2026-04-20T22:51:19Z https://ipsj.ixsq.nii.ac.jp/oai

oai:ipsj.ixsq.nii.ac.jp:00209762 2025-01-19T18:23:58Z 1164:5159:10515:10530

Comparison of End-to-End Models for Joint Speaker and Speech Recognition Comparison of End-to-End Models for Joint Speaker and Speech Recognition Kak, Soky Sheng, Li Masato, Mimura Chenhui, Chu Tatsuya, Kawahara Kak, Soky Sheng, Li Masato, Mimura Chenhui, Chu Tatsuya, Kawahara SP1 In this paper, we investigate the effectiveness of using speaker information on the performance of speaker- imbalanced automatic speech recognition (ASR). We identify major speakers and combine other speakers who have a small size of speech, and make a systematic comparison of three methods that use speaker information for ASR including speaker attribute augmentation (SAug), multi-task learning (MTL), and adversarial learning (AL). We conduct experiments on a large spontaneous speech corpus of the Extraordinary Chambers in the Courts of Cambodia (ECCC) and an open Khmer text-to-speech corpus. As a result, we find that the use of speaker clustering information improves ASR performance including new speakers. Moreover, AL achieves better performance and more robustness in the speaker-independent setting compared to the other methods. It reduces errors of the baseline model by 4.32%, 5.46%, and 16.10% for the closed test, open test, and out-of-domain test, respectively. In this paper, we investigate the effectiveness of using speaker information on the performance of speaker- imbalanced automatic speech recognition (ASR). We identify major speakers and combine other speakers who have a small size of speech, and make a systematic comparison of three methods that use speaker information for ASR including speaker attribute augmentation (SAug), multi-task learning (MTL), and adversarial learning (AL). We conduct experiments on a large spontaneous speech corpus of the Extraordinary Chambers in the Courts of Cambodia (ECCC) and an open Khmer text-to-speech corpus. As a result, we find that the use of speaker clustering information improves ASR performance including new speakers. Moreover, AL achieves better performance and more robustness in the speaker-independent setting compared to the other methods. It reduces errors of the baseline model by 4.32%, 5.46%, and 16.10% for the closed test, open test, and out-of-domain test, respectively. technical report 情報処理学会 2021-02-24 application/pdf 研究報告音声言語情報処理（SLP） 24 2021-SLP-136 1 5 2188-8663 AN10442647 https://ipsj.ixsq.nii.ac.jp/record/209762/files/IPSJ-SLP21136024.pdf eng