Skip to main content

Research & Open Benchmarks

Research and open benchmarks

We turn the methods behind our collection and verification work into reproducible benchmarks, protocols and datasets. The measurement conventions are public, ready for peers to test and reuse.

Research outputs
6 items
Years covered
2023–2026
Open benchmarks
3
Open datasets
1

01 / Index

Research and open assets, ordered by year

Every entry includes method notes and reuse instructions. Benchmarks and protocols can be reproduced directly; dataset entries require a license.

2026

1 output

  • BenchmarkTOMCUBE Research

    MC-Sync: A Synchronization Quality Benchmark for Multimodal Corpora

    Lin Che · Zhou Yuan · Chen Yan

    A reproducible protocol that measures audio, video and sensor synchronization error across six capture chains, based on 12,480 measured frames.

    1.4 ms

    Median sync error

    12,480

    Measured frames

    6

    Capture chain types

    View sample data

2025

2 outputs

  • BenchmarkTOMCUBE Research

    Dialect Speech Recognition Robustness Benchmark

    Lin Che · Chen Yan

    Word error rate across 12 dialect sites and three noise conditions, with practical guidance on capture-side robustness.

    12

    Dialect sites

    3

    Noise conditions

    +18.6 pp

    WER rise in street noise

    View dataset
  • ProtocolTOMCUBE Research

    Classroom Multimodal Annotation Protocol v2

    Zhou Yuan · Lin Che

    A layered annotation protocol for classroom corpora, from discourse acts and gesture to teaching events, with tag sets, boundary rules and double-blind review.

    κ = 0.89

    Median cross-layer agreement

    4

    Annotation layers

    36

    Pilot lessons

    View service

2024

2 outputs

  • BenchmarkTOMCUBE Research

    Clinical Conversation De-identification Benchmark

    Su Yao · Chen Yan

    Entity recall and utility loss for de-identification pipelines on clinician–patient transcripts, with a configuration that keeps both privacy and research value.

    99.2%

    Entity recall

    < 3%

    Downstream task loss

    1,200

    Sample transcripts

    View solution
  • DatasetTOMCUBE Research

    READ: An Open Chinese Reading Eye-Tracking Dataset

    Zhou Yuan · Lin Che

    Word-level eye-tracking records from 320 readers, including first fixation, regressions, reading times and text difficulty annotations.

    320

    Participants

    186,000

    Word-level records

    1,000 Hz

    Sampling rate

    View dataset

2023

1 output

  • PaperTOMCUBE Research

    Cross-Modal Alignment of Classroom Signals: An Empirical Study

    Lin Che · Zhou Yuan · Su Yao

    How teacher gesture, gaze and speech prosody couple in time during classroom explanation, with an analysis workflow others can reuse.

    24

    Classroom samples

    6,800

    Annotated segments

    120 ms

    Gesture lead over stress

02 / Citation & licensing

How to cite

Our research and benchmarks may be cited, reproduced and redistributed, subject to the same attribution and version rules.

03 / Open research program

Publish the methods, so results can be tested

Open research does not mean releasing all data. It means explaining measurement conventions, annotation rules and verification workflows so peers can judge whether a conclusion holds.

Open checklist

  • Benchmark evaluation protocols and measurement scripts
  • Partially anonymized data samples
  • Annotation manuals and codebooks
  • Verification report templates and field dictionaries

All materials are versioned and release notes are recorded on each output entry. Sample data lets you inspect structure and fields directly.

View sample data
01

Open benchmarks

Evaluation protocols, measurement scripts and result tables are public. Anyone can reproduce a conclusion with the same method, or publicly challenge it with a counterexample.

02

Joint research

We define research questions with partner teams, share collection capacity and annotation resources, and agree credit up front. Source data is handled under the ethics agreement.

03

Data access

Anonymized samples, variable dictionaries and analysis scripts are released on request. Data involving individuals or minors requires ethics review and a signed agreement.

Open research

Reuse a benchmark, or start a joint study with us

Tell us your research question, target population, data scale and timeline. We will come back with an executable study design, sampling plan and budget frame.

Reply within one business day · NDA and data licensing supported · Dedicated academic track