validated Random Forest model is proposed as a standardized objective metric for speech BCI waveform quality assessment.
• This metric provides a reliable tool for benchmarking, facilitating comparisons, and accelerating advancements in speech neuroprosthetics.