Designing a Speech Corpus for the Development and Evaluation of Dictation Systems in Latvian

Designing a Speech Corpus for the Development and Evaluation of Dictation Systems in Latvian Mārcis Pinnis1, Askars Salimbajevs1, Ilze Auziņa2 1Tilde, Vienibas gatve 75a, Riga, Latvia 2Institute of Mathematics and Computer Science, University of Latvia, 29 Raina Blvd., Riga, Latvia E-mail: marcis.pinnis@tilde.lv, askars.salimbajevs@tilde.lv, ilze.auzina@lumii.lv Abstract In this paper the authors present a speech corpus designed and created for the development and evaluation of dictation systems in Latvian. The corpus consists of over nine hours of orthographically annotated speech from 30 different speakers. The corpus features spoken commands that are common for dictation systems for text editors. The corpus is evaluated in an automatic speech recognition scenario. Evaluation results in an ASR dictation scenario show that the addition of the corpus to the acoustic model training data in combination with language model adaptation allows to decrease the WER by up to relative 41.36% (or 16.83% in absolute numbers) compared to a baseline system without language model adaptation. Contribution of acoustic data augmentation is at relative 12.57% (or 3.43% absolute). Keywords: speech recognition corpus, dictation systems, Latvian 1. Introduction Automatic speech recognition (ASR) technologies for Latvian have a relatively short history because even three years ago (i.e., in 2013 and before) there was no orthographically annotated speech corpus, which could be used for ASR purposes, available. However, there have been attempts to develop ASR systems for broadcast speech recognition (Oparin et al, 2013) in the Quaero project (Lamel, 2012) using acoustic model bootstrapping. Since the creation of the first orthographically and phonetically annotated speech corpus for Latvian (Pinnis et al., 2014) – the Latvian Speech Recognition Corpus (LSRC), ASR technologies for Latvian have been actively researched (e.g., Salimbajevs & Pinnis, 2014, Salimbajevs & Strigins, 2015b), and different application scenarios that have resulted in practical applications with ASR capabilities (e.g., Vīra & Vasiļjevs, 2014; Salimbajevs & Strigins, 2015a; Znotiņš & Dargis, 2014; Znotiņš et. al., 2015) have been investigated. However, the technology has yet to reach a level where it is applicable in dictation scenarios in text editors. For the development of dictation systems, it was necessary to create a specific corpus that would: 1) better capture the speaking characteristics of speakers when dictating text to a computer and 2) contain spoken commands common to dictation scenarios (e.g., punctuation, formatting, special symbol, and action commands). Therefore, in this paper, we present the Dictated Speech Corpus (DSC) that has been created to address these requirements. Speech corpora creation with spoken commands and speech recognition system development with spoken command support has been investigated also in related research. For instance, Paul and Baker (1992) and Bernstein and Danielson (1992) created speech corpora for English continuous speech recognition with both verbalised and non-verbalised punctuation marks. Digalakis et al. (2003) created a speech recognition system for Greek that could handle special symbol and formatting commands, however they were introduced only in the speech recognition system and were omitted from their speech corpus. Enravi (2012) in his thesis stressed the importance of spoken command support in dictation systems. …

Content truncated to 500 words