An Application for Building a Polish Telephone Speech Corpus

An Application for Building a Polish Telephone Speech Corpus Bartosz Zi´ołko1,2, Piotr ˙Zelasko1, Ireneusz Gawlik1,3, Tomasz Pe¸dzima¸˙z1,2, Tomasz Jadczyk2 1AGH University of Science and Technology, Poland, Department of Computer Science, Electronics and Telecommunications, al. Mickiewicza 30, Krak´ow, Poland, www.dsp.agh.edu.pl 2Techmo, Krak´ow, Poland, techmo.pl 3Grupa Allegro Sp. z o.o., Poland bartosz.ziolko@techmo.pl, pzelasko@agh.edu.pl, igawlik@live.com, {pedzimaz, jadczyk}@techmo.pl Abstract The paper presents our approach towards building a tool for speech corpus collection of a specific domain content. We describe our iterative approach to the development of this tool, with focus on the most problematic issues at each working stage. Our latest version synchronizes VoIP call management and recording with a web application providing content. The tool was already used and applied for Polish to gather 63 hours of automatically annotated recordings across several domains. Amongst them, we obtained a continuous speech corpus designed with an emphasis on optimal phonetic diversification in relation to the phonetically balanced National Corpus of Polish. We evaluate the usefulness of this data against the GlobalPhone corpus in the task of training an acoustic model for a telephone speech ASR system and show that the model trained on our balanced corpus achieves significantly lower WER in two grammar-based speech recognition tasks - street names and public transport routes numbers. Keywords: speech recognition, language resources, corpus recording tools 1. Introduction Speech corpora acquisition is a significant problem in voice applications development, however, its magnitude is often under-appreciated. In this paper, we address this problem by exploring different means of how speech recordings can be collected. The process we describe is iterative - as we will show, at each iteration the previous approach is scruti- nised and improved upon, which allows for acquisition of the recordings on a larger scale than possible before, while improving user experience, and thus the quality of the data collected. Specifically, we describe a tool designed to improve our Automatic Speech Recognition (ASR) system perfor- mance, specialised for conversations over telephone - SAR- MATA (Zi´ołko et al., 2015). To that end, we focus on im- proving the training corpus by collecting more and more data, as Polish is still quite limited in the availability of speech corpora regarding their sizes and diversity. Our main assumptions were: • training data for telephone ASR should be recorded through a telecommunication channel in order to con- dition the ASR on the channel effects met in a produc- tion environment, • speech content should be as diversified as possible, • it is faster and more cost efficient to produce speech for a given transcription than to annotate existing recordings. Our ASR is based on a Deep Neural Network (DNN) ar- chitecture (Zi´ołko et al., 2015). At the moment, we are able to train the DNN with approx. 4 million parameters on as little as around 100 hours of speech. However, the de facto standard in the industry is to use sizeable training corpora, which total durations are measured in hundreds or even thousands of hours (Amodei and others, Decem- ber 2015; …

Content truncated to 500 words