C is the number of correctly recognized words, I is the number of incorrect recognitions and N is the number of total words.
A simulation of 200 runs is performed. To this end, a two small but well-known databases are used. The simulations are performed in two phases: the training phase and the testing phase. During the training phase, different words are estimated by GMM models and stored in the codebook to be used later during the testing phase. During the testing phase, unknown words are presented by GMMs and compared with the Gaussian mixture configurations available in the codebook. Then, the most similar one is considered the recognised word.
The first database used is 16-bit, mono at a sampling rate of 44 KHz. Word level labelling at millisecond accuracy is done manually, this database contains 05 different speakers and 5 words chosen randomly and similar in terms of …show more content…
Hence, the accuracy of the results is guaranteed. According to the same results, an adjustable threshold could be fixed on KL divergence, in order to know whether the word is already available on the codebook or not. In this case, the KL measure is very important as it leads to detect automatically that the word does not exist in the codebook. A reference signal of this word should be trained to be added as a new vocabulary to the reference codebook. By this way, the reference codebook could be extended, the database used is improved and the errors in the speech recognition could be avoided automatically.
The effect of noisy environments on the recognition performance is also analysed. Many simulations in different noise levels (signal to noise ratios, SNR) are performed by adding white noise to the MFCCs. We have observed that the percent accuracy rate of the recognition decreases slightly when the noise becomes important, but the recognition results are still acceptable as presented in figure