Modeling DNN as human learner

Ni, Junrui

Permalink

https://hdl.handle.net/2142/106027 Copy

Description

Title

Modeling DNN as human learner

Author(s)

Ni, Junrui

Contributor(s)

Hasegawa-Johnson, Mark

Issue Date

2019-12

Keyword(s)

• Phoneme Category Adaptation
• Human Perceptual Learning
• Deep Neural Networks
• Time-course
• Second Language Learning
• Articulatory Feature Detection

Abstract

In previous experiments, human listeners demonstrated that they had the ability to adapt to unheard, ambiguous phonemes after some initial, relatively short exposures. At the same time, previous work in the speech community has shown that pre-trained deep neural network-based (DNN) ASR systems, like humans, also have the ability to adapt to unseen, ambiguous phonemes after retuning their parameters on a relatively small set. In the first part of this thesis, the time-course of phoneme category adaptation in a DNN is investigated in more detail. By retuning the DNNs with more and more tokens with ambiguous sounds and comparing classification accuracy of the ambiguous phonemes in a held-out test across the time-course, we found out that DNNs, like human listeners, also demonstrated fast adaptation: the accuracy curves were step-like in almost all cases, showing very little adaptation after seeing only one (out of ten) training bins. However, unlike our experimental setup mentioned above, in a typical lexically guided perceptual learning experiment, listeners are trained with individual words instead of individual phones, and thus to truly model such a scenario, we would require a model that could take the context of a whole utterance into account. Traditional speech recognition systems accomplish this through the use of hidden Markov models (HMM) and WFST decoding. In recent years, bidirectional long short-term memory (Bi-LSTM) trained under connectionist temporal classification (CTC) criterion has also attracted much attention. In the second part of this thesis, previous experiments on ambiguous phoneme recognition were carried out again on a new Bi-LSTM model, and phonetic transcriptions of words ending with ambiguous phonemes were used as training targets, instead of individual sounds that consisted of a single phoneme. We found out that despite the vastly different architecture, the new model showed highly similar behavior in terms of classification rate over the time course of incremental retuning. This indicated that ambiguous phonemes in a continuous context could also be quickly adapted by neural network-based models. In the last part of this thesis, our pre-trained Dutch Bi-LSTM from the previous part was treated as a Dutch second language learner and was asked to transcribe English utterances in a self-adaptation scheme. In other words, we used the Dutch model to generate phonetic transcriptions directly and retune the model on the transcriptions it generated, although ground truth transcriptions were used to choose a subset of all self-labeled transcriptions. Self-adaptation is of interest as a model of human second language learning, but also has great practical engineering value, e.g., it could be used to adapt speech recognition to a lowr-resource language. We investigated two ways to improve the adaptation scheme, with the first being multi-task learning with articulatory feature detection during training the model on Dutch and self-labeled adaptation, and the second being first letting the model adapt to isolated short words before feeding it with longer utterances.

Type of Resource

text

Language

en

Permalink

http://hdl.handle.net/2142/106027

Owning Collections