**7.1.2 The genetic algorithm in L3**

Consecutive executions of the GA produced variable Fisher's ratios. It was observed here that the population size is critical, since a population of 20 chromosomes produced Fisher's ratios between 35 and 42, while a population of 120 individuals easily produced Fisher's ratios above 58. It was noticed during experimentation that specific values for restrictions R1 and R2 also have a strong influence in the outcome.

Comparing the results over a traditional approach with neural networks and cepstral coefficients it is evident a higher performance and, more important, the system exhibits realtime operation and very low computational effort compared to neural networks and realtime computation of the Cepstral coefficients.

### **7.1.3 Results in the L0 English vocabulary**

The genetic algorithm was executed 30 times, varying population size, probability of mutation, restrictions, number of sub-bands, and other parameters. The initial number of sub-bands was 8, then it was scaled to 9 and 12; in this scenario, the Fisher's ratio varied from 23 (8 sub-bands, population size = 20) to 48 (12 sub-bands, population size = 200). The resulting centers and bandwidths were:

> [ ] [ ] C 250 446 648 1283 1483 1776 2018 2506 2737 3197 3383 3833 BW 5 138 187 157 139 43 207 106 105 98 148 224 = =

#### **Intra-class repeatability and inter-class differences**

The performance in the training set was 100%; the performance in the testing set was 99%. The correct classification per word was {100% 100% 98% 100% 97% 100% 98% 99%} for the respective words {*'faster','slower','left','right','stop','forward','reverse','brake'*}, respectively. Figure 10 shows the normalized espectra of two utterances of the word 'faster' and the word 'slower'. In both cases notice the repeatability in the frequency domain, as well as the difference between both sets of spectra.

Fig. 10. Normalized spectra of the words "faster" and "slower".

#### **7.1.4 Results of the real-time implementation**

258 Bio-Inspired Computational Algorithms and Their Applications

The genetic algorithm was executed 30 times, and the maximum Fisher's ratio obtained was

BF = [254 180 526 132 744 118 1196 141 1483 115 2082 86 2295 171 2828 46]

BW 180 132 118 141 115 86 171 46

The recognition rates in the training and testing sets were 100% and 99%, respectively. In real conditions the correct classification rate was 93.5% in 40 repetitions of each word to the

Consecutive executions of the GA produced variable Fisher's ratios. It was observed here that the population size is critical, since a population of 20 chromosomes produced Fisher's ratios between 35 and 42, while a population of 120 individuals easily produced Fisher's ratios above 58. It was noticed during experimentation that specific values for restrictions

Comparing the results over a traditional approach with neural networks and cepstral coefficients it is evident a higher performance and, more important, the system exhibits realtime operation and very low computational effort compared to neural networks and real-

The genetic algorithm was executed 30 times, varying population size, probability of mutation, restrictions, number of sub-bands, and other parameters. The initial number of sub-bands was 8, then it was scaled to 9 and 12; in this scenario, the Fisher's ratio varied from 23 (8 sub-bands, population size = 20) to 48 (12 sub-bands, population size = 200). The

[ ]

[ ] C 250 446 648 1283 1483 1776 2018 2506 2737 3197 3383 3833

The performance in the training set was 100%; the performance in the testing set was 99%. The correct classification per word was {100% 100% 98% 100% 97% 100% 98% 99%} for the respective words {*'faster','slower','left','right','stop','forward','reverse','brake'*}, respectively. Figure 10 shows the normalized espectra of two utterances of the word 'faster' and the word 'slower'. In both cases notice the repeatability in the frequency domain, as well as the

BW 5 138 187 157 139 43 207 106 105 98 148 224

[ ] [ ] C 254 526 744 1196 1483 2082 2295 2828

**7.1 Results and discussion** 

microphone.

**7.1.1 Results in the L3 Spanish vocabulary** 

From which the corresponding center and bandwidth were:

= =

R1 and R2 also have a strong influence in the outcome.

time computation of the Cepstral coefficients.

**7.1.3 Results in the L0 English vocabulary** 

resulting centers and bandwidths were:

difference between both sets of spectra.

**Intra-class repeatability and inter-class differences** 

= =

62. The resulting best chromosome was:

**7.1.2 The genetic algorithm in L3**

A minimum distance classifier was implemented in a digital signal processor TMS320LF2407 for each of the four lexicons **L**={L0, L1, L2, L3} from Texas Instruments, in order to verify the performance using in a nearly real-life application. The voiced/no-voiced segmentation was performed using a push-button to start and finish capturing the voice. The DSP has a built-in 10-bit analog to digital converter facilitating the interfacing task. The digital filters used were IIR topology, elliptic type, 8th order. The filter coefficients (*A, B*) were calculate using the Matlab® Software. The analog-to-digital conversion was set-up to acquire one sample every T seconds, (T=1/6000); each time a sample came into de device, the filters actuated and the respective output was squared and accumulated to calculate the energy of the signal. Scaling issues had to be solved since the model was created in a real valued [-1 , 1] scale, while the DSP just "see" integer values. Once a whole command was processed, it was just a matter of a few miliseconds to apply the minimum distance classifier and provide the classification. The correct classification rate was in this case of the order of 94.5%, in a total of 1200 repetitions of the words in **L**.
