Roy Gan

—

Unclaimed · AI-structured profileOpenAlex

Streaming speech AI (TTS/ASR) and embodied robot planning researcher with a concentrated burst of 2026 arXiv output

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) 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—

This axis measures cumulative publication volume and research productivity over time. The current snapshot only shows 13 works indexed in 2026 with no fuller career-level output count available, so Evolve does not yet compute an output score for this researcher.

Impact—

This axis captures citation-based influence such as accumulated citations and h-index-type signals. The record currently carries no usable citation evidence — h-index and i10-index are both recorded as 0 — so impact is left unscored rather than estimated.

Focus0/100

Focus scores 0 because the field distribution is perfectly flat: speech processing (text-to-speech), speech processing (ASR), and embodied AI / robot learning each account for 33.3% of the record. Topic counts echo the split — robot manipulation and learning (8) and reinforcement learning in robotics (7) on one side, speech recognition and synthesis (4), speech and audio processing (4), and multimodal applications (4) on the other — a two-front agenda across speech and embodied robotics rather than a single-field concentration.

Collaboration75/100

Collaboration scores 75, grounded in 42 distinct coauthors across the record and a stable inner circle of 12 frequent collaborators, including Zehan Liu, Rime Wen, Kaiqi Fu, and Qian Wang. All three reviewed papers carry middle-author positions without corresponding-author roles, consistent with contributing within larger team-based projects rather than leading a small independent group.

Activity77/100

Activity scores 77, driven by an entirely recent record: 13 works indexed in 2026 and 13 works over the last three years — the same figure, meaning the whole indexed output falls within the current year. All three reviewed papers appeared as arXiv preprints in 2026, indicating a burst of near-simultaneous submission activity rather than a long publication history.

Reproducibility32/100

Reproducibility scores 32, the mean across three reviewed papers. X-Planner and X2Streaming-ASR are rated low — the former lacks ablations and significance testing, the latter raises open comparability questions — while X2-NativeCursor reaches medium, though its gains are teacher-relative and measured on synthetic speech. The pattern shows strong reported numbers with limited reproducibility safeguards.

Research summary

Roy Gan is a name-resolved profile built from public OpenAlex bibliographic data; no verified institution, country, or ORCID is attached to the record. The indexed output is unusually concentrated: 13 works appear in 2026 and recent3yWorks is also 13, meaning the entire indexed record falls within the current year. Research is split evenly across three fields — speech processing (text-to-speech), speech processing (automatic speech recognition), and embodied AI / robot learning — each at 33.3% of the field distribution. Three Evolve-reviewed arXiv preprints span this range: X-Planner (event-structured task planning with parallel latent chain-of-thought for robots), X2-NativeCursor (native-token text progress tracking for streaming codec TTS, MAE 0.151 vs 1.253), and X2Streaming-ASR (a GRPO-trained wait/emit policy cutting commit latency to 24-97 ms from 409-585 ms baselines). All three record middle-author positions without corresponding-author roles. Topic data concentrates on robot manipulation and learning (8) and reinforcement learning in robotics (7), with speech topics at 4 each. Collaboration is broad: 42 distinct coauthors and 12 recurring partners. Evolve's reproducibility scoring averages 32/100 across the three papers — two rated low, one medium — citing missing ablations, absent significance testing, and open comparability questions. Output and impact axes remain unscored pending sufficient data.

Collaboration pattern

With 42 distinct coauthors and 12 recurring collaborators — Zehan Liu, Rime Wen, Kaiqi Fu, Shawn Qin, Hao Wang, Carl Chen, Altman Lin, Lights Shi, Qian Wang, Howard Lu, Shalfun Li, and Porter Pan — Roy Gan works in sizeable team settings spanning both speech and robotics tracks. All three reviewed papers record middle-author positions without corresponding-author roles, pointing to a contributor role inside larger group efforts rather than sole or senior leadership.

Research keywords

task planningvision-language-action modelsembodied intelligencerobot manipulationlatent chain-of-thought reasoningstreaming text-to-speechnative speech codec tokenstext progress trackingstreaming ASRcommit latencyreinforcement learning (GRPO)multimodal machine learning

Highlights

Streaming ASR commit latency cut to 24-97 ms with a GRPO wait/emit policy
X2Streaming-ASR trains a three-stage GRPO wait/emit policy over a decoder-only language model, reaching 24-97 ms commit latency versus 409-585 ms baselines. Evolve flags the strong results alongside open comparability questions; reproducibility readiness is low.
Native-token observer tracks streaming-TTS text progress at MAE 0.151
X2-NativeCursor tracks text progress directly from native codec tokens for incremental-text streaming TTS, reporting MAE 0.151 against 1.253 for the comparison baseline. Evolve notes the gains are large but teacher-relative and evaluated on synthetic speech; reproducibility readiness is medium.
Event-structured planning with parallel latent chain-of-thought for robots
X-Planner introduces event-structured planning data and parallel latent chain-of-thought for embodied intelligence, reporting large real-robot wins. Evolve notes the absence of ablations and significance testing, and that it trails kimi3 offline; reproducibility readiness is low.

Timeline

2026X2Streaming-ASR posted on arXiv: GRPO-trained wait/emit streaming ASR with 24-97 ms commit latency versus 409-585 ms baselines.
2026X2-NativeCursor posted on arXiv: native-token text progress tracking for incremental-text streaming codec TTS (MAE 0.151 vs 1.253).
2026X-Planner posted on arXiv: event-structured task planning with parallel latent chain-of-thought for embodied intelligence.

Papers in the Evolve catalog

Frequent co-authors

Zehan Liu×2 Rime Wen×2 Kaiqi Fu×2 Shawn Qin×2 Hao Wang×2 Carl Chen×1 Altman Lin×1 Lights Shi×1 Qian Wang×1 Howard Lu×1 Shalfun Li×1 Porter Pan×1

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 OpenAlex bibliographic data; no verified institution, country, or ORCID is attached to this record.
Bibliometrics come from the OpenAlex snapshot only; citation coverage appears sparse (h-index and i10-index recorded as 0), so the output and impact axes are unscored.
All reviewed papers are 2026 arXiv preprints with middle-author positions; no peer-reviewed venue, first-author, or corresponding-author records are available in this scope.
Reproducibility is averaged over only 3 reviewed papers, and the flat three-field split (33.3% each) is what depresses the focus score.