I'm Nay, currently a PhD candidate at Stanford Linguistics. My research focuses on data-centric approaches to creating speech and language technologies for digitally underserved languages and user populations. In recent work, I have been examining how state-of-the-art models trained for major languages can be adapted for speech information retrieval applications related to endangered language documentation and revitalisation.
- Experiments with wav2vec 2.0 models involving only 10 minutes of transcribed speech: https://github.com/fauxneticien/w2v2-10min-replication
- A pipeline to isolate and transcribe one language in mixed-language speech: https://github.com/CoEDL/vad-sli-asr
- A reference implementation for a Query-by-Example Spoken Term Detection service: https://github.com/parledoct/qbestdocks
- Evaluation of feature extraction methods for query-by-example spoken term detection with low resource languages: https://github.com/fauxneticien/qbe-std_feats_eval
- Query by example spoken term detection using bottleneck features and a convolutional neural network: https://github.com/fauxneticien/bnf_cnn_qbe-std
- An R package for tidying lexicographical data in backslash-coded formats: https://github.com/CoEDL/tidylex