Mathematical methods in automatic speech recognition
Keywords:
speech corpus, automatic speech recognition, low resource languageAbstract
We present the mathematical methods which are used in the process of Automatic Speech Recognition. The presentation is divided in three parts. We start with a short overview of the vocal tract and the corresponding acoustics equations. Afterwards we introduce the digital signal processing, which is performed over the speech signal in order to extract Mel-frequency cepstrum coefficients, corresponing to the articulation configuration. In the second part we present an approach for acoustic modeling based on time-delayed deep neural networks. We discuss the methodology for the machine learning of the acoustic model. In the third part we describe the use of finite-state f-transducers for representing the language model. For decoding the signal we shortly present the Viterby and the beam-search algorithm over a Hidden Markov Model represented as a f-transducer. Finally, we show experimental results for automatic speech recognition of Bulgarian language.
References
L. Deng, Y. Liu. Deep Learning in Natural Language Processing, Springer, 2018.
D. Geneva, G. Shopov, S. Mihov. Building an ASR Corpus Based on Bulgarian Par- liament Speeches. In: Statistical Language and Speech Processing (Eds C. Mart´ın-Vide, M. Purver, S. Pollak). Springer International Publishing, 2019, 188–197.
X. Huang, A. Acero, H.-W. Hon. Spoken Language Processing: A Guide to Theory, Algorithm, and System Development. Prentice hall PTR, 2001. [4] F. Jelinek. Statistical Methods for Speech Recognition. MIT press, 1997.
V. Peddinti, D. Povey, S. Khudanpur. A time delay neural network architecture for efficient modeling of long temporal contexts. In: INTERSPEECH, 2015.
D. Yu, L. Deng., Automatic Speech Recognition. Springer, 2016.