Phoneme transcription of endangered languages: an evaluation of recent ASR architectures in the single speaker scenario

Findings of the Association for Computational Linguistics: ACL 2022, pages 2301 - 2308 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Phoneme transcription of endangered languages: an evaluation of recent ASR architectures in the single speaker scenario Gilles Boulianne CRIM (Centre de recherche informatique de Montréal) gilles.boulianne@crim.ca Abstract Transcription is often reported as the bottle- neck in endangered language documentation, requiring large efforts from scarce speakers and transcribers. In general, automatic speech recognition (ASR) can be accurate enough to accelerate transcription only if trained on large amounts of transcribed data. However, when a single speaker is involved, several studies have reported encouraging results for phonetic transcription even with small amounts of train- ing. Here we expand this body of work on speaker-dependent transcription by comparing four ASR approaches, notably recent trans- former and pretrained multilingual models, on a common dataset of 11 languages. To au- tomate data preparation, training and evalua- tion steps, we also developed a phoneme recog- nition setup which handles morphologically complex languages and writing systems for which no pronunciation dictionary exists. We find that fine-tuning a multilingual pretrained model yields an average phoneme error rate (PER) of 15% for 6 languages with 99 min- utes or less of transcribed data for training. For the 5 languages with between 100 and 192 minutes of training, we achieved a PER of 8.4% or less. These results on a num- ber of varied languages suggest that ASR can now significantly reduce transcription efforts in the speaker-dependent situation common in endangered language work. 1 Introduction Recent progress in automatic speech recognition (ASR) was made by training neural networks on increasingly large amounts of annotated data. To significantly reduce the efforts needed to transcribe endangered languages, ASR must reach sufficient accuracy when trained on relatively much smaller amounts of transcribed data. Already several re- search efforts have been dedicated specifically to ASR for low-resource languages, such as the IARPA BABEL program1 and the NIST OpenASR Challenge2. However, creating an ASR system for a task like speaker-independent phonetic tran- scription is still difficult and requires amounts of transcription that are very large in the context of en- dangered languages. For example, Shi et al. (2021) recently concluded that at least 50 hours of training data are needed for this task, comparing ESPnet and HMM-based models on two languages. In language documentation, field recordings are seldom made with a large number of speakers, but rather with a few speakers and for long du- rations (Amith et al., 2021). In these conditions, small amounts of transcribed data from a single speaker might be enough to train a phoneme rec- ognizer with sufficient accuracy to automatically transcribe the remaining recordings from the same speaker. Concentrating on the single speaker sce- nario, Adams et al. (2018) evaluated a CTC-based LSTM model on Na and Chatino, and showed en- couraging results for automated phoneme transcrip- tion as well as the effectiveness of this approach for linguistic work on endangered languages; they also created the open-source phonemic transcrip- …

Content truncated to 500 words