Portrait of Shree Harsha.

PhD student · Speech technology · Conversational AI

Shree Harsha

I research speech interaction dynamics, SpeechLLM evaluation, and conversational AI.

My current work studies enchronic speech dynamics with self-supervised learning, voice-based bias evaluation, and failure modes in speech foundation models. I am a WASP PhD Student at the KTH Royal Institute of Technology, Stockholm, Sweden. I am supervised by Éva Székely and Gustav Eje Henter.

Publications Google Scholar GitHub

Research

Current Directions

Current focus JSALT 2026

Evaluating full-duplex conversation as interaction

I am developing evaluation methods for full-duplex conversational systems that use human-human conversation as the reference. We plan to use distributional metrics to measure human-human, human-AI, and AI-AI conversations.

It connects turn-taking, responsiveness, entrainment, affect, and distributional speech metrics to participant-level judgements of conversational success.

Final JSALT workshop presentation Watch on YouTube ↗

01

Enchronic speech dynamics

Modelling the moment-by-moment timing, turn-taking, and coordination through which people build meaning together.

02

Responsible SpeechLLM evaluation

Revealing voice-based bias and robustness failures through controlled variation, open-ended interaction, and human validation.

Selected work

Research in view

Selected papers on voice-based bias, interactive evaluation, and robustness in speech foundation models.

Heatmap and interval plot comparing SpeechLLM helpfulness across accents, models, and gender.
Helpfulness across accent × gender profiles

Interspeech 2026 · Accepted

The Voice Behind the Words

Quantifying intersectional bias in SpeechLLMs.

Study design comparing observational, direct interactive, and voice-converted vicarious evaluation.
Three positions for evaluating an AI interaction

CHI Extended Abstracts 2026

From Seeing it to Experiencing it

Interactive evaluation of intersectional voice bias.

t-SNE scatter plot and bar chart showing embedding shifts after voice quality conversion.
Representation shifts after voice conversion

Interspeech 2026 · Submitted

What Counts as Real?

New challenges for deepfake detection under speech restoration and voice-quality conversion.

Pipeline from speech prompts through model continuation and multi-part bias assessment.
Speech continuation evaluation framework

Identity-Aware AI · LREC 2026

Speak Your Mind

The speech continuation task as a probe of voice-based model bias.

Two SpeechLLM examples showing that unbiased multiple-choice behaviour does not consistently transfer to long-form advice.
Benchmark behaviour versus long-form generation

IEEE ICASSP 2026 · Accepted

Do Bias Benchmarks Generalise?

Evidence from voice-based evaluation of gender bias in SpeechLLMs, testing whether benchmark behaviour transfers to open-ended tasks.

Background

Education and Experience

  1. PhD student

    Researching enchronic speech dynamics using self-supervised learning and bias discovery/mitigation in conversational AI. Supervised by Éva Székely and Gustav Eje Henter at KTH Royal Institute of Technology, with support from WASP.

  2. Speech and AI systems

    Worked on speech assessment systems, voice activity detection, and industry experience with LLM/RAG solutions for aerospace repair-scheme creation at ALTEN Ltd. UK.

  3. MSc in Data Science & AI at the University of Liverpool

    Deepened my skills across machine learning, AI, and applied data-driven systems.

  4. Master's in Electrical Engineering, (Signal Processing) at IIT Bombay

    Built ASR models for literacy assessments in India, later used in the field by an NGO. Advised by Preeti Rao

Interests

Research Interests

  • Signal Processing
  • Speech Technology
  • Conversational AI
  • Machine Learning
  • Self-Supervised Learning
  • Responsible AI
  • SpeechLLMs
  • Voice Conversion
  • Bias Evaluation
  • Bias and Fairness
  • Deepfake Robustness

Contact

Contact