Audio/speech ML researcher: fairness auditing, adversarial robustness, and turn-taking in spoken dialogue models.
3
papers here
3
reviewed by Evolve
3
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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—
Publication output is not yet scored: the profile rests on three 2026 arXiv preprints, with no venue history or totals. This axis would measure total and annual publication volume and venue distribution.
Impact—
Impact is not yet scored: no citation counts, h-index, or citation-year history are available in the current snapshot. This axis would measure citations to the researcher's work and how that uptake develops over time.
Focus0/100
Focus is not yet scored: the three reviewed papers all sit in audio and spoken-dialogue machine learning, but too few indexed works exist to compute topical concentration. This axis would measure thematic breadth across fields.
Collaboration14/100
Collaboration scores 14 out of 100, based on exactly one distinct coauthor recorded in the bibliometric data: Iman Modarressi, a frequent coauthor on this work. All three catalogued 2026 papers are first-authored by Shamsaie, none as corresponding author, indicating a small two-person collaboration model so far.
Activity—
Activity is not yet scored: all three catalogued preprints carry 2026 dates, too short a window to establish a publication cadence. This axis would measure output rate and consistency of publishing over time.
Reproducibility55/100
Reproducibility is not yet scored: no code, data, or artifact-release information is recorded for any of the three papers. This axis would measure whether released artifacts let others independently verify the reported results.
Research summary
Kian Shamsaie is an early-career researcher in audio and speech machine learning whose profile is built from a small but thematically coherent set of three first-authored arXiv preprints, all dated 2026. The work addresses three complementary questions about audio-centric AI systems. First, an audit-and-repair study argues that representation entanglement can lower-bound demographic disparity in audio understanding models and turns that link into concrete fairness repairs. Second, a robustness study proposes psychoacoustically aligned latent smoothing to harden full-duplex speech-to-speech dialogue models against adversarial perturbations. Third, an evaluation study reframes turn-taking in full-duplex spoken dialogue models, arguing that neither silence nor overlap is intrinsically failure and introducing an intent-conditioned evaluation protocol. Together, the papers span fairness, robustness, and interaction-quality evaluation for spoken dialogue systems. Bibliometric context is thin: the collaboration dimension scores 14 out of 100 with exactly one distinct coauthor recorded, Iman Modarressi, and no citation, output-volume, or venue data is yet available. The name is resolved from public bibliographic data; no institutional affiliation, country, or ORCID appears in the current snapshot. This is therefore an early-stage profile: the thematic focus on trustworthy spoken-dialogue AI is clear, but quantitative impact and output cannot yet be characterized.
Collaboration pattern
The recorded collaboration footprint is minimal: one distinct coauthor, Iman Modarressi, appears as the frequent coauthor across all three 2026 first-authored arXiv papers. The collaboration dimension scores 14 out of 100, consistent with a tight two-person working model in which Shamsaie leads as first author and no corresponding-author role is recorded. No institutional or multi-lab collaborations are visible in the current snapshot.
Research keywords
Highlights
Entanglement lower-bounds demographic disparity in audio models
An audit-and-repair study of audio understanding models shows representation entanglement can impose a lower bound on demographic disparity, and uses that link to design demographic fairness repairs.
Psychoacoustically aligned latent smoothing for speech robustness
Proposes latent smoothing aligned with psychoacoustics to strengthen adversarial robustness of full-duplex speech-to-speech dialogue models.
Intent-conditioned evaluation of turn-taking in full-duplex dialogue
Argues that neither silence nor overlap is intrinsically failure in full-duplex spoken dialogue models, and introduces an intent-conditioned protocol for evaluating turn-taking.
Timeline
- 2026First-authored arXiv paper audits and repairs demographic fairness in audio understanding models via an entanglement-disparity link.
- 2026First-authored arXiv paper proposes psychoacoustically aligned latent smoothing for adversarial robustness of full-duplex speech-to-speech dialogue models.
- 2026First-authored arXiv paper introduces intent-conditioned evaluation of turn-taking, treating neither silence nor overlap as inherent failure.
Papers in the Evolve catalog
- first author
When Entanglement Lower-Bounds Disparity: Auditing and Repairing Demographic Fairness in Audio Understanding Models
- first author
Psychoacoustically Aligned Latent Smoothing for Adversarial Robustness of Full-Duplex Speech-to-Speech Dialogue Models
- first author
Neither Silence nor Overlap Is Failure: Intent-Conditioned Evaluation of Turn-Taking in Full-Duplex Spoken Dialogue Models
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.
- Name resolved from public bibliographic data; no institution, country, or ORCID is attached, so identity disambiguation is limited.
- All three papers are arXiv preprints dated 2026; no peer-reviewed venue, citation, or reproducibility-artifact information is available.
- Five of the six radar dimensions (output, impact, focus, activity, reproducibility) lack sufficient underlying data and are left unscored.
- The profile draws only on papers in the Evolve catalog; other work by this researcher may exist outside this snapshot.