bian, xuefei (2025) Improving OOV Word Recognition in End-to-End ASR via Lightweight Domain Adaptation with a TED-LIUM2 N-gram Language Model. Master thesis, Voice Technology (VT).
|
Text
MA_5836670_X_Bian.pdf Download (276kB) | Preview |
Abstract
Automatic Speech Recognition (ASR) systems have achieved strong performance on read and clean speech. However, ASR systems still face major challenges when recognizing spontaneous speech, especially for low frequency or domain specific out of vocabulary (OOV) words. Most current research focuses on improving overall recognition accuracy in non-spontaneous speech, but few studies explore whether it is possible to improve OOV recognition and overall performance in spontaneous speech without retraining the acoustic model. This study aims to fill this gap. I trained an n-gram language model (LM) using a small amount of spontaneous speech data from TED-LIUM2 and applied it to a pre trained Wav2Vec2.0 acoustic model using shallow fusion. This study tested three settings: no LM, a TED LIUM2 trained LM, and a pretrained LM based on LibriSpeech. I evaluated the performance in terms of Word Error Rate (WER), OOV recall, and three types of recognition errors: substitution, insertion, and deletion. Results show that the TED LIUM2 LM increased the OOV recall rate by 21.75% compared to the no LM baseline, confirming its ability to help recognize domain specific words. However, this LM also caused a increase in overall WER from 25.04% to 44.28%, mainly due to substitution errors caused by incomplete sentence structure, redundant content, and missing context in the training data. In contrast, the LibriSpeech LM achieved a better balance between WER (21.22%) and OOV recall (40.41%). This work contributes by providing the first systematic evaluation of using small scale spontaneous speech to train a language model for ASR without changing the acoustic model. It also highlights how the quality and structure of training text (such as context and sentence completeness) play a key role in ASR performance.
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Do, T.P. |
| Date Deposited: | 03 Sep 2025 14:48 |
| Last Modified: | 03 Sep 2025 14:48 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/670 |
Actions (login required)
 |
View Item |