← News

ResearchSources: arXiv · 20 Nov 2026

Xing-Audio-NVVASR transcribes speech with inline tags for non-verbal vocalizations

A 1.7B-parameter model trained on roughly 800,000 curated utterances reports the best composite score among nine systems on four human-annotated benchmarks, according to the paper's abstract.

In plain words

The researchers built a speech recognizer that also marks non-verbal sounds like breaths or laughter inside its transcripts. Because such events are rarely labeled well in training data, they used two models to check each other's labels and had humans confirm disputed cases. The result is one model trained on a large cleaned-up corpus.

What was announced

The paper describes a progressive data pipeline: two heterogeneous models cross-audit annotations against unified acoustic criteria, flagging disagreements for teacher reannotation, while breath detectors screen the corpus for omissions that model agreement might hide. Human-confirmed labels accumulate over five rounds, refining 170,000 utterances from public corpora and adding 630,000 mined utterances. The final model transcribes speech with inline tags for six non-verbal vocalization classes.

On four human-annotated benchmarks, the authors report that Xing-Audio-NVVASR ranks first among nine models in composite score and six-class Macro-F1, with a mean composite score of 64.46 — 6.03 points above StepAudio 3 Chat Preview — and breathing F1 exceeding its annotation teacher by 14.40 to 31.57 percentage points. The authors state they plan to release model weights, inference code, the four benchmarks, annotation guidelines, and evaluation code.

What it does not show

  • Evidence is limited to the paper's abstract; no independent evaluation or full-text detail is available.
  • Reported rankings and score margins come from the authors' own benchmarks and have not been confirmed elsewhere.
  • The abstract says the team plans to release weights and code; actual availability is not yet established.

For security and IT teams

Status
Model weights, inference code, benchmarks, annotation guidelines, and evaluation code are planned for release but not yet confirmed as available.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.