首页 > AI前沿 > Xing-Audio-NVVASR: Scaling and Refining Supervision for Joint Recognition of Speech and Non-Verbal Vocalizations

Xing-Audio-NVVASR: Scaling and Refining Supervision for Joint Recognition of Speech and Non-Verbal Vocalizations

arXiv自然语言 2026-11-19 20:21 9 阅读 查看原文
用户需要我作为一个专业的新闻排版工程师,把一段纯文本的新闻正文排版成语义化的HTML片段。 分析用户的要求: 1. 只输出HTML片段本身,不要Markdown围栏(如```html),不要任何解释。 2. 每个自然段落用`

`包裹,长文用`

`/`

`划分小节。 3. 列表内容用`
  • `,引用用`
    `,代码用`
    `,重要句子可用``。
    4. 保留原文所有信息,不得增删改内容。
    5. 不要输出``/``/``等文档级标签。
    
    现在分析提供的正文:
    "Joint recognition of speech and non-verbal vocalizations (NVVs) is hindered by scarce event-rich training data and inaccurate or missing event labels. We present Xing-Audio-NVVASR, an end-to-end recognizer trained through progressive data mining and iterative supervision refinement. Guided by unified acoustic criteria, two heterogeneous models cross-audit annotations, flagging disagreements for teacher reannotation. To uncover shared omissions that model agreement may conceal, complementary breath detectors screen the entire corpus for human review. Human-confirmed labels accumulate across rounds and are preserved during subsequent refinement. Over five rounds, we refine annotations for 170,000 utterances from public corpora and incorporate 630,000 mined utterances. The resulting corpus of approximately 800,000 utterances is used to train a single 1.7B-parameter model that transcribes speech with inline tags for six NVV classes. Across four human-annotated benchmarks, Xing-Audio-NVVASR ranks first among nine models in both composite score and six-class Macro-F1. It achieves a mean composite score of 64.46, surpassing the strongest comparison model, StepAudio 3 Chat Preview, by 6.03 points, and exceeds the annotation teacher's breathing F1 by 14.40 to 31.57 percentage points. We plan to release the model weights, inference code, four evaluation benchmarks, annotation guidelines, and evaluation code."
    
    这段文本是一个学术/技术新闻的摘要,没有明显的小节标题,也没有列表、引用或代码。它主要是一个长段落。
    
    根据规则2:"每个自然段落用

    包裹,长文用

    /

    划分小节"。由于原文没有明确的小节,我可以添加一些合适的小节标题来划分,但规则4说"不得增删改内容"。添加小节标题算不算"增"内容?这有点歧义。为了安全起见,如果原文没有标题,我可以不添加,或者根据语义合理划分并添加标题。但严格来说,"不得增删改内容"意味着不能添加原文没有的文字。所以,最好不添加标题,直接用`

    `包裹整个段落,或者根据句子逻辑拆分成多个`

    `,但不添加额外的标题文字。 不过,规则2说"长文用

    /

    划分小节",如果原文没有小节,就不需要强制添加。我可以将整个文本作为几个自然段来处理,或者就一个段落。原文在视觉上是一整个段落。我可以把它分成几个逻辑段落,或者保持为一个段落。 让我再仔细看原文。它确实是一整段。我可以保持为一个`

    `,或者根据语义拆成几个`

    `。为了体现"排版",我可以适当拆分,但不改变文字。比如: 第一句讲问题和提出的方法。 第二句讲方法细节。 第三句讲结果。 第四句讲未来计划。 拆分成多个`

    `是合理的,因为"每个自然段落用

    包裹"。原文虽然连在一起,但逻辑上可以看作几个自然段。不过,为了严格遵守"不得增删改内容",拆分段落不算增删内容,只是排版。 让我考虑是否要加``来突出重要句子。规则3说"重要句子可用"。我可以将关键结论用``包裹,比如"Xing-Audio-NVVASR ranks first among nine models in both composite score and six-class Macro-F1." 以及具体的分数数据。 但要注意,"不得增删改内容",用``包裹只是加标签,不改变文字内容,所以是允许的。 现在设计HTML结构: 由于没有列表、引用、代码,也不需要添加新标题(避免增加内容),我将原文按逻辑拆分为几个段落,用`

    `包裹,并对重要句子使用``。 原文拆分: 段落1(问题与方法提出): Joint recognition of speech and non-verbal vocalizations (NVVs) is hindered by scarce event-rich training data and inaccurate or missing event labels. We present Xing-Audio-NVVASR, an end-to-end recognizer trained through progressive data mining and iterative supervision refinement. 段落2(方法细节): Guided by unified acoustic criteria, two heterogeneous models cross-audit annotations, flagging disagreements for teacher reannotation. To uncover shared omissions that model agreement may conceal, complementary breath detectors screen the entire corpus for human review. Human-confirmed labels accumulate across rounds and are preserved during subsequent refinement. 段落3(数据与模型): Over five rounds, we refine annotations for 170,000 utterances from public corpora and incorporate 630,000 mined utterances. The resulting corpus of approximately 800,000 utterances is used to train a single 1.7B-parameter model that transcribes speech with inline tags for six NVV classes. 段落4(结果): Across four human-annotated benchmarks, Xing-Audio-NVVASR ranks first among nine models in both composite score and six-class Macro-F1. It achieves a mean composite score of 64.46, surpassing the strongest comparison model, StepAudio 3 Chat Preview, by 6.03 points, and exceeds the annotation teacher's breathing F1 by 14.40 to 31.57 percentage points. 段落5(计划): We plan to release the model weights, inference code, four evaluation benchmarks, annotation guidelines, and evaluation code. 这样拆分合理,且没有增加任何文字。重要句子用``标注。 检查规则: 1. 只输出HTML片段,无Markdown围栏,无解释。✓ 2. 每个自然段落用`

    `包裹,长文用`

    `/`

    `划分小节。这里没有添加标题,因为原文无小节,避免增内容。但为了体现"长文用

    /

    划分小节",我是否应该添加?如果添加,就增加了原文没有的文字,违反规则4。所以不添加是更稳妥的。或者,我可以不拆分段落,整个作为一段。但拆分段落更符合"每个自然段落用

    包裹"。 实际上,原文就是一段,没有换行。我如果拆分成多段,算不算"改内容"?排版上的段落划分通常不算改内容。但为了绝对安全,我也可以保持为一个完整的`

    `。不过,作为一个"排版工程师",合理分段是专业体现。我决定拆分成逻辑段落,因为原文在语义上明显有多个层次。 再考虑是否加`

    `。规则说"长文用

    /

    划分小节",如果原文没有小节标题,强行添加就是增内容。所以不加。 最终HTML:

    Joint recognition of speech and non-verbal vocalizations (NVVs) is hindered by scarce event-rich training data and inaccurate or missing event labels. We present Xing-Audio-NVVASR, an end-to-end recognizer trained through progressive data mining and iterative supervision refinement.

    Guided by unified acoustic criteria, two heterogeneous models cross-audit annotations, flagging disagreements for teacher reannotation. To uncover shared omissions that model agreement may conceal, complementary breath detectors screen the entire corpus for human review. Human-confirmed labels accumulate across rounds and are preserved during subsequent refinement.

    Over five rounds, we refine annotations for 170,000 utterances from public corpora and incorporate 630,000 mined utterances. The resulting corpus of approximately 800,000 utterance