Speech-to-speech researcher building open full-duplex voice models with native tool calling
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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—
Not scored: the OpenAlex snapshot records no reliable publication-volume signal beyond three works in 2026, with h- and i10-indexes at zero. This axis would measure cumulative research output — papers per year across venues and career stage.
Impact—
Not scored: no citation counts, h-index, or venue-level influence are available in the OpenAlex snapshot, so influence cannot be estimated. This axis would measure citations, citation velocity, and uptake of the researcher's speech models by others.
Focus0/100
Radar scores focus at 0, but the underlying field distribution is narrow, splitting evenly (0.5/0.5) between "speech processing and conversational AI" and "speech processing / spoken dialogue systems (cs.CL)" — two adjacent labels. Topic counts reinforce this: speech and audio processing (3), speech and dialogue systems (3), speech recognition and synthesis (2), multimodal ML applications (1). Visible work sits almost entirely in full-duplex speech-to-speech with tool calling.
Collaboration83/100
Collaboration scores 83, backed by 64 distinct coauthors in the recent window — a wide footprint. Both reviewed papers are middle-author, non-corresponding contributions, consistent with team-scale research. The recurring roster includes Edresson Casanova, Zhehuai Chen, Slyne Deng, Jagadeesh Balam, Elena Rastorgueva, and Travis Bartley, indicating a stable core group repeating across speech-to-speech and dialogue-system projects.
Activity40/100
Activity scores 40. The OpenAlex counts-by-year series contains a single entry — 3 works in 2026 — and recent3yWorks is 3, so observable recent output is real but thin. Both visible 2026 papers appeared on arXiv within the same narrow area, suggesting concentrated sprints around one problem rather than sustained high-volume publishing.
Reproducibility38/100
Reproducibility scores 38, the mean over 2 scored papers. NemotronLabs VoiceChat rates medium readiness, described as an open model, while the frontend-backend architecture paper rates low — Evolve notes strong benchmark results but no significance testing and low reproducibility, and VoiceChat's design claims are left untested.
Research summary
Nikhil Srihari is a speech-processing researcher whose visible work centers on full-duplex speech-to-speech (S2S) modeling and on giving spoken dialogue systems the ability to act. OpenAlex records three works in 2026, two of which Evolve has reviewed — both arXiv preprints in the same narrow lane: adding tool calling to full-duplex speech models. The first, NemotronLabs VoiceChat, is described by Evolve as the first open full-duplex S2S model with native tool calling, with strong turn-taking and barge-in behavior, though argument accuracy remains weak (42.2%, Pass@1 33.0%) and some design claims are untested. The second proposes a frontend-backend architecture in which a delegation token lets the speech frontend hand tool calls to a backend, reporting strong benchmark results but with no significance testing. Both papers list Srihari as a middle author rather than first or corresponding, consistent with large-team research: OpenAlex credits 64 distinct coauthors in the recent window, and the recurring roster (Edresson Casanova, Zhehuai Chen, Jagadeesh Balam, Elena Rastorgueva, Travis Bartley, and others) indicates a stable core group. Radar scores collaboration 83, activity 40, and reproducibility 38 — the mean over two scored papers — while output and impact cannot be scored because citation and volume data are unavailable in the snapshot.
Collaboration pattern
OpenAlex credits 64 distinct coauthors in the recent window, and radar scores collaboration at 83 — a wide footprint. Both reviewed papers are middle-author, non-corresponding contributions, the signature of team-scale research. The recurring roster — Edresson Casanova, Zhehuai Chen, Slyne Deng, Oluwatobi Olabiyi, Elena Rastorgueva, Viet Anh Trinh, Jagadeesh Balam, Travis Bartley, Sanjay Chauhan, Chen Chen — suggests a stable core group spanning speech synthesis, recognition, and dialogue rather than a small single-PI lab.
Research keywords
Highlights
First open full-duplex speech-to-speech model with native tool calling
Delegation-token architecture for tool calls in duplex speech models
Honest limits: untested claims and thin evaluation
Both reviews carry caveats — VoiceChat's argument accuracy (42.2%, Pass@1 33.0%) and untested design claims, and the architecture paper's lack of significance testing — marking evaluation rigor as the clearest gap in this line of work.
A concentrated 2026 sprint on voice agents
OpenAlex records all 3 works in 2026 within speech processing and conversational AI, and both reviewed papers attack the same problem: giving full-duplex speech models the ability to call tools mid-conversation.
Timeline
- 2026OpenAlex records 3 works in 2026, all within speech processing and conversational AI.
- 2026Co-authored NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with native tool calling (middle author, arXiv).
- 2026Co-authored a frontend-backend architecture adding tool calls to full-duplex speech models via a delegation token (middle author, arXiv).
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.
- Bibliometrics come from OpenAlex only; the snapshot reports h-index and i10-index of 0, so the output and impact dimensions are left unscored rather than estimated.
- Institution, country, and ORCID are absent; this profile is name-resolved from public bibliographic data and may merge or split individuals with the same name.
- Only 3 works appear in the recent 3-year window, both reviewed papers are arXiv preprints, and no peer-reviewed venues are recorded.
- The reproducibility score (38) is a mean over 2 papers — one medium, one low readiness — too small a base for a stable rating.