Contact

ONE VOICE. MANY DIMENSIONS.

HUMAN SIGNALS
BRIGHTER FUTURES
SPEECH · MOVEMENT · HEALTH

TEACHING & SUPERVISION

Educating the next generation
of engineers and researchers.

I lecture in mechatronics at Deakin’s School of Engineering and supervise research in clinical speech measurement, information-theoretic machine learning and wearable sensing.

Learn about teaching

FROM THE RESEARCH

Recent work & updates

All publications ↗
ACCEPTED · INTERSPEECH

What information do speech features capture?

Quantifying dimensional independence through an information-theoretic framework.

↗
PREPRINT

Exploring the geometry of speech representations

Topological disentanglement and confound detection in speech representation spaces.

↗
CONFERENCE PAPER · IEEE EMBC

Quantum approaches for dysphonia assessment

Investigating assessment methods for small speech datasets.

↗
Deakin University  |  Geelong, AustraliaLast updated 20 September 2026
BASED AT Deakin UniversitySchool of EngineeringNetworked Sensing &
Biomedical Engineering Lab
Geelong, Australia

01 / RESEARCH

A voice is more
than a signal.

Speech carries language, emotion and information about the body—all at once. My work asks how we can separate them, and how confidently we can measure each.

I’m a Lecturer in Mechatronics and Deputy Director of the Networked Sensing and Biomedical Engineering Lab at Deakin University. My research connects information-theoretic machine learning with clinical speech measurement, building on my PhD work in cerebellar ataxia.

01

Understanding what speech encodes

Separating physiological, emotional and linguistic information in speech representations, with bounds on what each measurement can tell us.

02

Bringing measurement into the clinic

Connecting acoustic analysis and biomedical sensing with the practical demands of reliable, repeatable assessment.

03

Studying neurological speech changes

Investigating speech and motor function in cerebellar ataxia, and the relationships between affected domains.

RESEARCH IN PRACTICE

From a question
to a method.

Three ways to follow the work: the question, the approach, and the evidence behind it.

01 / REPRESENTATION LEARNING

What is a speech feature actually measuring?

Question
Can we distinguish physiological information from emotion and language?
Approach
Estimate mutual information with lower and upper bounds, then compare with non-parametric tests.
Read the evidence
The 2026 manuscript studies handcrafted features across six corpora.
Explore the manuscript →
02 / ACOUSTIC MEASUREMENT

Make speech timing interpretable.

Question
How can a waveform reveal changes in syllabic timing?
Approach
Extract an amplitude envelope and detect prominent peaks to describe temporal patterns.
Explore the method
Try the speech-timing demonstration, then follow the published study for its clinical context.
Try speech timing →
03 / BIOMEDICAL SENSING

Look beyond a single measurement.

Question
What can different motor domains contribute to assessment?
Approach
Combine instrumented tasks across speech, limbs, gait and balance, then analyse the contributing features.
Read the evidence
The open-access study describes the measurement framework and its evaluation.
Read the study →

TOOLS & METHODS

The methods behind
the questions.

Speech analysis
Acoustic features · Cepstral analysis · Speech envelopes · Syllabic timing
Representation learning
Mutual information · Statistical dependence · Clustering · Confound analysis
Sensing & evaluation
Inertial sensing · Instrumented motor tasks · Feature selection · Clinical-score comparison

02 / PUBLICATIONS & MANUSCRIPTS

Read the work.
Explore the evidence.

Full list on Google Scholar ↗

Selected work in clinical speech, representation learning and quantitative assessment.

Published & accepted work 04 selected papers

2026

Accepted · Interspeech 2026

Quantifying Dimensional Independence in Speech: An Information-Theoretic Framework for Disentangled Representation Learning

B. Kashyap, B. W. Schuller & P. N. Pathirana

How independent are the different kinds of information in a voice? This work combines bounded mutual-information estimates with non-parametric validation across six speech corpora.

Research context

The analysis concerns handcrafted acoustic features and the datasets studied; it does not establish universal independence across all speech models.

