Can LLMs Help Localize Fake Words in Partially Fake Speech?
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866917333482602496 |
|---|---|
| author | Zhang, Lin Thebaud, Thomas Cai, Zexin Khudanpur, Sanjeev Povey, Daniel García-Perera, Leibny Paola Wiesner, Matthew Andrews, Nicholas |
| author_facet | Zhang, Lin Thebaud, Thomas Cai, Zexin Khudanpur, Sanjeev Povey, Daniel García-Perera, Leibny Paola Wiesner, Matthew Andrews, Nicholas |
| contents | Large language models (LLMs), trained on large-scale text, have recently attracted significant attention for their strong performance across many tasks. Motivated by this, we investigate whether a text-trained LLM can help localize fake words in partially fake speech, where only specific words within a speech are edited. We build a speech LLM to perform fake word localization via next token prediction. Experiments and analyses on AV-Deepfake1M and PartialEdit indicates that the model frequently leverages editing-style pattern learned from the training data, particularly word-level polarity substitutions for those two databases we discussed, as cues for localizing fake words. Although such particular patterns provide useful information in an in-domain scenario, how to avoid over-reliance on such particular pattern and improve generalization to unseen editing styles remains an open question. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_11205 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Can LLMs Help Localize Fake Words in Partially Fake Speech? Zhang, Lin Thebaud, Thomas Cai, Zexin Khudanpur, Sanjeev Povey, Daniel García-Perera, Leibny Paola Wiesner, Matthew Andrews, Nicholas Audio and Speech Processing Sound Large language models (LLMs), trained on large-scale text, have recently attracted significant attention for their strong performance across many tasks. Motivated by this, we investigate whether a text-trained LLM can help localize fake words in partially fake speech, where only specific words within a speech are edited. We build a speech LLM to perform fake word localization via next token prediction. Experiments and analyses on AV-Deepfake1M and PartialEdit indicates that the model frequently leverages editing-style pattern learned from the training data, particularly word-level polarity substitutions for those two databases we discussed, as cues for localizing fake words. Although such particular patterns provide useful information in an in-domain scenario, how to avoid over-reliance on such particular pattern and improve generalization to unseen editing styles remains an open question. |
| title | Can LLMs Help Localize Fake Words in Partially Fake Speech? |
| topic | Audio and Speech Processing Sound |
| url | https://arxiv.org/abs/2603.11205 |