Xiangang Li

Output—

Not scored: this snapshot carries no publication-volume data. The output axis would measure total research production — number of papers, venues, and arXiv preprints over time. Only two Evolve-reviewed entries exist so far, which is too thin to characterize.

Impact—

Not scored: citation counts, h-index, and per-paper influence are absent from this snapshot. The impact axis would measure how widely the cataloged speech-processing work is cited and adopted; no citation figure is available to quote here.

Focus0/100

The snapshot's field distribution splits exactly 50/50 between automatic speech recognition and audio-language models for conversational AI, producing a focus score of 0. Read against the catalog, the spread is not scattered topic-hopping: both halves come from one Qwen-Audio line that moves from transcription-grade MoE ASR (3.0) toward full-duplex agentic voice interaction (3.1), a deliberate expansion within speech processing rather than across unrelated fields.

Collaboration83/100

Collaboration is the strongest scored axis at 83, backed by 62 distinct coauthors in the raw record and a stable recurring core of twelve named collaborators — including Xu Li, Zhifu Gao, Zhigao Chen, Peixin Chen, and Chuanmeng Bian. Both cataloged papers place Li in a middle author slot with no corresponding-author credit, consistent with contribution inside a large industrial speech team rather than small-group or solo work.

Activity—

Not scored: the snapshot provides no publication-cadence or recency metrics. The activity axis would measure output rhythm over time — papers per year and gaps between releases. Only two 2026 entries are cataloged, insufficient to characterize a pace.

Reproducibility20/100

Evolve's reproducibility read averages 20 across the 2 cataloged papers, both rated low readiness. The verdicts are specific about why: Qwen-Audio-3.1-Realtime reports extensive agentic voice gains from distillation and GRPO but ships no ablations, confidence intervals, or independent verification, while the 3.0-ASR report's state-of-the-art Chinese/English and multilingual claims rest largely on unverifiable internal test sets. Engineering claims currently outrun what outsiders can replicate.