Conversational Speech Recognition Needs Data? Experiments with Austrian German

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 4684–4691 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 4684 Conversational Speech Recognition Needs Data? Experiments with Austrian German Julian Linke1, Philip N. Garner2, Gernot Kubin1, Barbara Schuppler1 1Signal Processing and Speech Communication Laboratory, Graz University of Technology 2Idiap Research Institute 1Inffeldgasse 16c, A-8010 Graz, Austria 2Rue Marconi 19, 1920 Martigny, Switzerland 1{linke, kubin, b.schuppler}@tugraz.at, 2phil.garner@idiap.ch Abstract Conversational speech represents one of the most complex of automatic speech recognition (ASR) tasks owing to the high inter-speaker variation in both pronunciation and conversational dynamics. Such complexity is particularly sensitive to low-resourced (LR) scenarios. Recent developments in self-supervision have allowed such scenarios to take advantage of large amounts of otherwise unrelated data. In this study, we characterise an (LR) Austrian German conversational task. We begin with a non-pre-trained baseline and show that fine-tuning of a model pre-trained using self-supervision leads to improvements consistent with those in the literature; this extends to cases where a lexicon and language model are included. We also show that the advantage of pre-training indeed arises from the larger database rather than the self-supervision. Further, by use of a leave-one-conversation out technique, we demonstrate that robustness problems remain with respect to inter-speaker and inter-conversation variation. This serves to guide where future research might best be focused in light of the current state-of-the-art. Keywords: Speech Recognition, Conversational Speech, Austrian German, Low-Resource, Wav2vec2.0, Kaldi 1. Introduction Solving automatic speech recognition (ASR) tasks for conversational speech is crucial especially for social robots interacting with humans or automatic transcrip- tions of multimedia meetings (Popescu-Belis et al., 2012). Two humans who interact spontaneously with each other introduce complex inter- and intra-speaker variation depending on for instance the speaker’s atti- tude towards the listener and the speaking task (Wright, 2006). Especially casual face-to-face conversations are characterized by a large amount of speaker-dependent pronunciation variation, by disfluencies, and by broken words or incomplete utterance structures. The resulting high degree of variation on all linguistic levels affects the acoustic model, the lexicon and language model of an ASR system. Given the high variation in spontaneous conversations, the amount of annotated training data needed for ASR experiments to enable generalization for an unseen test set can sometimes be misleading in the sense that avoiding the data sparseness problem appears not to be possible, especially in case of spontaneous speech (Furui et al., 2005; Furui, 2009). Such studies give in- sights into the relationship between data size for acous- tic model training and WER in case of Japanese spon- taneous speech recognition: Utilizing 1/8 of available data (63.75h) for acoustic model training results in a WER of approx. 27% whereby training with the entire data (510h) gives an improvement of approx. 2%, but still no convergence. In this paper, we deal with conversational speech from the “Graz corpus of Read And Spontaneous Speech” (GRASS) (Schuppler et al., 2014a), which contains about 19h (or 19 conversations) of Austrian German conversations introducing …

Content truncated to 500 words