Towards reconstructing intelligible speech from the human auditory cortex

Abstract

Auditory stimulus reconstruction is a technique that finds the best approximation of the acoustic stimulus from the population of evoked neural activity. Reconstructing speech from the human auditory cortex creates the possibility of a speech neuroprosthetic to establish a direct communication with the brain and has been shown to be possible in both overt and covert conditions. However, the low quality of the reconstructed speech has severely limited the utility of this method for brain-computer interface (BCI) applications. To advance the state-of-the-art in speech neuroprosthesis, we combined the recent advances in deep learning with the latest innovations in speech synthesis technologies to reconstruct closed-set intelligible speech from the human auditory cortex. We investigated the dependence of reconstruction accuracy on linear and nonlinear (deep neural network) regression methods and the acoustic representation that is used as the target of reconstruction, including auditory spectrogram and speech synthesis parameters. In addition, we compared the reconstruction accuracy from low and high neural frequency ranges. Our results show that a deep neural network model that directly estimates the parameters of a speech synthesizer from all neural frequencies achieves the highest subjective and objective scores on a digit recognition task, improving the intelligibility by 65% over the baseline method which used linear regression to reconstruct the auditory spectrogram. These results demonstrate the efficacy of deep learning and speech synthesis algorithms for designing the next generation of speech BCI systems, which not only can restore communications for paralyzed patients but also have the potential to transform human-computer interaction technologies.

Figure 1

Schematic of the speech reconstruction method. (A) Subjects listened to natural speech sentences. The population of evoked neural activity in the auditory cortex of the listener was then used to reconstruct the speech stimulus. The responsive electrodes in an example subject are shown in red. High and low frequency bands were extracted from the neural data. Two types of regression models and two types of speech representations were used, resulting in four combinations: linear regression to auditory spectrogram (light blue), linear regression to vocoder (dark blue), DNN to auditory spectrogram, and DNN to vocoder (dark red). (B) The input to all models was a 300 ms sliding window containing both low frequency (LF) and the high-gamma envelope (HG). The DNN architecture consists of two modules: feature extraction and feature summation networks. Feature extraction for auditory spectrogram reconstruction was a fully connected neural network (FCN). For vocoder reconstruction, the feature extraction network consisted of an FCN concatenated with a locally connected network (LCN). The feature summation network is a two-layer fully connected neural network (FCN). (C) Vocoder parameters consist of spectral envelope, fundamental frequency (f0), voicing, and aperiodicity (total of 516 parameters). An autoencoder with a bottleneck layer was used to reduce the 516 vocoder parameters to 256. The bottleneck features were then used as the target of reconstruction algorithms. The vocoder parameters were calculated from the reconstructed bottleneck features using the decoder part of the autoencoder network.

Figure 2

Deep neural network architecture (A) An original auditory spectrogram of a speech sample is shown on top. The reconstructed auditory spectrograms of the four models are shown below. (B) Magnitude power of frequency bands during an unvoiced (t = 1.4 sec) and a voiced speech sound (t = 1.15 sec, shown with dashed lines in A) for original (top) and the four reconstruction models.

Figure 3

Subjective evaluation of the reconstruction accuracy. (A) The behavioral experiment design used to test the intelligibility and the quality of the reconstructed digits. Eleven subjects listened to digit sounds (zero to nine) spoken by two male and two female speakers. The subjects were asked to report the digit, the quality on the mean-opinion-scale, and the gender of the speaker. (B) The intelligibility score for each model defined as the percentage of correct digits reported by the subject. (C) The quality score on the MOS scale. (D) The speaker gender identification rate for each model. (E) The digit confusion patterns for each of the four models. The DNN vocoder shows the least amount of confusion among the digits.

Figure 4

Objective intelligibly scores for different models. (A) The average ESTOI score based on all subjects for the four models. (B) Coverage and the location of the electrodes and ESTOI score for each of the five subjects. In all subjects, the ESTOI score of the DNN vocoder was higher than in the other models.

Figure 5

Effect of neural frequency range, number of electrodes, and stimulus duration on reconstruction accuracy. (A) The reconstruction ESTOI score based on high gamma, low frequency, and high gamma and low frequency combined. (B) The accuracy of reconstruction when the number of electrodes increases from one to 128. For each condition, 20 random subsets were chosen. (C) The accuracy of reconstruction when the duration of the training data increases. Each condition is the average of 20 random subsets.