UT Austin speech and multimodal ML researcher: SpeechLLM reasoning, voice-agent benchmarks, and speech intelligibility
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
23
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—
No output score is available yet because work-level output records are still incomplete in the underlying data. This axis would measure cumulative publication volume and its distribution over years; only partial aggregate indicators are currently on record.
Impact—
No impact score is available yet. This axis would measure citation-based influence — citations per work, field-normalized impact, and uptake by other researchers — once work-level citation records are complete in the pipeline.
Focus0/100
The focus score is 0, reflecting an evenly spread field distribution rather than narrow concentration: field weights split equally (0.33 each) across conversational speech AI, SpeechLLM research, and computational psycholinguistics/speech technology. Topic counts likewise span speech recognition and synthesis (87), speech and audio processing (80), multimodal ML applications (41), music and audio (26), and phonetics and phonology (17) — broad coverage of adjacent speech areas rather than one specialized track.
Collaboration58/100
The collaboration score is 58. The recent publication window records 17 distinct co-authors, and the recurring collaborator list is dense with speech-community researchers, including Shinji Watanabe, Hung-yi Lee, Kai-Wei Chang, Abdelrahman Mohamed, Siddhant Arora, Shu-wen Yang, Guan-Ting Lin, Puyuan Peng, Yi-Jen Shih, Shih-Yun Shan Kuan, and Eunjung Yeo. On the three Evolve-reviewed papers he sits in the last-author position, consistent with a senior, supervisory role in group projects.
Activity100/100
Activity scores 100, the maximum. OpenAlex counts-by-year show sustained acceleration: 3 works in 2016 rising to 13 in 2022, 16 in 2023, 18 in 2024, 13 in 2025, and 21 in 2026 — the highest annual count on record — with 52 works in the most recent three-year window. Three 2026 arXiv preprints in the Evolve catalog corroborate the pace. The trajectory indicates a researcher operating at peak productivity.
Reproducibility55/100
Reproducibility scores 55, computed as the mean over 3 papers assessed by Evolve. All three — MP-Bench, the consonant-intelligibility masking study, and RetroThinker — carry a medium reproducibility-readiness rating. Evolve's verdicts temper replication confidence: fully synthetic clean audio for MP-Bench, correlations awaiting human validation for the masking study, and single-benchmark evidence for RetroThinker.
Research summary
David Harwath is a researcher at The University of Texas at Austin working across speech and language processing, speech and audio processing, and multimodal machine learning. His OpenAlex-recorded profile shows an h-index of 23 and an i10-index of 38, with output that has accelerated sharply — from single-digit annual counts through the 2010s to 13 works in 2022, 16 in 2023, 18 in 2024, and 21 in 2026, including 52 works in the most recent three-year window. His topic profile is anchored in speech recognition and synthesis (87 works) and speech and audio processing (80), extending into multimodal machine learning applications (41), music and audio processing (26), and phonetics and phonology (17). Three 2026 arXiv papers reviewed by Evolve illustrate the range: MP-Bench, the first audio benchmark for voice agents as multiparty conversation participants; RetroThinker, which trains streaming SpeechLLMs to self-correct their chain-of-thought while remaining forward-only, gaining 11 accuracy points on spoken GSM8K at similar latency; and an ASR-based acoustic-masking study quantifying consonant contributions to word intelligibility across four languages. He sits in the last-author position on all three, working within a broad network of speech-community collaborators.
Collaboration pattern
Harwath's recent window records 17 distinct co-authors, and his recurring collaborator list clusters around the speech and SpeechLLM community: Shinji Watanabe, Hung-yi Lee, Kai-Wei Chang, Abdelrahman Mohamed, Siddhant Arora, Shu-wen Yang, Guan-Ting Lin, Puyuan Peng, Yi-Jen Shih, Shih-Yun Shan Kuan, and Eunjung Yeo. On all three Evolve-reviewed papers he holds the last-author position without being the corresponding author, suggesting senior oversight of group projects run with a wide, speech-focused network.
Research keywords
Highlights
First benchmark for voice agents in multiparty conversation
RetroThinker: streaming SpeechLLMs learn to self-correct
Quantifying consonants' role in word intelligibility
Timeline
- 2010Earliest year in the OpenAlex record, with 1 indexed work.
- 2019Annual output reaches 8 indexed works, a step up from 3–4 per year in the preceding years.
- 2022Output accelerates to 13 indexed works, beginning a multi-year climb in publication rate.
- 202418 indexed works in a single year, the highest annual count to that point.
- 202621 indexed works, the record high, including three Evolve-reviewed arXiv preprints: MP-Bench, RetroThinker, and the consonant-intelligibility masking study.
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.
- Bibliometric aggregates (h-index, i10-index, per-year counts, topics, co-author counts) come from OpenAlex; identity is resolved from public bibliographic data.
- Output and impact dimension scores are not yet computed; only h-index (23) and i10-index (38) are available as aggregate indicators.
- The three reviewed papers are 2026 arXiv preprints; Evolve verdicts are editorial assessments rather than peer review, and none lists Harwath as corresponding author.
- Co-author affiliations are not provided in the data and are not verified here.