Common Phone: A Multilingual Dataset for Robust Acoustic Modelling

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 763–768 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 763 Common Phone: A Multilingual Dataset for Robust Acoustic Modelling Philipp Klumpp1, Tomás Arias-Vergara1,2, Paula-Andrea Pérez-Toro1,2 Elmar Nöth1, Juan Rafael Orozco-Arroyave1,2 1Friedrich-Alexander-Universität Erlangen-Nürnberg, 2Universidad de Antioquia 1Lehrstuhl für Mustererkennung, Martensstrasse 3, 91058 Erlangen, Germany 2Calle 67 # 53-108, UdeA Campus Principal, Medellín, Colombia philipp.klumpp@fau.de Abstract Current state of the art acoustic models can easily comprise more than 100 million parameters. This growing complexity demands larger training datasets to maintain a decent generalization of the final decision function. An ideal dataset is not necessarily large in size, but large with respect to the amount of unique speakers, utilized hardware and varying recording conditions. This enables a machine learning model to explore as much of the domain-specific input space as possible during parameter estimation. This work introduces Common Phone, a gender-balanced, multilingual corpus recorded from more than 11,000 contributors of Mozilla’s Common Voice project. It comprises around 116 hours of speech enriched with automatically generated phonetic segmentation. A Wav2Vec 2.0 acoustic model was trained with the Common Phone to perform phonetic symbol recognition and validate the quality of the generated phonetic annotation. The architecture achieved a PER of 18.1 % on the entire test set, computed with all 101 unique phonetic symbols, showing slight differences between the individual languages. We conclude that Common Phone provides sufficient variability and reliable phonetic annotation to help bridging the gap between research and application of acoustic models. Keywords: Speech Dataset, Acoustic Modelling, Multilingual Corpus 1. Introduction In the past two years, Wav2Vec (Schneider et al., 2019) and Wav2Vec 2.0 (Baevski et al., 2020) have leveraged the state-of-the-art of acoustic models to a new level, with the latter being able to achieve a Phoneme Error Rate (PER) of 8.3 % on the TIMIT (Garofolo et al., 1993) test set. TIMIT is among the most important cor- pora for acoustic model evaluation, specifically because of the precise, manually annotated, phonetic reference. All speakers had been recorded in equal acoustic con- ditions using the same microphone. For acoustic model validation, these conditions can be considered perfect. However, the constricted acoustic environment also im- plies limited robustness to altered recording conditions. The corpus is alsonot gender-balanced, with 439( 70 %) of the 630 speakers being male. For any application- driven project, the total number of speakers is quite small as well for modern standards. A large variability of training samples is key to enable a deep architecture to explore as much of the input space as possible. Any perturbation in said space could re- sult in the model performance to collapse (Wang et al., 2018). Deviations from the known input-space could be so small that it would be impossible for a human rater to perceive them, thus they might be exploited in scenarios like adversarial attacks (Schönherr et al., 2018; Hu et al., 2019). Robustness could be improved during train- ing through techniques …

Content truncated to 500 words