arXiv cs.CL· Artem Saloev, Nicolas Ballier, Erin Pacquetet·· 12 小时前AI 评分18
研究 Mimi 编解码器首个码本（CB0）的音素表征
Probing the phon* representations of the Mimi codebooks
AI 导读
研究用强制对齐、熵分析和 Phone-Normalized Mutual Information（PNMI）考察 Mimi 编解码器首个码本（CB0）的音素表征，因 Mimi 固定 80 ms 分词窗口常跨多个音素，ABX 评测并不适用。
正文
Abstract:Neural audio codecs are increasingly used to encode speech for integration with large language models. This paper investigates the phonetic representations in the first codebook (CB0) of the Mimi codec. ABX style evaluation is poorly suited to Mimi fixed 80 ms tokenization window, which often spans multiple phonemes; we instead use forced alignment, entropy analysis, and Phone-Normalized Mutual Information (PNMI). Using TIMIT speech (58/55 speakers, SA1/SA2), individual tokens are strongly associated with specific phonemes (PNMI = 0.84, above random), yet phonemes are realized by multiple competing tokens rather than a one-to-one mapping.
| Comments: | Accepted at the 9th International Conference on Natural Language and Speech Processing (ICNLSP-2026) this https URL |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.35432 [cs.CL] |
| (or arXiv:2610.35432v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.35432 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Nicolas Ballier [view email]
[v1]
Mon, 26 Oct 2026 13:36:53 UTC (391 KB)
来源：arXiv cs.CL · arxiv.org