Md. Sah Salam

Fuzzy Phoneme Classification Using Multi-speaker Vocal Tract Length Normalization

ABSTRACT The overall success of automatic speech recognition (ASR) depends on efficient phoneme recognition performance and quality of speech signal received in ASR. However, dissimilar inputs of speakers affect the overall recognition performance. One of the main problems that affect recognition performance is inter-speaker variability. Vocal tract length normalization (VTLN) is introduced to compensate inter-speaker variation on the speaker signal by applying speaker-specific warping of the frequency scale of a filter bank. Instead of measuring the performance on word level with speaker-specific warping, this research focuses on direct tackling at the phoneme level and applying VTLN on all speakers’ speech signals to analyse the best setting for the highest recognition performance. This research seeks to compare each phoneme recognition results from warping factor between 0.74 and 1.54 with 0.02 increments on nine different ranges of frequency warping boundary. The warp factor and frequency warping range that provides the highest phoneme recognition performance is applied on word recognition. The results show an improved performance in phoneme recognition by 0.7% and spoken word recognition by 0.5% using warp factor of 1.40 on frequency range of 300–5000 Hz in comparison to baseline results.

Speech segmentation using divergence algorithm with Zero Crossing property

Divergence algorithm is a statistical segmentation approach which finds segmentation point via detection of abrupt changes without any previous information of the acoustic signal. The approach could get high match of segmentation but also gives a lot of false segmentation points. This work introduced a property based on the usage of Zero Crossing Rate (ZCR) in enhancing segmentation by divergence algorithm. The work starts via optimizing divergence algorithm segmentation performance via parameters tuning. Then the proposed property based on ZCR is applied to divergence algorithm to reduce insertion points. The results of tuning divergence parameters achieved match rate of 99.4% at time tolerance of 0.09 seconds with 69% insertion rate occurrences in comparisons to reference points. The result in applying the introduced ZCR property to divergence algorithm shows that tuning of some ZCR property parameters could reduce insertion between 4% to 45%. However, it would also reduce the match rate. Nevertheless, the method could reduced insertion rate by 5.5% while maintaining match rate of 99.4%.

Recognition of Holy Quran Recitation Rules Using Phoneme Duration

Many of the Holy Quran Recitation rules (which referred to Tajweed) depend on the phoneme duration, such as Medd, Ghunnah and some letters characters. This paper proposes a model to recognize the Medd rule and Ghunnah using the phoneme duration. The data has collected from 10 Holy Quran reciter to calculate the Medd and Ghunnah duration in the correct recitation. Then, the proposed model used to recognize the Medd and Ghunnah as Tajweed rules in Quran recitation. This model can support the existing traditional methods of Quran learning process by proposing a method that help the unskilled reciters to learn a part of Tajweed rules (Medd and Ghunnah). The reciters will be able to recite the Holy Quran by themselves and the system will correct them if there is any mistake in Al-Medd and Ghunnah rules.

Insertion reduction in speech segmentation using neural network

Statistical approach with non-fixed overlapping window size is able to make good identification of discontinuity in speech signal without further knowledge upon the phonetic sequence. This however, leads to increase number of insertion and thus increase confusion in recognition. This paper present a fusion between statistical and connectionist approach namely divergence algorithm and MLP neural network to improved segmentation by reducing insertions. The experiment conducted on Malay semi-spontaneous connected digit in classroom environment. The digit strings were manually segmented and trained using neural network with three set of data. The first training set trained without silence pattern, the second include silence while the last set introduced both silence and false pattern in the training. The experimental result on digit string segmentation shows number of insertion reduction of more than 5 times in comparison using divergence alone with increment of accuracy up to 40%.. The drawback however, the number of omission also increases to more than 10 times. Nevertheless, match segmentation rate still above 85%.