Output details

Correctness-Adjusted Unsupervised Discriminative Acoustic Model Adaptation

<22> Techniques for adapting acoustic models to match new environments are essential for robust speech recognition. However, discriminative adaptation approaches fail in situations where transcribed adaptation data is unavailable. This paper represents a new understanding of the underlying causes of this problem and presents a novel solution that operates by employing discriminative wide-range error classifiers. This approach is the first to show consistent gains for discriminative adaptation over non-discriminative baselines. The work was funded by Cisco Systems (Unsupervised Domain Adaptation) as part of a project for developing speech transcription for teleconferencing systems.