Edward F. Chang

University of California, San Francisco
Country flagUnited States

Research Interests

ApplyKite TwinLearns as you apply

Make your next move.

Prepare your application

Review drafts before you send them.

Explore related searches

Articles (21)

Reconstructing speech from neural recordings is crucial for understanding speech coding and developing brain-computer interfaces (BCIs). However, existing methods trade off acoustic richness (pitch, prosody) for linguistic intelligibility (words, phonemes). To overcome this limitation, we propose a dual-path framework to concurrently decode acoustic and linguistic representations. The acoustic pathway uses a long-short term memory (LSTM) decoder and a high-fidelity generative adversarial network (HiFi-GAN) to reconstruct spectrotemporal features. The linguistic pathway employs a transformer adaptor and text-to-speech (TTS) generator for word tokens. These two pathways merge via voice cloning to combine both acoustic and linguistic validity. Using only 20 minutes of electrocorticography (ECoG) data per subject, our approach achieves highly intelligible synthesized speech (mean opinion score = 4.0/5.0, word error rate = 18.9%). Our dual-path framework reconstructs natural and intelligible speech from ECoG, resolving the acoustic-linguistic trade-off.

Year:

2026

Reconstructing speech from neural recordings is crucial for understanding speech coding and developing brain-computer interfaces (BCIs). However, existing methods trade off acoustic richness (pitch, prosody) for linguistic intelligibility (words, phonemes). To overcome this limitation, we propose a dual-path framework to concurrently decode acoustic and linguistic representations. The acoustic pathway uses a long-short term memory (LSTM) decoder and a high-fidelity generative adversarial network (HiFi-GAN) to reconstruct spectrotemporal features. The linguistic pathway employs a transformer adaptor and text-to-speech (TTS) generator for word tokens. These two pathways merge via voice cloning to combine both acoustic and linguistic validity. Using only 20 minutes of electrocorticography (ECoG) data per subject, our approach achieves highly intelligible synthesized speech (mean opinion score = 4.0/5.0, word error rate = 18.9%). Our dual-path framework reconstructs natural and intelligible speech from ECoG, resolving the acoustic-linguistic trade-off.

Year:

2026

Reconstructing speech from neural recordings is crucial for understanding speech coding and developing brain-computer interfaces (BCIs). However, existing methods trade off acoustic richness (pitch, prosody) for linguistic intelligibility (words, phonemes). To overcome this limitation, we propose a dual-path framework to concurrently decode acoustic and linguistic representations. The acoustic pathway uses a long-short term memory (LSTM) decoder and a high-fidelity generative adversarial network (HiFi-GAN) to reconstruct spectrotemporal features. The linguistic pathway employs a transformer adaptor and text-to-speech (TTS) generator for word tokens. These two pathways merge via voice cloning to combine both acoustic and linguistic validity. Using only 20 minutes of electrocorticography (ECoG) data per subject, our approach achieves highly intelligible synthesized speech (mean opinion score = 4.0/5.0, word error rate = 18.9%). Our dual-path framework reconstructs natural and intelligible speech from ECoG, resolving the acoustic-linguistic trade-off.

Year:

2025

Fluent speech depends on precisely timed motor commands that coordinate rapid transitions between successive articulatory gestures. Using direct cortical recordings, we identified a prominent sensorimotor theta oscillation (6–10 Hz) that supports this coordination. During articulation, premotor speech circuits exhibited enhanced theta phase coherence, with elevated population activity near theta troughs. The oscillation’s frequency remained remarkably stable across varying speech rates and cognitive states, consistent with an intrinsically generated rhythm. Vocal-tract kinematics revealed pulse-like movements at 6–10 Hz, tightly coupled to cortical theta phase. At a mesoscopic scale, theta cycles structured sequential sensorimotor activations encoding articulatory gestures, with syllable identity optimally decodable following theta troughs. These findings identify theta oscillations as an intrinsic timing mechanism that coordinates the distributed and synergistic motor control underlying fluent speech.

Year:

2025

Reconstructing speech from neural recordings is crucial for understanding speech coding and developing brain-computer interfaces (BCIs). However, existing methods trade off acoustic richness (pitch, prosody) for linguistic intelligibility (words, phonemes). To overcome this limitation, we propose a dual-path framework to concurrently decode acoustic and linguistic representations. The acoustic pathway uses a long-short term memory (LSTM) decoder and a high-fidelity generative adversarial network (HiFi-GAN) to reconstruct spectrotemporal features. The linguistic pathway employs a transformer adaptor and text-to-speech (TTS) generator for word tokens. These two pathways merge via voice cloning to combine both acoustic and linguistic validity. Using only 20 minutes of electrocorticography (ECoG) data per subject, our approach achieves highly intelligible synthesized speech (mean opinion score = 4.0/5.0, word error rate = 18.9%). Our dual-path framework reconstructs natural and intelligible speech from ECoG, resolving the acoustic-linguistic trade-off.

