Comparison of Language Models Trained on Written Texts and Speech Transcripts in the Context of Automatic Speech Recognition

Abstract

Abstract-We investigate whether language models used in automatic speech recognition (ASR) should be trained on speech transcripts rather than on written texts. By calculating log-likelihood statistic for part-of-speech (POS) n-grams, we show that there are significant differences between written texts and speech transcripts. We also test the performance of language models trained on speech transcripts and written texts in ASR and show that using the former results in greater word error reduction rates (WERR), even if the model is trained on much smaller corpora. For our experiments we used the manually labeled one million subcorpus of the National Corpus of Polish and an HTK acoustic model. Index Terms-automatic speech recognition, morphosyntactic language model, written and spoken language comparison

Statistical analysis of orthographic and phonemic language corpus for word-based and phoneme-based Polish language modelling

“…This article contains a general statistics of Polish language that can be useful for a variety of language and speech processing applications, including automatic speech recognition with language models [55].…”

First Automatic Fongbe Continuous Speech Recognition System: Development of Acoustic Models and Language Models

“…they describe relations between words (or other tokens), thus enabling to choose most probable sequences. This proves to be especially useful in speech recognition, where acoustical models usually produce a number of hypotheses, and reranking them according to a language model can substantially improve recognition rates [10] To compare the performance of our Fongbe recognition system, we built two language models (LM) using the same text corpus. The first language model (LM1) is built with the original texts after normalization and contain different tonal vowels.…”