2020

Journal article · Annals of Biomedical Engineering

Quantitative Assessment of Speech in Cerebellar Ataxia Using Magnitude and Phase Based Cepstrum

B. Kashyap, P. N. Pathirana, M. Horne, L. Power & D. Szmulewicz

An acoustic approach to characterising speech in cerebellar ataxia, examining information in both the magnitude and phase of the speech signal.

Research context

The study investigates quantitative speech assessment in a clinical research setting. Its results should be interpreted within the study population and recording conditions.

2020

Journal article · Biomedical Signal Processing and Control

Automated Topographic Prominence based quantitative assessment of speech timing in Cerebellar Ataxia

B. Kashyap, M. Horne, P. N. Pathirana, L. Power & D. Szmulewicz

Uses the prominence of peaks in a speech envelope to investigate syllabic timing. The method connects interpretable signal features with quantitative assessment of ataxic speech.

Research context

The browser instrument below illustrates a related signal-processing idea. It is a demonstration, not a reproduction of the clinical validation or a diagnostic tool.

2020

Journal article · Scientific Reports · Open access

Objective Assessment of Cerebellar Ataxia: A Comprehensive and Refined Approach

B. Kashyap, D. Phan, P. N. Pathirana, M. Horne, L. Power & D. Szmulewicz

Brings speech, limb movement, gait and balance into a multi-domain assessment framework using instrumented neurological tasks and statistical analysis.

Research context

The study combines sensor-derived features and clinical scores to examine which measurements contribute to assessment. The published methods describe the cohort and validation.

Additional publications 14 works · Deakin & Google Scholar

2025

Journal Article · International Journal of Machine Learning and Cybernetics 16 (9), 6065-6080

Image encoded time series classification of small datasets: an innovative architecture using deep learning ensembles

PL Indrasiri, B Kashyap, PN Pathirana

2023

Conference Paper · IEEE EMBC

A Study of Upper-Limb Motion using Kinematic Measures for Clinical Assessment of Cerebellar Ataxia

LL Abeysekara, B Kashyap, C Kolambahewage, PN Pathirana, M Horne, et al.

2023

Conference Paper · IEEE EMBC

An Enhanced Classification Framework for Limited IoHT Time Series Data Using Ensemble Deep Learning and Image Encoding

PL Indrasiri, B Kashyap, PN Pathirana

2023

Journal Article · IEEE Internet of Things Journal 11 (7), 12115-12131

FedXPro: Bayesian Inference for Mitigating Poisoning Attacks in IoT Federated Learning

PL Indrasiri, DC Nguyen, B Kashyap, PN Pathirana, YC Eldar

2023

Journal Article · IEEE Transactions on Neural Systems and Rehabilitation Engineering 31, 4839-4850

Machine Learning Based Scoring System to Predict the Risk and Severity of Ataxic Speech Using Different Speech Tasks

B Kashyap, PN Pathirana, M Horne, L Power, DJ Szmulewicz

2021

Journal Article · IEEE Access 9, 135343-135353

Modeling the Progression of Speech Deficits in Cerebellar Ataxia Using a Mixture Mixed-Effect Machine Learning Framework

B Kashyap, PN Pathirana, M Horne, L Power, DJ Szmulewicz

2020

Conference Paper · IEEE EMBC

A sensor-based comprehensive objective assessment of motor symptoms in Cerebellar Ataxia

B Kashyap, D Phan, PN Pathirana, M Horne, L Power, D Szmulewicz

2019

Conference Paper · IEEE EMBC

Automated tongue-twister phrase-based screening for Cerebellar Ataxia using Vocal tract Biomarkers∗

B Kashyap, PN Pathirana, M Horne, L Power, D Szmulewicz

2017

Conference Paper · IEEE BMEiCON

A constrained nonlinear optimization solution for 3D orientation estimation of the human limb

D Phan, B Kashyap, PN Pathirana, A Seneviratne

Preprints Work in progress

2026

Preprint · arXiv

Geometric Analysis of Speech Representation Spaces: Topological Disentanglement and Confound Detection

B. Kashyap & P. N. Pathirana

Examines how emotional, linguistic and pathological speech features cluster, and where they overlap. The question is whether apparent clinical differences can be confused with language or accent.