Year:

2025

Objective. Neuroprostheses can restore communicative ability to people with paralysis by decoding intended speech motor movements from the sensorimotor cortex (SMC). However, overlapping neural populations in the SMC are also engaged in visual and auditory perceptual processing. The nature of these shared motor and perceptual activations and their potential to interfere with decoding are particularly relevant questions for speech neuroprostheses, as reading and listening are essential daily functions. Approach. In two participants with vocal-tract paralysis and anarthria (ClinicalTrials.gov; NCT03698149), we developed an online electrocorticography (ECoG) based speech-decoding system that maintained accuracy and specificity to intended speech, even during common daily tasks like reading and listening. Offline, we studied the spectrotemporal characteristics and spatial distribution of reading, listening, and attempted-speech responses across our participants’ ECoG arrays. Main results. Across participants, the speech-decoding system had zero false-positive activations during 63.2 min of attempted speech and perceptual tasks, maintaining accuracy and specificity to volitional speech attempts. Offline, though we observed shared neural populations that responded to attempted speech, listening, and reading, we found they leveraged different neural representations with differentiable spectrotemporal responses. Shared populations localized to the middle precentral gyrus and may have a distinct role in speech-motor planning. Significance. Potential neuroprosthesis users strongly desire reliable systems that will retain specificity to volitional speech attempts during daily use. These results demonstrate a decoding framework for speech neuroprostheses that maintains this specificity and further our understanding of shared perceptual and motor activity on the SMC.

Year:

2025

Direct cortical stimulation (DCS) is the clinical gold standard for identifying functional cortex in the human brain, which is essential for the safe removal of brain lesions. Defining the electro-physiological properties of DCS-positive cortical regions may facilitate the identification of critical language regions, thereby permitting safe glioma resections in communities without access. Leveraging a multicenter electrophysiologic dataset of DCS-positive language regions spatially matched with subdural arrays, we analyzed regions identified as functionally critical (DCS+) versus functionally non-critical (DCS-) during intraoperative language mapping. In IDH-mutant gliomas, DCS+ regions exhibited significantly greater speech-related neural activity and enhanced encoding and decoding of linguistic and semantic features. We demonstrate that resting-state classifiers distinguish DCS+ from DCS- regions in IDH-mutant tumors. Task-based and resting-state electrophysiologic distinctions were pathology-specific and not present in IDH-wildtype glioblastomas. These findings may accelerate DCS mapping by guiding surgeons to priority regions, improving efficiency, and patient outcomes.

Year:

2025

When we hear continuous speech, we perceive it as a series of discrete words, despite the lack of clear boundaries in the acoustic signal. The superior temporal gyrus (STG) encodes phonetic elements like consonants and vowels, but how it extracts whole words as perceptual units remains unclear. Using high-density cortical recordings, we investigated how the brain represents auditory word forms—integrating acoustic-phonetic, prosodic, and lexical features—while participants listened to spoken narratives. Our results show that STG neural populations exhibit a distinctive reset in activity at word boundaries, marked by a brief, sharp drop in cortical activity. Between these resets, the STG consistently encodes distinct acoustic-phonetic, prosodic, and lexical information, supporting the integration of phonological features into coherent word forms. Notably, this process tracks the relative elapsed time within each word, independent of its absolute duration, providing a flexible temporal scaffolding for encoding variable word lengths. We observed similar word form dynamics in the deeper layers of a self-supervised artificial speech network, suggesting a potential convergence with computational models. Additionally, in a bistable word perception task, STG responses were aligned with participants’ perceived word boundaries on a trial-by-trial basis, further emphasizing the role of dynamic encoding in word recognition. Together, these findings support a new dynamical model of auditory word forms, highlighting their importance as perceptual units for accessing linguistic meaning.

