Phone Based Keyword Spotting for Transcribing Very Low Resource Languages

Phone Based Keyword Spotting for Transcribing Very Low Resource Languages ´Eric Le Ferrand,1,2 Steven Bird,1 and Laurent Besacier2 1Northern Institute, Charles Darwin University, Australia 2Laboratoire Informatique de Grenoble, Universit´e Grenobles Alpes, France Abstract We investigate the efficiency of two very differ- ent spoken term detection approaches for tran- scription when the available data is insufficient to train a robust speech recognition system. This work is grounded in a very low-resource language documentation scenario where only a few minutes of recording have been tran- scribed for a given language so far. Experi- ments on two oral languages show that a pre- trained universal phone recognizer, fine-tuned with only a few minutes of target language speech, can be used for spoken term detec- tion through searches in phone confusion net- works with a lexicon expressed as a finite state automaton. Experimental results show that a phone recognition based approach pro- vides better overall performances than Dy- namic Time Warping when working with clean data, and highlight the benefits of each meth- ods for two types of speech corpus. 1 Introduction Efforts are made across Australia to preserve, doc- ument and revitalize Aboriginal languages. These languages exist primarily in spoken form, and even if there often is an official orthography available, it is not widely used by local people. Making record- ings of speakers has been a widespread practice for documenting traditional knowledge. However, such recordings are often not transcribed, making them hard to access. Manual transcription is time consuming and is often described as a bottleneck (Brinckmann, 2009). While automatic speech recognition (ASR) has seen great improvements in recent years (Povey et al., 2011; Watanabe et al., 2018), it relies on a large amount of annotated data. Attempts to build ASR systems for low-resource languages end up with high word error rate or single-speaker models making them of limited use in Indigenous contexts (Gupta and Boulianne, 2020a,b). Such methods assume that everything should be transcribed. Bird (2020) describes a sparse tran- scription model where we only transcribe the words we can confidently recognize, using word-spotting, while leaving the transcription of more difficult sec- tions for later, perhaps when a speaker is available (Bird, 2020). Based on this model, Le Ferrand et al. (2020) proposed a workflow which combines spo- ken term detection and a human-in-the-loop to sup- port transcription in under-resourced settings. Such a workflow avoids the use of a language model which requires too much textual data, data that we cannot find in most Aboriginal contexts, and which only needs a few spoken terms to be anno- tated. While they show through their simulation the capability of iterative transcription in remote com- munities, the precision of their method depends on the quality of the spoken queries, and the density of the resulting transcription is limited by the size of the lexicon. Automatic phone recognition has seen progress with minimal data (Gupta and Boulianne, 2020b; Li et al., 2020). While Bird (2020) argues that pho- netic transcriptions do not stand in for …

Content truncated to 500 words