Speech and audio AI researcher at CUHK-Shenzhen: audio language models, TTS evaluation, and deepfake detection.
4
papers here
4
reviewed by Evolve
0
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—
No publication-volume metric is available for this axis yet. It would measure total research output, such as paper counts per year, venue types and authorship roles. The four Evolve-reviewed arXiv papers, all dated 2026, are currently the only output evidence in the dataset.
Impact—
Citation counts, h-index and downstream influence are absent from the input data, so this axis cannot be scored. Once citation records are linked, it would measure how widely the four 2026 papers are cited, reused and built upon by other work.
Focus0/100
The focus score is 0, and the raw field distribution explains why: the four catalog fields each hold exactly a 0.25 share — audio machine learning (speech, sound and music understanding), speech and audio signal processing (deepfake detection), speech processing/audio AI (TTS evaluation), and speech and natural language processing. Coverage is perfectly even with no dominant specialization, a deliberately broad portfolio across adjacent speech and audio subfields rather than a single concentrated line.
Collaboration70/100
The collaboration score of 70 is computed from 33 distinct coauthors, and the frequent-coauthor list — including Yuxiang Wang, Zifan Guan, Longyu Lu, Junan Zhang and Meiguang Jin — indicates a broad, recurring working network. All four catalog papers are multi-author efforts in which this researcher appears in middle or last author positions, never as corresponding author, consistent with a team-embedded contributor sharing credit across a lab-sized group.
Activity—
The dataset provides no publication cadence, career-length or recency statistics beyond four papers all dated 2026, so ongoing activity cannot be scored. This axis would measure publishing rhythm and sustained engagement over time.
Reproducibility29/100
Reproducibility scores a mean of 29 across 4 assessed papers, matching the raw reproMean of 29 and reproPapers of 4. Per-paper readiness is low for EvoAudio (no error bars or code), the think-aloud spoken language model (proprietary data, no interleaving baseline) and the prosody-judgment TTS work (single-seed results on private data), and medium for CoRELoop, which offers strong ablations but no external SOTA comparison. This is the profile's clearest weakness.
Research summary
Zhizheng Wu is a speech and audio AI researcher affiliated with the Chinese University of Hong Kong, Shenzhen, whose Evolve-reviewed catalog of four 2026 arXiv papers spans four subfields in equal measure: audio machine learning for speech, sound and music understanding; speech and audio signal processing for deepfake detection; speech processing and audio AI for TTS evaluation; and speech and natural language processing. The through-line is reinforcement learning and self-improvement applied to audio systems: EvoAudio builds a closed loop that synthesizes verifiable audio QA and reports gains up to 6.3 points on five LALMs over 13 rounds; a live-streaming TTS study distills Gemini into an open prosody judge trained with span-local GRPO; and Spoken Language Models that Think Aloud cuts reasoning silence from 12.8s to 0.36s at near-chain-of-thought accuracy. CoRELoop applies LoRA-based recurrent refinement to a frozen self-supervised deepfake detector, reducing pooled EER from 4.85% to 3.74%. Collaboratively, 33 distinct coauthors and a collaboration score of 70 point to a team-embedded working pattern, with this researcher in middle or last author positions and never as corresponding author. Reproducibility is the open question: a mean readiness of 29 across four assessed papers, three rated low and one medium, reflects private data, missing code and single-seed results.
Collaboration pattern
The collaboration score of 70 rests on 33 distinct coauthors. The recurring group includes Yuxiang Wang, Zifan Guan, Longyu Lu, Junan Zhang, Meiguang Jin, Junfeng Ma, Kunyu Feng, Li Wang, Wan Lin, Shengbo Cai and Yingda Shen, suggesting a stable lab cluster. Across the four 2026 catalog papers this researcher holds last-author placement twice (EvoAudio, CoRELoop) and middle positions otherwise, never corresponding author — a senior-collaborator rather than sole-lead pattern.
Research keywords
Highlights
Recursive self-improvement for audio understanding
Spoken language models that think aloud
Open prosody judge for live-streaming TTS
Parameter-efficient recurrent refinement for deepfake detection
Timeline
- 2026CoRELoop: parameter-efficient controlled recurrent refinement for audio deepfake detection (arXiv, last author).
- 2026Multi-Dimensional Prosody Judgment for live-streaming speech synthesis (arXiv, middle author).
- 2026Spoken Language Models that Think Aloud: asynchronous think-aloud decoding for real-time voice interaction (arXiv, middle author).
- 2026EvoAudio: recursive self-improvement for audio understanding (arXiv, last author).
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 scores are computed from OpenAlex-derived data; no citation, h-index or venue-level statistics were available.
- All four reviewed papers are 2026 arXiv preprints; peer-reviewed publication status is unknown.
- This profile is name-resolved from public bibliographic data; no ORCID was provided, so disambiguation among same-name authors is not guaranteed.
- Reproducibility and focus scores reflect Evolve's assessment of the four catalog papers only; other works were not evaluated.