Sergey D. Stavisky

This paper fine-tunes an ASR model on human-transcribed synthesized speech to better align its word error rate with human perception, achieving higher correlation than zero-shot ASR.

This paper introduces a multi-user transformer-based decoder for intracortical speech BCIs, achieving lower word error rates with less training data than individual models.

Brain2voice 2.0 improves voice synthesis from neural signals, achieving 5.24% word error rate, 8x better than prior methods, enabling real-time intelligible speech for paralyzed individuals.

A paralyzed man with ALS used a long-term intracortical BCI for over two years to communicate, control a computer, and work independently.

This paper tests deep neural ensembles in real-time speech BCIs, showing reduced word error rates and introducing a computationally efficient pseudoensembling method.

The paper shows that premotor cortex uses a compositional neural geometry to plan words by combining phoneme representations, enabling efficient planning of novel sequences.

This study decodes speech mode and loudness from ventral precentral gyrus activity in two participants, achieving high accuracy in predicting speech parameters and developing a closed-loop decoder for expressive speech neuroprostheses.

This paper demonstrates cross-brain transfer learning for intracortical BCIs decoding speech and handwriting, showing improved performance with limited training data and specific decoder design.

This paper decodes intended speech from a person with anarthria and locked-in syndrome using an intracortical brain-computer interface, showing that phonemes, words, and higher-order language units can be decoded above chance.

This paper demonstrates a brain-to-voice neuroprosthesis that synthesizes speech from neural activity, enabling real-time voice modulation with closed-loop feedback.

The paper shows error detection signals in speech motor cortex during neuroprosthesis use, enabling error identification with 86% accuracy and revealing distinct patterns for different error types.

This paper demonstrates that neural activity from ventral motor cortex can enable high-performance cursor control and click for brain-computer interfaces.

The paper studies neural encoding of inner speech in motor cortex and shows it can be decoded, but also identifies a neural dimension that can prevent unintentional decoding of private thoughts.

The paper maps human motor cortex using microelectrode recordings, finding whole-body representations and challenging the traditional motor homunculus model.

BRAND provides a platform for closed-loop experiments with deep networks, achieving low latency and supporting multiple programming languages.

This paper describes a new high-density Neuropixels probe enabling single-cell resolution recordings from over 200 cortical neurons in humans during surgery, identifying eight distinct single unit classes.

This study examines how force and grasp are represented in motor cortex of tetraplegic individuals, finding that grasp affects force decoding and that grasp is more accurately classified than force.

This study decodes 39 English phonemes from neural signals in the dorsal precentral gyrus using intracortical electrode arrays, achieving 33.9% classification accuracy with a recurrent neural network.

The paper records neural activity in the motor cortex of people with paralysis during speech and non-speaking movements, showing that neurons modulate during speech and revealing common neural dynamics between speech and arm movements.

The paper shows that neural dynamics can be accurately estimated without spike sorting by using multi-unit threshold crossings, enabling new analyses and electrode array design.

This paper introduces LFADS, a deep learning method to infer single-trial neural population dynamics from spiking data, demonstrating improved accuracy in predicting behavior and extracting dynamics across datasets.