University of Maryland researcher advancing steerable speech dialogue models and audio-language understanding.
4
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
No publication-volume data is present in this snapshot, so output cannot be scored. This axis would measure total publications, venues, and authorship roles over time — a productivity baseline for peer comparison.
Impact—
No citation or h-index data is included, so impact is not scored. This axis would measure how widely the work is cited and adopted — citation totals, per-paper influence, and downstream uptake by other researchers.
Focus0/100
The focus score is 0, computed from an even field distribution: 0.5 weight in speech processing/conversational AI and 0.5 in speech processing/audio-language models (machine learning). The two lanes are closely related, and the same 2026 cluster — duplex dialogue, audio-language post-training, paralinguistics — spans both, so the score indicates breadth across adjacent speech subfields rather than disconnected work.
Collaboration69/100
Scores 69, grounded in 31 distinct coauthors. All four cataloged 2026 papers place this researcher in a middle author position with no corresponding role, consistent with contributing inside a large group. The recurring core is Dinesh Manocha and Sreyan Ghosh, joined by a rotating set of collaborators such as Utkarsh Tyagi, Vaibhavi Lokegaonkar, Vishnu Raj, and Sonal Kumar across dialogue, audio-language, and music-generation projects.
Activity—
No year-by-year publication counts are available, so activity cannot be scored. This axis would measure publishing cadence and momentum over time; the four 2026 arXiv catalog entries hint at an active period but cannot quantify it.
Reproducibility55/100
Mean readiness is 55, computed over 2 papers with Evolve ratings — ParA-LLM and SteerDuplex, both medium. The reviews flag reproducibility-relevant limits: ParA-LLM rests on a circular self-built benchmark with weak statistics, and SteerDuplex has thin baselines and single-judge scoring. A 55 therefore describes work that introduces new benchmarks and reports gains honestly, but still needs independent evaluation for full verification.
Research summary
Ramani Duraiswami is a researcher at the University of Maryland, College Park whose cataloged work sits in speech processing, split evenly (0.5/0.5 in the field distribution) between conversational AI and audio-language models. The 2026 arXiv record shows a coherent program around speech-language interfaces: ParA-LLM unifies paralinguistic and acoustic understanding, reaching 43.53% on its own ParA-Bench against GPT-4o-Audio's 36.03 with a +7.49 gain on MMAR; SteerDuplex uses SFT plus reinforcement learning to lift the steerability of the Moshi duplex speech model by 44.5 percentage points on a new SteerBench; VIBE targets video instruction-aligned background music generation; and TEMPO develops temporally-grounded multi-task post-training for large audio-language models. Evolve's reviews credit credible, honestly-reported gains while flagging evaluation caveats — a circular benchmark and weak statistics for ParA-LLM, thin baselines and single-judge scoring for SteerDuplex. Across 31 distinct coauthors, this researcher publishes in middle positions without corresponding authorship, within a large group anchored by Dinesh Manocha and Sreyan Ghosh. Radar scores are 69 for collaboration, 55 for reproducibility (two medium-readiness papers), and 0 for focus, the last reflecting the even split across two adjacent speech subfields rather than scattered work. Output, impact, and activity cannot be scored from the available data.
Collaboration pattern
Collaboration scores 69 on the radar, with 31 distinct coauthors in the snapshot. All four cataloged 2026 papers are arXiv efforts with middle authorship and no corresponding role, fitting a senior contributor within a large group. The stable core is Dinesh Manocha and Sreyan Ghosh; a wider ring of co-authors — including Utkarsh Tyagi, Ramaneswaran Selvakumar, Sonal Kumar, Vaibhavi Lokegaonkar, and Vishnu Raj — rotates across projects spanning dialogue models, audio-language models, and music generation.
Research keywords
Highlights
ParA-LLM: unified paralinguistic and acoustic speech understanding
SteerDuplex: +44.5pp steerability for duplex speech dialogue
TEMPO: temporally-grounded multi-task post-training for large audio-language models
A 2026 arXiv paper on post-training large audio-language models with temporal grounding; it is cataloged without an Evolve review, so no verdict, keywords, or reproducibility rating are available.
VIBE: video instruction-aligned background music generation
A 2026 arXiv paper generating background music aligned with video instructions; it is in the catalog without an Evolve review, keywords, or reproducibility assessment.
Timeline
- 2026Posts ParA-LLM on arXiv, unifying paralinguistic and acoustic understanding with 43.53% on its own ParA-Bench versus 36.03 for GPT-4o-Audio.
- 2026Posts SteerDuplex on arXiv, lifting duplex speech dialogue steerability by 44.5 percentage points on the new SteerBench via SFT plus reinforcement learning.
- 2026Posts TEMPO on arXiv, studying temporally-grounded multi-task post-training for large audio-language models.
- 2026Posts VIBE on arXiv, targeting video instruction-aligned background music generation.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Radar scores and coauthor counts derive from an OpenAlex-based bibliometric snapshot; no citation counts, h-index, or yearly output totals are included.
- Catalog coverage is limited to four 2026 arXiv papers; two carry Evolve reviews with full metadata, and two are catalog-only entries without verdicts or keywords.
- The reproducibility score averages only 2 rated papers (both medium readiness), so it is a narrow estimate.
- Identity is anchored to ORCID 0000-0002-5596-8460; the profile is name-resolved from public bibliographic data and does not cover earlier career outputs.