Topics

2017.05.22

【Awards and Commendations】Shohei TOYAMA, Department of Electrical Engineering and Information Systems, wins a Students' Best Presentation Award at Spring meeting of 2017.

On 28th Spring 2017, Shohei TOYAMA, Department of Electrical Engineering and Information Systems, wins a Students' Best Presentation Award at Spring meeting of 2017.

Integration of Acoustic Features into Language Models for Spontaneous Speech Recognition
http://www.asj.gr.jp/recommending/index.html#gakusei

NN-based language models for speech recognition were conventionally built only by using large text corpora. However, spoken words depend on non-linguistc factors in speech such as speakers' gender, age, and identity. Further, various environments also affect what kind of expressions are used for conversation. In this work, acoustic features related to these factors are introduced and testedto increase the word prediction performance of language models for speech recognition.

 

 

 

The current results are based on fundamental investigation of what kind of acoustic features can improve the performance of language models. It was shown that the word prediction performance was certainly improved but it did not necessarily lead to speech recognition performance improvement. We're going to do more sophisticated investigation of what kind of acoustic features are more related to word selectionand attempt to improve the performance of automatic speech recognition.