Single-channel, speaker-independent speech separation methods have recently seen great progress. However, the accuracy, latency, and computational cost of such methods remain insufficient. The majority of the previous methods have formulated the separation problem through the time-frequency representation of the mixed signal, which has several drawbacks, including the decoupling of the phase and magnitude of the signal, the suboptimality of time-frequency representation for speech separation, and the long latency in calculating the spectrograms. To address these shortcomings, we propose a fully-convolutional time-domain audio separation network (Conv-TasNet), a deep learning framework for end-to-end time-domain speech separation. Conv-TasNet uses a linear encoder to generate a representation of the speech waveform optimized for separating individual speakers. Speaker separation is achieved by applying a set of weighting functions (masks) to the encoder output. The modified encoder representations are then inverted back to the waveforms using a linear decoder. The masks are found using a temporal convolutional network (TCN) consisting of stacked 1-D dilated convolutional blocks, which allows the network to model the long-term dependencies of the speech signal while maintaining a small model size. The proposed Conv-TasNet system significantly outperforms previous time-frequency masking methods in separating two-and three-speaker mixtures. Additionally, Conv-TasNet surpasses several ideal time-frequency magnitude masks in two-speaker speech separation as evaluated by both objective distortion measures and subjective quality assessment by human listeners. Finally, Conv-TasNet has a significantly smaller model size and a shorter minimum latency, making it a suitable solution for both offline and real-time speech separation applications. This study therefore represents a major step toward the realization of speech separation systems for real-world speech processing technologies.

How humans solve the cocktail party problem remains unknown. However, progress has been made recently thanks to the realization that cortical activity tracks the amplitude envelope of speech. This has led to the development of regression methods for studying the neurophysiology of continuous speech. One such method, known as stimulus-reconstruction, has been successfully utilized with cortical surface recordings and magnetoencephalography (MEG). However, the former is invasive and gives a relatively restricted view of processing along the auditory hierarchy, whereas the latter is expensive and rare. Thus it would be extremely useful for research in many populations if stimulus-reconstruction was effective using electroencephalography (EEG), a widely available and inexpensive technology. Here we show that single-trial (≈60 s) unaveraged EEG data can be decoded to determine attentional selection in a naturalistic multispeaker environment. Furthermore, we show a significant correlation between our EEG-based measure of attention and performance on a high-level attention task. In addition, by attempting to decode attention at individual latencies, we identify neural processing at ∼200 ms as being critical for solving the cocktail party problem. These findings open up new avenues for studying the ongoing dynamics of cognition using EEG and for developing effective and natural brain-computer interfaces.

During speech perception, linguistic elements such as consonants and vowels are extracted from a complex acoustic speech signal. The superior temporal gyrus (STG) participates in high-order auditory processing of speech, but how it encodes phonetic information is poorly understood. We used high-density direct cortical surface recordings in humans while they listened to natural, continuous speech to reveal the STG representation of the entire English phonetic inventory. At single electrodes, we found response selectivity to distinct phonetic features. Encoding of acoustic properties was mediated by a distributed population response. Phonetic features could be directly related to tuning for spectrotemporal acoustic cues, some of which were encoded in a nonlinear fashion or by integration of multiple cues. These findings demonstrate the acoustic-phonetic representation of speech in human STG.

Humans possess a remarkable ability to attend to a single speaker’s voice in a multi-talker background1–3. How the auditory system manages to extract intelligible speech under such acoustically complex and adverse listening conditions is not known, and, indeed, it is not clear how attended speech is internally represented4,5. Here, using multi-electrode surface recordings from the cortex of subjects engaged in a listening task with two simultaneous speakers, we demonstrate that population responses in non-primary human auditory cortex encode critical features of attended speech: speech spectrograms reconstructed based on cortical responses to the mixture of speakers reveal the salient spectral and temporal features of the attended speaker, as if subjects were listening to that speaker alone. A simple classifier trained solely on examples of single speakers can decode both attended words and speaker identity. We find that task performance is well predicted by a rapid increase in attention-modulated neural selectivity across both single-electrode and population-level cortical responses. These findings demonstrate that the cortical representation of speech does not merely reflect the external acoustic environment, but instead gives rise to the perceptual aspects relevant for the listener’s intended goal.

Speaking is one of the most complex actions we perform, yet nearly all of us learn to do it effortlessly. Production of fluent speech requires the precise, coordinated movement of multiple articulators (e.g., lips, jaw, tongue, larynx) over rapid time scales. Here, we used high-resolution, multi-electrode cortical recordings during the production of consonant-vowel syllables to determine the organization of speech sensorimotor cortex in humans. We found speech articulator representations that were somatotopically arranged on ventral pre- and post-central gyri and partially overlapping at individual electrodes. These representations were temporally coordinated as sequences during syllable production. Spatial patterns of cortical activity revealed an emergent, population-level representation, which was organized by phonetic features. Over tens of milliseconds, the spatial patterns transitioned between distinct representations for different consonants and vowels. These results reveal the dynamic organization of speech sensorimotor cortex during the generation of multi-articulator movements underlying our ability to speak.

Direct brain recordings from neurosurgical patients listening to speech reveal that the acoustic speech signals can be reconstructed from neural activity in auditory cortex.

Robust speech processing in multi-talker environments requires effective speech separation. Recent deep learning systems have made significant progress toward solving this problem, yet it remains challenging particularly in real-time, short latency applications. Most methods attempt to construct a mask for each source in time-frequency representation of the mixture signal which is not necessarily an optimal representation for speech separation. In addition, time-frequency decomposition results in inherent problems such as phase/magnitude decoupling and long time window which is required to achieve sufficient frequency resolution. We propose Time-domain Audio Separation Network (TasNet) to overcome these limitations. We directly model the signal in the time-domain using an encoder-decoder framework and perform the source separation on nonnegative encoder outputs. This method removes the frequency decomposition step and reduces the separation problem to estimation of source masks on encoder outputs which is then synthesized by the decoder. Our system outperforms the current state-of-the-art causal and noncausal speech separation algorithms, reduces the computational cost of speech separation, and significantly reduces the minimum required latency of the output. This makes TasNet suitable for applications where low-power, real-time implementation is desirable such as in hearable and telecommunication devices.

Despite the overwhelming success of deep learning in various speech processing tasks, the problem of separating simultaneous speakers in a mixture remains challenging. Two major difficulties in such systems are the arbitrary source permutation and unknown number of sources in the mixture. We propose a novel deep learning framework for single channel speech separation by creating attractor points in high dimensional embedding space of the acoustic signals which pull together the time-frequency bins corresponding to each source. Attractor points in this study are created by finding the centroids of the sources in the embedding space, which are subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings. The proposed model is different from prior works in that it implements an end-to-end training, and it does not depend on the number of sources in the mixture. Two strategies are explored in the test time, K-means and fixed attractor points, where the latter requires no post-processing and can be implemented in real-time. We evaluated our system on Wall Street Journal dataset and show 5.49% improvement over the previous state-of-the-art methods.