首页 > AI前沿 > Asymmetric Correspondence Learning for Lexical and Style Embeddings: Interpreting Latent Space via Acoustic Features

Asymmetric Correspondence Learning for Lexical and Style Embeddings: Interpreting Latent Space via Acoustic Features

arXiv自然语言 2026-11-06 19:04 19 阅读 查看原文
用户提供了一段新闻/学术文本,要求将其排版为语义化的HTML片段。 规则: 1. 只输出HTML片段本身,不要Markdown围栏,不要解释。 2. 每个自然段落用`

`包裹,长文用`

`/`

`划分小节(虽然这段文本不长,可以只用`

`,或者加上一个小标题)。 3. 列表用`

  • `,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。
    
    分析正文内容:
    "Self-supervised speech embeddings jointly encode lexical, speaker, and prosodic information, complicating feature isolation. We introduce Asymmetric Correspondence Learning (ACL), distilling HuBERT features into separate Acoustic Word Embeddings (AWEs) and Acoustic Style Embeddings (ASEs) via a split-latent correspondence bottleneck. Through probing and clustering evaluations on single-word utterances, we demonstrate that substantial variance from extrinsic sources remains encoded within each latent space, despite linear probing indicating the information has been stripped. Whilst the AWE space exhibits cohesive lexical clustering (AP$\approx$0.90, Fisher$\approx$6.28), linear probing on the ASE space achieves 97.9\% speaker accuracy across a diffuse topology (Fisher$\approx$1.01), yet preserves substantial lexical variance (Fisher$\approx$0.86) that remains linearly unrecoverable (0.17\%). To explain these findings, we apply group-level permutation feature importance over eGeMAPS acoustic features, finding that AWEs partition by vocal tract filters and ASEs by $F_0$ dynamics. These findings suggest that standard diagnostic benchmarks may be insufficient for verifying feature isolation, as substantial speech structure can be retained even when it remains entirely hidden from probes."
    
    这其实是一段学术论文的摘要。
    我可以把它分成几个段落,或者就一个段落。为了符合“长文用h2/h3”的规则,我可以加一个h2作为标题,比如“Abstract”或者直接不适用h2/h3,只用p。不过题目说“长文用”,这段不算长,可以直接用p。或者把句子稍微拆分成几个p?不,规则说“每个自然段落用

    包裹”,原文是一个自然段。 等等,原文中有一些数学公式和特殊字符,如 `AP$\approx$0.90`, `Fisher$\approx$6.28`, `97.9\%`, `$F_0$`, `0.17\%`。这些需要保留在HTML中。 重要句子可以使用 ``。比如最后一句结论可以使用 ``。 让我们看一下原文结构: 1. 背景与提出方法:Self-supervised speech embeddings ... through probing and clustering evaluations on single-word utterances ... stripped. 2. 实验结果:Whilst the AWE space ... unrecoverable (0.17\%). 3. 分析与结论:To explain these findings ... hidden from probes. 如果保持原本的一个自然段,就只用一个`

    `。 或者拆成三个`

    `:

    Self-supervised speech embeddings jointly encode lexical, speaker, and prosodic information, complicating feature isolation. We introduce Asymmetric Correspondence Learning (ACL), distilling HuBERT features into separate Acoustic Word Embeddings (AWEs) and Acoustic Style Embeddings (ASEs) via a split-latent correspondence bottleneck. Through probing and clustering evaluations on single-word utterances, we demonstrate that substantial variance from extrinsic sources remains encoded within each latent space, despite linear probing indicating the information has been stripped.

    Whilst the AWE space exhibits cohesive lexical clustering (AP$\approx$0.90, Fisher$\approx$6.28), linear probing on the ASE space achieves 97.9\% speaker accuracy across a diffuse topology (Fisher$\approx$1.01), yet preserves substantial lexical variance (Fisher$\approx$0.86) that remains linearly unrecoverable (0.17\%).

    To explain these findings, we apply group-level permutation feature importance over eGeMAPS acoustic features, finding that AWEs partition by vocal tract filters and ASEs by $F_0$ dynamics. These findings suggest that standard diagnostic benchmarks may be insufficient for verifying feature isolation, as substantial speech structure can be retained even when it remains entirely hidden from probes.

    检查特殊字符: `AP$\approx$0.90` -> 在HTML中可以直接保留,或者转义?题目说“不得增删改内容”,所以直接保留原字符串。 `97.9\%` -> 保留 `97.9\%` `$F_0$` -> 保留 `$F_0$` 检查是否有需要列表或引用的地方?没有。 代码块?没有。 所以最终输出应该是三个 `

    ` 标签包裹的内容,最后一句用 `` 包裹。 再读一遍规则: 1.只输出HTML片段本身,不要Markdown围栏、不要解释; 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。
      
      原文是一个自然段,如果强行拆分算不算“增删改内容”?通常按句群拆分段落是可以的,但最安全的是保持原有的一个自然段落,只用一个`

      `。 不过,作为排版工程师,把长摘要拆成几个语义段落是常见的排版工作。我就拆成3个`

      `吧,或者就1个`

      `。 为了严格遵守“不得增删改内容”,改变段落划分可能被视为修改了文本结构。所以只用1个`

      `是最稳妥的。 或者加一个`

      `?原文没有标题,加标题属于“增”。所以不加标题。 那就直接:

      Self-supervised speech embeddings jointly encode lexical, speaker, and prosodic information, complicating feature isolation. We introduce Asymmetric Correspondence Learning (ACL), distilling HuBERT features into separate Acoustic Word Embeddings (AWEs) and Acoustic Style Embeddings (ASEs) via a split-latent correspondence bottleneck. Through probing and clustering evaluations on single-word utterances, we demonstrate that substantial variance from extrinsic sources remains encoded within each latent space, despite linear probing indicating the information has been stripped. Whilst the AWE space exhibits cohesive lexical clustering (AP$\approx$0.90, Fisher$\approx$6.28), linear probing on the ASE space achieves 97.9\% speaker accuracy across a diffuse topology (Fisher$\approx$1.01), yet preserves substantial lexical variance (Fisher$\approx$0.86) that remains linearly unrecoverable (0.17\%). To explain these findings, we apply group-level permutation feature importance over eGeMAPS acoustic features, finding that AWEs partition by vocal tract filters and ASEs by $F_0$ dynamics. These findings suggest that standard diagnostic benchmarks may be insufficient for verifying feature isolation, as substantial speech structure can be retained even when it remains entirely hidden from probes.

      等等,如果用户期望看到排版工程师的“语义化”能力,我把整个文本放在一个`

      `里,里面用``突出重点,可能略显单薄。但根据规则4“保留原文所有信息,不得增删改内容”,拆分段落如果不改变单词,只是改变`

      `标签的位置,算不算增删改?通常不算,因为HTML标签本身就是排版的一部分。 我还是分成三个`

      `吧,逻辑更清晰: 段落1:背景+方法+初步结论 段落2:具体实验数据 段落3:分析+最终结论

      Self-supervised speech embeddings jointly encode lexical, speaker, and