Ifeanyi Desmond Dike

Port Harcourt, Nigeria

I’m an AI engineer and applied mathematician interested in how machine learning systems learn reliable structure from complex, temporally-ordered data — whether that data is spoken dialogue, video, or financial time series. My primary research focus is conversational AI — specifically the problems of turn-taking, multi-speaker dynamics, and real-time speech understanding that arise when voice AI systems meet the messiness of actual human speech.

This interest is grounded in production experience. As founding AI engineer at Datastrut AI, I’ve built voice AI agents and multi-speaker collaboration systems for the construction industry. At Cenphi, I built multimodal pipelines that jointly reason over speech, video, and language. These systems work, but they expose hard research gaps — particularly around evaluation methodology for interactive voice agents and robustness to domain-specific disfluency — that engineering alone cannot close.

My current research proposes ATR (Acoustic Turn-over Rate) (paper, code), a VAD-based replacement for the ASR-dependent Turn-Over Rate metric used in Full-Duplex-Bench (Lin et al., 2025). Two standard ASR backends diverge by up to 23.5 percentage points on the same audio when computing TOR — ATR replaces the transcription step with voice activity detection and reduces backend disagreement to 7.6 pp across three independent VAD systems. Earlier work includes two preprints reproducing the Full-Duplex-Bench pipeline for Gemini: the v1.0 reproduction quantifies ASR backend sensitivity in TOR computation, and the v1.5 reproduction evaluates overlap handling and prosodic adaptation across four scenarios.

I hold a First Class Honours degree in Applied Mathematics (GPA 4.63/5.0) from the Federal University of Technology Akure and co-authored four IEEE-indexed publications in 2018. I’m currently pursuing an MSc in Financial Engineering at WorldQuant University and actively seeking a PhD position in ML.

I’m particularly interested in connecting with faculty working on speech-language models, full-duplex dialogue systems, multimodal AI, and related areas involving temporal structure in complex data. If that intersects your work and you are considering students, feel free to reach out.

news

May 23, 2026	Published second preprint: Evaluating Overlap Handling in Gemini 3.1 Flash Live — Full-Duplex-Bench v1.5. Gemini correctly handles user interruptions 72% of the time and shows systematic prosodic adaptation after overlap events.
May 20, 2026	Published preprint: Reproducing Full-Duplex-Bench — reproduction of Lin et al. (2025) on Apple Silicon + ASR backend sensitivity analysis. Code at GitHub.
May 17, 2026	Started reproduction and extension of Full-Duplex-Bench (Lin et al., 2025) — evaluating turn-taking and overlap handling in voice AI systems. See the project page for details.

selected publications

Preprint
ATR: An Acoustic Turn-over Rate Metric for Evaluating Overlap Handling in Full-Duplex Spoken Dialogue Systems

Ifeanyi Desmond Dike

Jun 2026

Abs DOI Bib Code

Turn-over Rate (TOR), introduced in Full-Duplex-Bench, measures how often a full-duplex spoken dialogue model yields the speaking floor during overlapping speech. TOR is computed from ASR transcripts, and this dependency is important: two commonly-used ASR backends produce TOR estimates that diverge by up to 23.5 percentage points on the same audio, making cross-study comparisons unreliable. We propose the Acoustic Turn-over Rate (ATR), which replaces the ASR step with voice activity detection (VAD). Evaluated on Moshi across all four Full-Duplex-Bench v1.5 scenarios, three independent VAD backends agree within 7.6 percentage points on user interruption — roughly three times more consistent than ASR-based TOR on the same audio. ATR requires no language model, runs locally, and is language-agnostic.
@misc{dike2026atr, title = {ATR: An Acoustic Turn-over Rate Metric for Evaluating Overlap Handling in Full-Duplex Spoken Dialogue Systems}, author = {Dike, Ifeanyi Desmond}, year = {2026}, month = jun, publisher = {Zenodo}, doi = {10.5281/zenodo.20548903}, url = {https://doi.org/10.5281/zenodo.20548903}, }
Preprint
Evaluating Overlap Handling in Gemini 3.1 Flash Live: A Reproduction of Full-Duplex-Bench v1.5