Year:

2025

Neuroprostheses can restore communicative ability to people with paralysis by decoding intended speech movements from the sensorimotor cortex (SMC). However, overlapping neural populations in the SMC are also engaged in visual and auditory processing. The nature of these shared motor and perceptual activations and their potential to interfere with decoding are particularly relevant questions for speech neuroprostheses, as reading and listening are essential daily functions. In two participants with vocal-tract paralysis and anarthria (ClinicalTrials.gov; NCT03698149 ), we developed an online electrocorticography (ECoG) based speech-decoding system that maintained accuracy and specificity to intended speech, even during common daily tasks like reading and listening. Offline, we studied the spectrotemporal characteristics and spatial distribution of reading, listening, and attempted-speech responses across our participants’ ECoG arrays. Across participants, the speech-decoding system had zero false-positive activations during 63.2 minutes of attempted speech and perceptual tasks, maintaining accuracy and specificity to volitional speech attempts. Offline, though we observed shared neural populations that responded to attempted speech, listening, and reading, we found they leveraged different neural representations with differentiable spectrotemporal responses. Shared populations localized to the middle precentral gyrus and may have a distinct role in speech-motor planning. Potential neuroprosthesis users strongly desire reliable systems that will retain specificity to volitional speech attempts during daily use. These results demonstrate a decoding framework for speech neuroprostheses that maintains this specificity and further our understanding of shared perceptual and motor activity on the SMC.

Year:

2025

Introduction Treatment-refractory obsessive-compulsive disorder (trOCD) is a complex network disorder that may require personalized treatment strategies due to disease heterogeneity. A multi-site, multi-stage, double-blinded, randomized crossover clinical trial is underway, using stereo electroencephalography (sEEG) to guide selection of multi-nodal targets for deep brain stimulation (DBS) for trOCD. Objectives To describe the clinical trial design, emphasizing personalized surgical targeting strategies that ensure the feasibility and precision of sEEG electrode placement, and enable adequate sampling of relevant targets in trOCD for network evaluation and modulation. Methods Adults with severe trOCD (Yale-Brown Obsessive Compulsive Scale ≥ 28) who meet eligibility criteria are enrolled in this three-stage clinical trial ( NCT05623306 ). Stage 1 involves SEEG electrode implantation in trOCD implicated regions and inpatient evaluation. Individualized probabilistic-tractography-guided target refinement is performed for surgical planning. Multimodal recordings are taken while participants stay in the psychiatric monitoring unit for 12 days. In stage 2, up to four permanent DBS electrodes are implanted followed by stimulation optimization. Stage 3 is the randomized, double blinded cross over phase. Expected Outcomes Safety, feasibility and preliminary efficacy will be assessed in this ongoing study. We anticipate that the use of sEEG to guide selection of multi-nodal targets for DBS will be safe, feasible and result in clinically meaningful improvements in symptom severity and functional impairment in trOCD. Discussion We present the clinical protocol of sEEG-guided investigation of brain networks involved in trOCD and describe our tractography-guided surgical targeting strategy designed to optimize individualized network engagement and neuromodulation.

Year:

2025

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Year:

2024

Collaborators (20)

Evan R. Williams

University of California

UNITED STATES
View details

Liming Qiu

University of Pennsylvania

UNITED STATES
View details

Tal Sharf

Assistant Professor

University of California, Santa Cruz

UNITED STATES
View details

David Schaffer

Professor/Director

University of California, Berkeley

UNITED STATES
View details

James McParland

University College London Hospitals NHS Foundation Trust

UNITED KINGDOM
View details

Maria Luisa Mandelli

Associate Professor

University of California, San Francisco

UNITED STATES
View details

Shawn L. Hervey‐Jumper

Professor

University of California, San Francisco

UNITED STATES
View details

Mircea Teodorescu

Associate Professor

University of California, Santa Cruz

UNITED STATES
View details

Loren Frank

Howard Hughes Medical Institute

UNITED STATES
View details

Karunesh Ganguly

Professor

University of California, San Francisco

UNITED STATES
View details

Matthew Leonard

Associate Professor

University of California, San Francisco

UNITED STATES
View details

Manish K Aghi

University of California, San Francisco

UNITED STATES
View details

Sofie Salama

Professor

University of California, Santa Cruz

UNITED STATES
View details

Christos Ganos

Director

-

GERMANY
View details

John Andrews

University of California, San Francisco

UNITED STATES
View details

Robert Currie

University of California, Santa Cruz

UNITED STATES
View details

Tomasz J. Nowakowski

University of California

UNITED STATES
View details

Desmond J. Oathes

Associate Professor

University of Pennsylvania

UNITED STATES
View details

Nolan Williams

Associate Professor

Stanford University

UNITED STATES
View details

Jennifer Listgarten

University of California

UNITED STATES
View details