Saurav Keshari Aryal

Research profile

Research summary

Saurav Keshari Aryal is a researcher at Howard University (US), name-resolved from public OpenAlex bibliographic data (ORCID 0000-0001-9815-9295). The record shows an early-career trajectory with a steep productivity ramp: 22 works in the last three years and 16 indexed works in 2026 alone, against 1 in 2020 and 6 in 2023, while citation standing remains thin (h-index 4, i10-index 2). Research splits evenly across two fields — affective computing / wearable physiological sensing and speech processing / machine learning (0.5/0.5) — with OpenAlex topics led by topic modeling (14), sentiment analysis and opinion mining (8), hate speech and cyberbullying detection (6), NLP techniques (5), and emotion and mood recognition (5). Two 2026 arXiv papers reviewed by Evolve illustrate both lines: a BiMamba2 masked discrete-unit prediction entry for the Unsupervised Speech in the Wild Challenge that tops speaker clustering (ARI 0.735, above all 4 baselines) while language identification lags badly (F1 0.073), and a comparative LSTM/TCN/Transformer study of physiological emotion recognition on WESAD and EmoWear. Evolve rates both papers medium for reproducibility readiness (mean 55 across 2 papers). Collaboration is compact — 4 distinct co-authors, with Hagos, Burge, Subedi, and Prioleau recurring — and authorship roles on the reviewed papers are non-leading (middle and last, neither corresponding). Overall: a fast-moving applied ML/NLP contributor whose citation impact has not yet caught up with output.

Research keywords

Highlights

• BiMamba2 challenge entry leads speaker clustering

  Last-authored 2026 arXiv entry for the Unsupervised Speech in the Wild Challenge applies BiMamba2 masked discrete-unit prediction with k-means pseudo-labels, topping speaker clustering at ARI 0.735 against all 4 baselines — while language identification lags badly at F1 0.073. Evolve calls the report honest but ablation-free.

• Comparative deep learning for physiological emotion recognition

  Middle-authored 2026 arXiv study comparing LSTM, TCN, and Transformer architectures for emotion recognition from wearable physiological signals on WESAD and EmoWear, with thorough ablations across sensor modalities — though Evolve flags unexplained wrist-only table contradictions and no significance tests.

• Sharp output ramp into 2026

  Indexed output climbs from 1 work (2020) to 16 works in 2026, with 22 works in the last three years — the profile's highest radar score (activity 91) — spanning both the affective-computing and speech-processing halves of the record.

Timeline

1. 2020First works appear in the OpenAlex record (1 indexed work).
2. 2022Output grows to 4 indexed works.
3. 20236 indexed works in the OpenAlex record.
4. 20256 indexed works, holding the 2023 pace.
5. 2026Most productive year: 16 indexed works, including two arXiv preprints reviewed by Evolve — the BiMamba2 speech challenge entry and the wearable emotion-recognition study.

Papers in the Evolve catalog

• 2026-09-23

  BiMamba2 Masked Discrete-Unit Prediction for Multilingual Speech Representation for Unsupervised Speech in the Wild Challenge

• 2026-09-17

  From Stress to Affect: Multimodal Deep Learning for Physiological Emotion Recognition Across Wearable Sensor Modalities

Frequent co-authors