Paper ID | SPE-55.6 | ||
Paper Title | How Phonotactics Affect Multilingual and Zero-shot ASR Performance | ||
Authors | Siyuan Feng, Delft University of Technology, Netherlands; Piotr Żelasko, Laureano Moro-Velázquez, Johns Hopkins University, United States; Ali Abavisani, Mark Hasegawa-Johnson, University of Illinois at Urbana-Champaign, United States; Odette Scharenborg, Delft University of Technology, Netherlands; Najim Dehak, Johns Hopkins University, United States | ||
Session | SPE-55: Language Identification and Low Resource Speech Recognition | ||
Location | Gather.Town | ||
Session Time: | Friday, 11 June, 14:00 - 14:45 | ||
Presentation Time: | Friday, 11 June, 14:00 - 14:45 | ||
Presentation | Poster | ||
Topic | Speech Processing: [SPE-MULT] Multilingual Recognition and Identification | ||
IEEE Xplore Open Preview | Click here to view in IEEE Xplore | ||
Abstract | The idea of combining multiple languages' recordings to train a single automatic speech recognition (ASR) model brings the promise of the emergence of universal speech representation. Recently, a Transformer encoder-decoder model has been shown to leverage multilingual data well in IPA transcriptions of languages presented during training. However, the representations it learned were not successful in zero-shot transfer to unseen languages. Because that model lacks an explicit factorization of the acoustic model (AM) and language model (LM), it is unclear to what degree the performance suffered from differences in pronunciation or the mismatch in phonotactics. To gain more insight into the factors limiting zero-shot ASR transfer, we replace the encoder-decoder with a hybrid ASR system consisting of a separate AM and LM. Then, we perform an extensive evaluation of monolingual, multilingual, and crosslingual (zero-shot) acoustic and language models on a set of 13 phonetically diverse languages. We show that the gain from modeling crosslingual phonotactics is limited, and imposing a too strong model can hurt the zero-shot transfer. Furthermore, we find that a multilingual LM hurts a multilingual ASR system's performance, and retaining only the target language's phonotactic data in LM training is preferable. |