Creating Lithuanian and Latvian Speech Corpora from Inaccurately Annotated Web Data

Creating Lithuanian and Latvian Speech Corpora from Inaccurately Annotated Web Data Askars Salimbajevs Tilde, Vienibas gatve 75A, Riga, Latvia University of Latvia, Raina bulv 19, Riga, Latvia askars.salimbajevs@tilde.lv Abstract This paper describes the method that was used to produce additional acoustic model training data for the less-resourced languages of Lithuanian and Latvian. The method uses existing baseline speech recognition systems for Latvian and Lithuanian to align audio data from the Web with imprecise non-normalised transcripts. From 690 hours of Web data (300h for Latvian, 390h for Lithuanian), we have created additional 378 hours of training data (186h for Latvian and 192 for Lithuanian). Combining this additional data with baseline training data allowed to significantly improve word error rate for Lithuanian from 40% to 23%. Word error rate for the Latvian system was improved from 19% to 17%. Keywords: speech recognition, web data, alignment, Lithuanian, Latvian, speech corpora, low-resourced languages 1. Introduction Training of an acoustic model for an automatic speech recognition (ASR) system requires large amounts of tran- scribed audio, especially for the state-of-the art deep neural network models. For Low Resource Languages (LRLs), where there may only be a few hours of transcribed au- dio, this is a very serious problem. For Latvian and Lithua- nian languages the situation is better, for example, there are a 100h Lithuanian speech corpus created for the LIEPA project1, an 84h corpus from BMMG(Alum¨ae and Tilk, 2016) and a 100h Latvian speech corpus(Pinnis et al., 2014). However, even this is not much, as neural network acoustic models can make use of many thousands of hours of data and achieve significant improvement in recognition accuracy. In recent years, improvements in data storage and network- ing technology have made it feasible to provide Internet users with access to large amounts of multimedia content. This content can be automatically collected and processed for the purpose of training statistical models. However, in many cases, this content is not structured or organised in an accurate and machine-readable form. For example, both the Lithuanian Parliament (Seimas) and Latvian Parliament (Saeima) websites contain a large archive of video recordings of parliamentary sessions and edited transcripts. One may want to collect this data and use it for the development of a general purpose speech recogni- tion system. This data can also be used for adapting exist- ing ASR for transcription of Seimas sessions. Edited tran- scripts can be used as the training data for the language model, and video recordings can be used for adapting the acoustic model. Unfortunately, edited transcripts do not have any timing in- formation. Also, these transcripts are non-normalised. This means that numbers, dates, percent signs, etc. are written 1LIEPA (Services Controlled by Lithuanian Voice), Vilnius University, LEU, Siauliai University, Institute of Lithuanian lan- guage. https://www.xn–ratija-ckb.lt/liepa with digits and symbols, not as words. Converting these tokens back to words is complicated and error-prone for in- flected languages like Latvian and Lithuanian. While speaking humans can make grammar mistakes, re- peat words, make corrections and restart whole sentences. …

Content truncated to 500 words