PhD student · Speech technology · Conversational AI
I research speech interaction dynamics, SpeechLLM evaluation, and responsible 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 supervised by Éva Székely and Gustav Eje Henter, supported by the WASP program, and based at KTH Royal Institute of Technology.
About
Enchronic speech dynamics are the moment-by-moment patterns that make conversation work: when people take turns, how quickly they respond, how meaning is built together, and how local social context shapes what comes next.
My research uses self-supervised learning to study these dynamics and to ask how SpeechLLMs and conversational AI systems can better reflect the timing, variation, and responsibility that real interaction requires.
Research
Problem. Conversation unfolds in real time, but many speech systems treat utterances as isolated clips.
Why it matters. Timing, turn-taking, and shared meaning are central to natural interaction and to conversational AI that behaves responsibly.
Methods. Speech timing analysis, interaction-focused modeling, and self-supervised representations of spoken audio.
Problem. Multiple-choice bias benchmarks can miss how SpeechLLMs behave in open-ended, real-use settings.
Why it matters. Context-aware measurements are needed when voice, accent, gender, and interaction style shape model behavior.
Methods. Controlled acoustic variation, voice conversion, long-form evaluation, LLM-as-a-judge methods, and human preference validation.
Problem. Systems can appear strong on proxy tasks while failing under transformations, restoration, conversion, or out-of-distribution speech.
Why it matters. Robust speech AI needs evaluations that reveal what models encode, what users perceive, and what downstream classifiers actually measure.
Methods. Speech continuation probes, voice-conversion studies, deepfake-detection stress tests, and representation analysis.
Publications
2026
SH Bokkahalli Satish, C Minixhofer, M Teleki, J Caverlee, O Klejch, P Bell, GE Henter, É Székely.
2026
SH Bokkahalli Satish, M Teleki, C Minixhofer, O Klejch, P Bell, É Székely.
2026
SH Bokkahalli Satish, M Teleki, C Minixhofer, O Klejch, P Bell, É Székely.
2026
SH Bokkahalli Satish, H Lameris, J Gustafson, É Székely.
2026
SH Bokkahalli Satish, GE Henter, É Székely.
2026
SH Bokkahalli Satish, H Lameris, O Perrotin, GE Henter, É Székely.
2025
H Lameris, SH Bokkahalli Satish, J Gustafson, É Székely.
2025
SH Bokkahalli Satish, GE Henter, É Székely.
2025
SH Bokkahalli Satish, GE Henter, É Székely.
2021
SH Bokkahalli Satish, C Vitthal, K Sabu, P Rao.
Background
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.
Worked on speech assessment systems, voice activity detection, and LLM/RAG solutions for aerospace repair-scheme creation.
Deepened work across machine learning, AI, and applied data-driven systems.
Built ASR models for literacy assessments in India, later used in the field by an NGO.
Interests
Contact
I am happy to hear from people interested in speech technology, conversational AI, representation learning, or responsible evaluation.