Rail ticketing dialogue using HMM-based speech recognition

Abstract

ABSTRACT This paper studies the usage of Hidden Markov Model (HMM) in Hidden Markov Model Toolkit (HTK) for Malay language continuous speech recognition system. This study focuses on recognizing six different utterance structures of common Light Rail Transit (LRT) ticketing dialogue sentences spoken by passengers while purchasing train ticket at Putra LRT station, Malaysia. This study recognizes the Malay language speech input and produces recognition result based on the training and testing data. It was built as a prototype for future ticketing speech recognition system especially for LRT ticketing. The speech is sampled at 16 kHz with 16-bit resolution. Mel-Frequency Cepstrum Coefficients (MFCC) is used as feature extraction technique to describe speech signal. The Hidden Markov Model is applied as the speech classifier. Viterbi algorithm is chosen as a search algorithm in this study because it is able to find the most likely path from all other alternative paths. All the methods mentioned above are applied to produce more precise and faster results in speech recognition. Overall, this paper recognizes the Malay language speeches input for train ticketing system in monophones and triphones. The overall speech recognition rate results for monophones and triphones are achieved 91.41% and 88.95% respectively.

ABSTRAK Tujuan penulisan dan kajian tesis ini adalah mengaji kegunaan Hidden Markov Model (HMM) dalam Hidden Markov Model Toolkit (HTK) untuk mengecam dialog dalam Bahasa Melayu. Jumlah enam struktur ayat yang berkenaan dengan perbualan pembelian Sistem Transit Tren Ringan (LRT) tiket di Perhentian Putra LRT telah disediakan. Kajian ini khas reka untuk mengecam suara daripada pembeli dan menunjukkan keputusan pengecaman selepas melalui proses pengecaman suara. Sampel rakaman suara adalah 16kHz and 16 bit revolusi. Mel-Frequency Cepstrum Coefficients (MFCC) digunakan untuk proses penyarian sifat. Manakala HMM adalah untuk latihan rakaman semua dialog. Viterbi algorithm digunakan untuk mencari dan mengesan dialog yang paling sesuai dibaca oleh pembeli. Secara umumnya, teknik-teknik yang dikaji dalam kajian ini amat sesuai untuk mengecam suara selain menghasilkan keputusan dengan cepat. Kesimpulannya, kajian ini dapat membaca, mengecam dan menghasil keputusan daripada dialog semasa pembelian tiket di Putra LRT. Keputusan kajian bagi pengecam suara monophones menunjukkan 91.41% manakala pengecam suara untuk triphones menunjukkan 88.95%.