Ifeanyi Desmond Dike

May 2026

Abs DOI Bib Code

We reproduce the Full-Duplex-Bench v1.5 evaluation pipeline for Gemini 3.1 Flash Live Preview on Apple Silicon (M3 Max, no CUDA), covering four simulated overlap scenarios: background speech, talking to another person, user backchanneling, and user interruption. Gemini responds correctly to user interruption in 72% of cases and resumes after background noise and side conversation in 50–63% of cases. Pitch and speaking rate increase after overlap events while intensity drops, suggesting prosodic adaptation without dominance assertion. The v1.5 metric suite is more reproducible than v1.0 TOR-based metrics as it does not depend on ASR word boundary detection.
@misc{dike2026fullduplex_v15, title = {Evaluating Overlap Handling in Gemini 3.1 Flash Live: A Reproduction of Full-Duplex-Bench v1.5}, author = {Dike, Ifeanyi Desmond}, year = {2026}, month = may, publisher = {Zenodo}, doi = {10.5281/zenodo.20354457}, url = {https://doi.org/10.5281/zenodo.20354457}, }
Preprint
Reproducing Full-Duplex-Bench: Evaluation of Gemini 3.1 Flash Live on Turn-Taking and an Analysis of ASR Backend Sensitivity

Ifeanyi Desmond Dike

May 2026

Abs DOI Bib Code

We reproduced the Full-Duplex-Bench v1.0 evaluation pipeline for Gemini 3.1 Flash Live Preview and conducted a systematic analysis of how ASR backend choice affects downstream turn-taking metrics. TOR and response latency vary significantly across backends while JSD, backchannel frequency, and LLM-as-judge scores remain stable. Neither backend is universally closer to the paper’s numbers: AssemblyAI matches better on pause handling TOR while MLX Whisper-v3 is closer on backchannel TOR, smooth turn-taking, and latency.
@misc{dike2026fullduplex, title = {Reproducing Full-Duplex-Bench: Evaluation of Gemini 3.1 Flash Live on Turn-Taking and an Analysis of ASR Backend Sensitivity}, author = {Dike, Ifeanyi Desmond}, year = {2026}, month = may, publisher = {Zenodo}, doi = {10.5281/zenodo.20305268}, url = {https://doi.org/10.5281/zenodo.20305268}, }
IEEE IEMCON
Fast Prediction of Equivalent Model of Installed Patch Antenna Radiation Pattern

Happy Monday, Jian Li, Mordecai Raji, and 3 more authors

In 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), 2018

Abs Bib

This paper presents a fast prediction approach for equivalent modeling of installed patch antenna radiation patterns. The method enables efficient characterization of antenna performance in complex installation environments, reducing computational overhead while maintaining prediction accuracy.
@inproceedings{monday2018fastprediction, title = {Fast Prediction of Equivalent Model of Installed Patch Antenna Radiation Pattern}, author = {Monday, Happy and Li, Jian and Raji, Mordecai and Nneji, Grace and Dike, Ifeanyi and Nneji, Richard}, booktitle = {2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON)}, pages = {286--291}, year = {2018}, publisher = {IEEE}, url = {https://ieeexplore.ieee.org/document/8614869/}, }

IEEE IEMCON

An Improved e-Clearance Management System for Graduating Students in a University Environment

Grace Nneji, Jianhua Deng, Sarder Shamir Shakher, and 3 more authors

In 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), 2018

Abs Bib

@inproceedings{nneji2018eclearance,
  title = {An Improved e-Clearance Management System for Graduating Students in a University Environment},
  author = {Nneji, Grace and Deng, Jianhua and Shakher, Sarder Shamir and Monday, Happy and Agomuo, David and Dike, Ifeanyi},
  booktitle = {2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON)},
  pages = {74--80},
  year = {2018},
  publisher = {IEEE},
  url = {https://ieeexplore.ieee.org/document/8614875/},
}

IEEE IEMCON

Enhanced Attendance Management System: A Biometrics System of Identification Based on Fingerprint

Happy Monday, Ifeanyi Dike, Jian Li, and 3 more authors

In 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), 2018

Abs Bib

@inproceedings{monday2018biometrics,
  title = {Enhanced Attendance Management System: A Biometrics System of Identification Based on Fingerprint},
  author = {Monday, Happy and Dike, Ifeanyi and Li, Jian and Agomuo, David and Nneji, Grace and Ogungbile, Abel},
  booktitle = {2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON)},
  pages = {500--505},
  year = {2018},
  publisher = {IEEE},
  url = {https://ieeexplore.ieee.org/document/8614776/},
}

IEEE IEMCON

Design of an Improved Cost Effective Electronic Locking System

Happy Monday, Jian Li, Grace Nneji, and 3 more authors

In 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), 2018

Abs Bib

@inproceedings{monday2018locking,
  title = {Design of an Improved Cost Effective Electronic Locking System},
  author = {Monday, Happy and Li, Jian and Nneji, Grace and Ukwuoma, Chiagoziem and Dike, Ifeanyi and Nneji, Richard},
  booktitle = {2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON)},
  pages = {493--499},
  year = {2018},
  publisher = {IEEE},
  url = {https://ieeexplore.ieee.org/document/8614749/},
}