Research context

The manuscript evaluates geometric separation across six corpora and eight dataset combinations. Its arXiv record lists submission to Interspeech 2026, without a confirmed acceptance.

2026

Preprint · arXiv preprint arXiv:2601.02432

Quantifying Quanvolutional Neural Networks Robustness for Speech in Healthcare Applications

H Tran, B Kashyap, PN Pathirana

2025

Preprint · arXiv preprint arXiv:2503.22971

Enhancing federated learning through secure cluster-weighted client aggregation

K Ranaweera, AG Neiat, X Liu, B Kashyap, PN Pathirana

2025

Preprint · arXiv preprint arXiv:2510.05826

Leveraging Vision Transformers for Enhanced Classification of Emotions using ECG Signals

PL Indrasiri, B Kashyap, PN Pathirana

2024

Preprint · arXiv preprint arXiv:2412.02283

VR based emotion recognition using deep multimodal fusion with biosignals across multiple anatomical domains

PL Indrasiri, B Kashyap, C Kolambahewage, B Nakisa, K Ijaz, et al.

Thesis & other writing 2 works

2021

Thesis · Deakin University

Quantitative Assessment of Speech-Motor Deficits and Domain-Wise Manifestation in Cerebellar Ataxia

B Kashyap

03 / TEACHING & SUPERVISION

From understanding
to discovery.

Engineering begins with a question.
Research teaches us how to ask it well.

Alongside my research, I lecture in mechatronics at Deakin’s School of Engineering and supervise research in clinical speech measurement, information-theoretic machine learning and wearable sensing.

For prospective research students, these areas offer a meeting point between engineering methods and questions about human health. Get in touch with your background, the problem you want to investigate and what draws you to it.

Learning through measurementExplore the five browser-based instruments below to connect signals, observations and their limitations.
Discuss a research idea ↗

04 / THE OPEN INSTRUMENT BENCH

Make the signal visible.

Scroll through five working instruments. Try each one here using your own voice, movement or touch.

ILLUSTRATIVE SIGNAL / SYNTHETIC DATA00:00
FREQUENCY ↑   /   TIME →

One signal. Three layers.

3 layers visible

Toggle a layer to see its contribution in this illustration. These are illustrative channels, not a model of a person or a clinical result.

01 / TOUCH · 15 SECONDS

Finger tapping

Explore timing and rhythm with a simple tapping task. Tap the pad to begin.

No permissions needed. Use a finger, mouse, or the space bar.

Open separately ↗

02 / MICROPHONE · 5 SECONDS

Speech timing

See the rhythm of repeated syllables and explore how the detection threshold changes the result.

Microphone access is requested when you press Record.

Open separately ↗

03 / MICROPHONE · UP TO 30 SECONDS

Sustained vowel

Hold a comfortable “ah” and explore the resulting waveform and pitch trace.

Use a quiet space. Microphone access starts only when requested.

Open separately ↗

04 / PHONE MOTION · 10 SECONDS

Tremor spectrum

Observe the frequencies present in small movements of your phone.

Use a phone with a motion sensor. Keep the screen on during recording.

Open separately ↗

05 / CAMERA · 30 SECONDS

Fingertip pulse

Explore the small changes in light captured through your fingertip.

Best on a phone with a rear camera. Camera permission is requested on Start.

Open separately ↗

Research demonstrations, not medical devices. Sensor-based instruments need a compatible device and your permission. The illustrations on this page use synthetic data.

05 / LET’S CONNECT

Good science starts
with a conversation.

A research question. A collaboration.
A new direction to explore.

For enquiries about clinical speech, representation learning, biomedical sensing or research supervision, write to me.