NRC Canada speech AI researcher: deepfake detection, speech-model quantization, and auditable text-to-speech
3
papers here
3
reviewed by Evolve
3
first-author
1
corresponding
—
works · OpenAlex
18
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—
Not yet scored: Evolve lacks a verified complete works list for this OpenAlex identity, so publication volume cannot be computed. This axis would measure annual and cumulative output, venue mix, and authorship position across the career.
Impact—
Although h-index 18 and i10-index 27 are reported, citation coverage is judged insufficient for a defensible impact score. This axis would measure citation trajectory, field-normalized influence, and downstream uptake of the anti-spoofing and speech-synthesis work.
Focus0/100
The focus score of 0 reflects maximal breadth inside one domain: the field distribution splits evenly at one-third each across speech synthesis and spoken language systems, speech and audio machine learning, and deepfake detection. Topic counts reinforce the spread — speech recognition and synthesis (108), speech and audio processing (100), voice and speech disorders (40), dialogue systems (20), neural network applications (14) — yet all clusters remain within speech and audio.
Collaboration36/100
With a collaboration score of 36, this profile shows a compact, stable network: 5 distinct coauthors over the last three years, including recurring partners Jinxi Jin, Junhao Xu, Yujia Lu, Manuela Kunz, and Patrick Littell. All three Evolve-reviewed 2026 papers are first-authored at the National Research Council Canada, one as corresponding author, suggesting a lead-author role within a small core team rather than large consortium work.
Activity100/100
Activity scores a perfect 100, backed by the counts-by-year series: from 1 work in 2019 to 6 in 2020, 11 in 2021, 16 in 2022, 17 in 2023, 16 in 2024, then peaks of 24 in 2025 and 23 in 2026. The last three years alone account for 63 works, indicating a sustained high-tempo phase now centered on arXiv-first output.
Reproducibility85/100
Reproducibility averages 85 across 3 Evolve-assessed papers. The Qwen2-Audio quantization case study and the VoxReason benchmark both carry high reproducibility readiness — the latter ships a 100-case benchmark with a deterministic verifier — while the deepfake-detection decision-record paper is rated medium, with the mechanism behind its ASVspoof-5 gain (8.43% vs 11.91% EER) still unresolved. The pattern favors transparent protocols and honest negative results.
Research summary
Mengzhe Geng is a speech and audio machine learning researcher at the National Research Council Canada whose recent work concentrates on making speech AI systems trustworthy, measurable, and honest about their limits. Three 2026 arXiv papers — all first-authored, one as corresponding author — anchor the profile. The first is a quantization case study on Qwen2-Audio showing that text-space scores can miss how quantized models actually use waveforms, with a protocol covering mixed-precision allocation and packed memory on RAVDESS emotion recognition. The second proposes auditable decision records for speech deepfake detection, cutting matched ASVspoof-5 EER from 11.91% under fixed fusion to 8.43%, while openly noting the result still trails the best passive WavLM baseline at 6.71%. The third, VoxReason, contributes a 100-case benchmark that audits source-grounded speech plans before synthesis via cue edits and a deterministic verifier. OpenAlex bibliometrics (h-index 18, i10-index 27) show a steep ramp: yearly works rose from 1 in 2019 to 24 in 2025 and 23 in 2026, with 63 works in the last three years. Topic mass sits in speech recognition and synthesis (108), speech and audio processing (100), and voice and speech disorders (40). Collaboration stays compact at 5 distinct recent coauthors, indicating a lead-author style within a small core team. This profile is name-resolved from public bibliographic data.
Collaboration pattern
The collaboration score of 36 reflects a compact network: 5 distinct coauthors in the last three years, with Jinxi Jin, Junhao Xu, Yujia Lu, Manuela Kunz, and Patrick Littell recurring. All three Evolve-reviewed papers are first-authored at the National Research Council Canada — VoxReason also corresponding — pointing to a lead-author working style within a small, stable core team rather than large consortium projects.
Research keywords
Highlights
Quantization case study: text scores can miss waveform use
Auditable decision records improve speech deepfake detection
VoxReason: auditing source-grounded speech plans before synthesis
Timeline
- 2019First recorded works appear in the OpenAlex counts-by-year series (1 work).
- 2021Output accelerates to 11 works, up from 6 in 2020.
- 2023Sustained mid-phase with 17 works recorded in the year.
- 2025Peak activity with 24 works; the last three years total 63.
- 202623 works recorded; three first-authored arXiv papers enter the Evolve catalog, with VoxReason also corresponding-authored.
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 (identity: openalex); output and impact axes are unscored because citation coverage is judged insufficient.
- The profile is name-resolved from public bibliographic data anchored to ORCID 0000-0002-7886-439X; no biographical or award information was verified.
- Qualitative assessments rest on only three Evolve-reviewed 2026 arXiv preprints; the reproducibility average of 85 covers those 3 papers.
- h-index 18 and i10-index 27 are quoted as reported; venue- and field-normalized citation metrics were not available.