Using Automatic Speech Recognition in Spoken Corpus Curation

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6423–6428 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6423 Using Automatic Speech Recognition in Spoken Corpus Curation Jan Gorisch,1 Michael Gref,2 Thomas Schmidt1 1Leibniz-Institute for the German Language (IDS), Germany 2Fraunhofer Institute for Intelligent Analysis and Information Systems (IAIS), Germany gorisch@ids-mannheim.de, michael.gref@iais.fraunhofer.de, thomas.schmidt@ids-mannheim.de Abstract The newest generation of speech technology caused a huge increase of audio-visual data nowadays being enhanced with orthographic transcripts such as in automatic subtitling in online platforms. Research data centers and archives contain a range of new and historical data, which are currently only partially transcribed and therefore only partially accessible for systematic querying. Automatic Speech Recognition (ASR) is one option of making that data accessible. This paper tests the usability of a state-of-the-art ASR-System on a historical (from the 1960s), but regionally balanced corpus of spoken German, and a relatively new corpus (from 2012) recorded in a narrow area. We observed a regional bias of the ASR-System with higher recognition scores for the north of Germany vs. lower scores for the south. A detailed analysis of the narrow region data revealed – despite relatively high ASR-confidence – some specific word errors due to a lack of regional adaptation. These findings need to be considered in decisions on further data processing and the curation of corpora, e.g. correcting transcripts or transcribing from scratch. Such geography-dependent analyses can also have the potential for ASR-development to make targeted data selection for training/adaptation and to increase the sensitivity towards varieties of pluricentric languages. Keywords: oral corpora, automatic transcription, ASR, corpus curation, pluricentric, spoken German, Ripuarian 1. Introduction The Archive for Spoken German (AGD, Stift and Schmidt (2014), http://agd.ids-mannheim.de) is part of the CLARIN-D Centre IDS and specializes in data of spoken German, mostly corpora of natural interaction and data on varieties of German. The AGD develops and curates such corpora and makes them available to the scientific commu- nity. Besides detailed metadata on recordings and speakers, the key to accessibility are transcripts of the recordings al- lowing systematic queries and the application of other au- tomated methods. So far, the high cost of manual transcrip- tion results in large parts of the corpora remaining untran- scribed and therefore accessible only to a limited degree. Automatic Speech Recognition (ASR) is often claimed to be a way through that transcription bottleneck. The present paper describes some first steps in exploring whether or not that claim can be fulfilled and if so, for which type of data ASR is suitable and what factors influence the quality of ASR results. The latest developments in the field of ASR – artificial neu- ral networks, deep learning, and the intorduction of LF- MMI models (Povey et al., 2016) – have heaved speech technology from a level that was merely useful for limited vocabulary and clean audio tasks to a level where it could be applied to various recording conditions and large vocab- ulary challenges as can …

Content truncated to 500 words