Experiments in speech recognition using a modular MLP architecture for acoustic modelling

Phoneme Recognition on the TIMIT Database

“…The best results achieved were given by a combination of only five of the eight classifiers available. In 2003 Reynolds and Antoniou (Reynolds & Antoniou, 2003) proposed training a modular MLP. On a first level they trained the 39 phones but used different feature sets (MFCCs, perceptual linear prediction coefficients, LPC and combinations of them).…”

“…Table 5. Broad classes of phones used in the system proposed by Reynolds and Antoniou, (Reynolds & Antoniou, 2003).…”

“…The label of the winning model is compared with the corresponding TIMIT label and a hit or an error occurs. Nevertheless, phone classification allows a good evaluation of the quality of the acoustic modelling, since it computes the performance of the recognizer without the use of any kind of grammar, (Reynolds & Antoniou, 2003). Phone recognition obeys harder and more complex criteria.…”

Biologically inspired Continuous Arabic Speech Recognition

“…Therefore, many modern speech recognition systems perform the recognition process without prior segmentation. These systems tend to be based on the use of features extraction techniques such as MFCC [6,[11][12][13][14], FFT [15][16][17], HFCC [18], Linear Predictive Coefficient (LPC) [8] ... etc.…”

“…Although statistical approaches using HMMs have been the most successful for speech recognition, neural networks have also been applied to speech classification problems [6,8,12,15,[17][18][19][20]. A couple of research projects have shown the capability of using neural networks for Arabic phoneme recognition.…”