Leveraging End-to-End ASR for Endangered Language Documentation: An Empirical Study on Yolóxochitl Mixtec

Abstract

Transcription bottlenecks", created by a shortage of effective human transcribers are one of the main challenges to endangered language (EL) documentation. Automatic speech recognition (ASR) has been suggested as a tool to overcome such bottlenecks. Following this suggestion, we investigated the effectiveness for EL documentation of end-to-end ASR, which unlike Hidden Markov Model ASR systems, eschews linguistic resources but is instead more dependent on large-data settings. We open source a Yoloxóchitl Mixtec EL corpus. First, we review our method in building an end-to-end ASR system in a way that would be reproducible by the ASR community. We then propose a novice transcription correction task and demonstrate how ASR systems and novice transcribers can work together to improve EL documentation. We believe this combinatory methodology would mitigate the transcription bottleneck and transcriber shortage that hinders EL documentation.

Data-driven Model Generalizability in Crosslinguistic Low-resource Morphological Segmentation

“…In particular, for small corpora, the DNN architecture has been demonstrated to yield better results than statistical alternatives (e.g., subspace Gaussian mixture models) and other neural architectures (e.g., time delay neural networks) (Morris, 2021). We also found the DNN to be substantially more accurate than the endangered language end-to-end recipe (Shi et al, 2021) in ESPnet (Watanabe et al, 2018). 1 1 Our experiments using ESPnet and wav2vec 2.0 to finetune from multilingual models yielded inconsistent and weak Crucially, however, we note that our goal is not to improve upon current state-of-the-art for lowresource ASR but rather to examine what data partitioning strategies and evaluation methods lead to reliable estimates in low-resource settings with an already strong model architecture.…”

Combining Spectral and Self-Supervised Features for Low Resource Speech Recognition and Translation

“…In experiment SSL, the FBANK feature extractor (used in Base) is replaced by the pretained Hu-BERT model, which is fine-tuned during training. 7 Learnable combinations : Experiments Linear, Conv. and co-Att.…”

“…End-to-end models based on deep learning have demonstrated their superiority over conventional hidden Markov-based models on speech tasks for some corpora [1][2][3][4]. End-to-end models could be beneficial to low resource speech tasks because these models: (1) alleviate the need of language specific resources such as lexicons [5][6][7]. (2) can be trained multilingually to facilitate cross-lingual transfers between high resource and low resource languages through shared architecture and weights [8].…”

“…We also performed preliminary experiments over Wav2vec2 XLSR model[45], but it did not improve the results over HuBERT model so we continued the study only for HuBERT model 7. SSL based front-ends could be freezed, but the best performances were obtained when fine-tuning the models.…”

Proceedings of the Second Workshop on NLP Applications to Field Linguistics

“…The recent improvement of speech recognition for low-resource languages has also been seen as a way to mitigate the transcription bottleneck automatically transcribing large amount of untranscribed speech data (e.g. Foley et al, 2018;Shi et al, 2021;Adams et al, 2021).…”