雲间烟火Efficient Acoustic Front-End Processing for Tamil Speech Recognition using Modified GFCC Features
recorded at 16 KHz sampling rate using audacity software at a silence environment.
In this experiment, 60% of dataset is given for training and the remaining 40% of dataset are used for testing. The performances of these techniques are analyzed based on two metrics, namely, Word Recognition Rate (WRR) and Real Time Factor (RTF).The results obtained by applying the proposed five pass pre-processing and three modified GFCC techniques are shown in the following that the FWCMN-MTYW-GFCC-FF has increased the WRR up to 99.06% for the HMM technique. Also, the techniques have some significant improvement in reducing the processing time. The average testing time taken for MLP and SVM techniques has been reduced with the proposed techniques which is shown in Figure 7.
V. FINDINGS AND DISCUSSIONS
Table 1.
Table 1. Performance Evaluation of Proposed Techniques
From the above results, it is clear that the proposed front-end processing techniques are lead to competitive results for Tamil speech recognition. The usage of the above techniques has substantially improved the recognition rate for the dataset used. As per the experimental outcomes, it is shown that the five pass pre-processing technique has demonstrated the significant performance improvements for all the three speech recognition techniques adopted for this study. The improvements in WRR of 1.19%, 2.98% and 2.99% has been achieved for the HMM, MLP and SVM techniques respectively. In same way, the developed MTYW-GFCC features have increased the recognition accuracy for all the three speech recognition techniques involved in this research work.
Fig.7. Performance Evaluation of Proposed Technique based on Real
Time Factor
Likewise, very good results are achieved by using MTYW-GFCC-FF, where the WRR of 98.54% has been achieved for the HMM technique. Similar to the above proposed techniques, it is evident from the experiments
The main objective of this paper is, to propose a suitable pre-processing and feature extraction techniques for speaker independent isolated speech recognition for Tamil language. Accordingly, the merits of GFCC technique have been considered and improved using five-pass Pre-processing and modified feature extraction techniques.
In the proposed work, initially the five pass preprocessing technique has been applied and the improvements are tested with existing GFCC feature extraction method. Then, the GFCC feature extraction has been modified with multi taper and Yule walker AR power spectrum method. Subsequent improvements are achieved by using the MTYW-GFCC features for all the three techniques involved. Based on the improvements gained with the MTYW-GFCC features, the combinational features using formant frequencies were implemented. The combinational features were also improved the recognition accuracy for HMM technique. Furthermore, in order to reduce both speaker and channel variations, frequency warping and feature normalization techniques were implemented. The FWCMN-MTYW-GFCC-FF has provided very good results for HMM when compared with the MLP and SVM. The reason is that, the statistical methods can learn as much information as possible from the data to build unique model for each speech patterns. The learning based approaches works based on the fewer assumptions made on input data. It was found that continues improvements were gained with the proposed pre-processing and modified GFCC features. The highest recognition rate achieved with the existing GFCC using HMM, MLP and SVM technique are 96.5%, 92.8% and 92.9% respectively. The maximum accuracy of 99.06%, 96.15% and 96.14% was achieved with the HMM. From the outcome, it is proved that better results were achieved for all the speakers enrolled in this study.
VI. CONCLUSION AND FUTURE WORK
The main objective of this research work is to propose a novel preprocessing and feature extraction technique for speaker independent isolated speech recognition system for Tamil language. For Tamil speech recognition, only MFCC and LPC feature extraction techniques are implemented using HMM and Back propagation techniques. The most recent feature extraction techniques like GFCC and other machine learning techniques were not carried out for Tamil ASR. Hence, in this research work the GFCC and other machine learning techniques
