Direct decoding of speech from the brain is a faster alternative to current electroencephalography (EEG) speller-based brain-computer interfaces (BCI) in providing communication assistance to locked-in patients. Magnetoencephalography (MEG) has recently shown great potential as a non-invasive neuroimaging modality for neural speech decoding, owing in part to its spatial selectivity over other high-temporal resolution devices. Standard MEG systems have a large number of cryogenically cooled channels/sensors (200 - 300) encapsulated within a fixed liquid helium dewar, precluding their use as wearable BCI devices. Fortunately, recently developed optically pumped magnetometers (OPM) do not require cryogens, and have the potential to be wearable and movable making them more suitable for BCI applications. This design is also modular allowing for customized montages to include only the sensors necessary for a particular task. As the number of sensors bears a heavy influence on the cost, size, and weight of MEG systems, minimizing the number of sensors is critical for designing practical MEG-based BCIs in the future. In this study, we sought to identify an optimal set of MEG channels to decode imagined and spoken phrases from the MEG signals. Using a forward selection algorithm with a support vector machine classifier we found that nine optimally located MEG gradiometers provided higher decoding accuracy compared to using all channels. Additionally, the forward selection algorithm achieved similar performance to dimensionality reduction using a stacked-sparse-autoencoder. Analysis of spatial dynamics of speech decoding suggested that both left and right hemisphere sensors contribute to speech decoding. Sensors approximately located near Broca's area were found to be commonly contributing among the higher-ranked sensors across all subjects.

Neural speech decoding-driven brain-computer interface (BCI) or speech-BCI is a novel paradigm for exploring communication restoration for locked-in (fully paralyzed but aware) patients. Speech-BCIs aim to map a direct transformation from neural signals to text or speech, which has the potential for a higher communication rate than the current BCIs. Although recent progress has demonstrated the potential of speech-BCIs from either invasive or non-invasive neural signals, the majority of the systems developed so far still assume knowing the onset and offset of the speech utterances within the continuous neural recordings. This lack of real-time voice/speech activity detection (VAD) is a current obstacle for future applications of neural speech decoding wherein BCI users can have a continuous conversation with other speakers. To address this issue, in this study, we attempted to automatically detect the voice/speech activity directly from the neural signals recorded using magnetoencephalography (MEG). First, we classified the whole segments of pre-speech, speech, and post-speech in the neural signals using a support vector machine (SVM). Second, for continuous prediction, we used a long short-term memory-recurrent neural network (LSTM-RNN) to efficiently decode the voice activity at each time point via its sequential pattern-learning mechanism. Experimental results demonstrated the possibility of real-time VAD directly from the non-invasive neural signals with about 88% accuracy.

Speech production involves the synchronization of neural activity between the speech centers of the brain and the oral-motor system, allowing for the conversion of thoughts into meaningful sounds. This hierarchical mechanism is hindered due to partial or complete paralysis of the articulators for patients suffering from locked-in-syndrome. These patients are in dire need of effective brain-communication interfaces (BCIs), which can at least provide a level of communication assistance. In this study, we tried to decode overt (loud) speech directly from the brain via non-invasive magnetoencephalography (MEG) signals to build the foundation for a faster, direct brain to text mapping BCI. A shallow Artificial Neural Network (ANN) was trained with wavelet features of the MEG signals for this objective. Experimental results show that a direct speech decoding is possible from MEG signals. Moreover, we found that the jaw motion and MEG signals may have complimentary information for speech decoding.

Decoding speech directly from the brain has the potential for the development of the next generation, more efficient brain computer interfaces (BCIs) to assist in the communication of patients with locked-in syndrome (fully paralyzed but aware). In this study, we have explored the spectral and temporal features of the magnetoencephalography (MEG) signals and trained those features with convolutional neural networks (CNN) for the classification of neural signals corresponding to phrases. Experimental results demonstrated the effectiveness of CNNs in decoding speech during perception, imagination, and production tasks. Furthermore, to overcome the long training time issue of CNNs, we leveraged principal component analysis (PCA) for spatial dimension reduction of MEG data and transfer learning for model initialization. Both PCA and transfer learning were found to be highly beneficial for faster model training. The best configuration (50 principal coefficients + transfer learning) led to more than 10 times faster training than the original setting while the speech decoding accuracy remained at a similarly high level.