Semi-supervised and Weakly Supervised Training of Speech Recognition Models

Automatic Speech Recognition (ASR) is now available in the form of cloud services as well as deployable open-source tools. However, poor performance due to mismatch with the domain of end applications still limits their usage; especially with limited amount of labeled/unlabelled in-domain data. Efficiently exploiting this limited amount of data is vital for early development stages, under-resourced languages as well as privacy-critical applications. This seminar will present ASR training methods being developed under the COMPRISE project. The first part of the seminar will go through an error guided semi-supervised training approach which trains ASR acoustic models using lattices decoded from unlabelled speech data. The second part will present the ongoing work on a noisy channel approach for learning ASR language models from unlabelled speech.