Unsupervised Language Model Adaptation for Speech Recognition with no Extra Resources

Unsupervised Language Model Adaptation for Speech Recognition with no Extra Resources

Stať ve sborníku mimo WoS a Scopus

Classically, automatic speech recognition (ASR) modelsare decomposed into acoustic models and language models(LM). LMs usually exploit the linguistic structure ona purely textual level and usually contribute strongly toan ASR systems performance. LMs are estimated onlarge amounts of textual data covering the target domain.However, most utterances cover more specic topics, e.g.inuencing the vocabulary used. Therefore, it's desirableto have the LM adjusted to an utterance's topic. Previouswork achieves this by crawling extra data from theweb or by using signicant amounts of previous speechdata to train topic-specic LM on. We propose a wayof adapting the LM directly using the target utteranceto be recognized. The corresponding adaptation needsto be done in an unsupervised or automatically supervisedway based on the speech input. To deal withcorresponding errors robustly, we employ topic encodingsfrom the recently proposed Subspace MultinomialModel. This model also avoids any need of explicit topiclabelling during training or recognition, making the proposedmethod straight-forward to use. We demonstratethe performance of the method on the Librispeech corpus,which consists of read ction books, and we discussit's behaviour qualitatively.

Classically, automatic speech recognition (ASR) modelsare decomposed into acoustic models and language models(LM). LMs usually exploit the linguistic structure ona purely textual level and usually contribute strongly toan ASR systems performance. LMs are estimated onlarge amounts of textual data covering the target domain.However, most utterances cover more specic topics, e.g.inuencing the vocabulary used. Therefore, it's desirableto have the LM adjusted to an utterance's topic. Previouswork achieves this by crawling extra data from theweb or by using signicant amounts of previous speechdata to train topic-specic LM on. We propose a wayof adapting the LM directly using the target utteranceto be recognized. The corresponding adaptation needsto be done in an unsupervised or automatically supervisedway based on the speech input. To deal withcorresponding errors robustly, we employ topic encodingsfrom the recently proposed Subspace MultinomialModel. This model also avoids any need of explicit topiclabelling during training or recognition, making the proposedmethod straight-forward to use. We demonstratethe performance of the method on the Librispeech corpus,which consists of read ction books, and we discussit's behaviour qualitatively.

speech recognition

speech recognition

BENEŠ, K.; IRIE, K.; BECK, E.; SCHLÜTER, R.; NEY, H.

2020

18.03.2019

DEGA Head office, Deutsche Gesellschaft für Akustik

Rostock

978-3-939296-14-0

Proceedings of DAGA 2019

954

957

4

https://www.dega-akustik.de/publikationen/online-proceedings/

benes_DAGA_2019