雲间烟火Shruti and Bharti Chhabra (2016) have implemented a singer identification technique using Artificial Neural Network (ANN). The authors have focused on the basic and future works are discussed in section 6.
II. RELATED WORKS
This section presents the related research work carried out particularly for GFCC features used for various speech related applications. The research findings from our previous experiments by using GFCC features are also presented at the end of this section.
R. Schluter, L. Bezrukov, H. Wagner and H. Ney (2007) say that the Gammatone features lead to competitive results for large vocabulary speech recognition. Further, different methods to combine Gammatone features with a number of standard acoustic features such as MFCC, PLP, MF-PLP and Vocal Tract Length Normalization (VTLN) were investigated. Best results were obtained when combining all features using weighted ROVER, resulting in a relative improvement of about 12% in word error rate compared to the best single feature system [5].
Hui Yin, Volker Hohmann and Climent Nadeu (2011) have proposed a variety of features based on Gammatone filter banks. The phase modulation is represented by the sub band Instantaneous Frequency (IF) and it is explicitly used by concatenating envelope-based and IF-based features [6]. Experiments are done with Chinese mandarin digits corpus under both clean and multi-condition using HMM. Their results prove that the proposed features can improve recognition rates in both conditions compared to the MFCC-based recognizer.
Shaveta Sharma and Parminder Singh (2014) have done Speech Emotion Recognition using GFCC and BPNN for English speech data. The authors have considered two emotions SAD and HAPPY [7]. The experiments were done under matlab and the results prove that the BPNN with GFCC feature extraction method performs better. The authors have also discussed about extracting GFCC features for three emotions namely, sad, happy and angry (2015) [8].
P.K. Sahu, Astik Biswas, Anirban Bhowmick and Mahesh Chandra (2014) have discussed about auditory Equivalent Rectangular Bandwidth (ERB) based admissible wavelet packet features for TIMIT phoneme recognition [9]. The authors has introduced a new filter structure using admissible wavelet packet for English phoneme recognition. In the proposed filters, the Central frequencies of ERB scale are equally distributed along with the frequency response of human cochlea. The new sets of features derived from wavelet packet transform with multi resolution capabilities were found to be better
concepts of the feature extraction and classification techniques in speech identification system [10]. The total dataset of 45 songs are used by involving 9 singers and 5 songs of each singer. In their work, DCT is applied to derive cepstral features, GFCC is used for feature extraction and ANN is applied to classify. For the experiments, 88.9 % of singers were correctly identified using the above mentioned techniques.
Hari Krishna Maganti and Marco Matassoni (2014) have discussed about Auditory processing-based features for improving speech recognition in adverse acoustic conditions [11]. The proposed features incorporates a combination of gammatone filtering, modulation spectrum, non-linearity emulating the cochlear and the middle ear to improve robustness. The experimental results revealed that the proposed features provide reliable and considerable improvement in terms of robustness in different noise conditions by using standard Aurora-4 large vocabulary database. Their future work is to focus on evaluating the performance of the proposed features for reverberant environments and large vocabulary tasks.
Shaik Shafee and B.Anuradha (2016) have done an experiment for speaker identification and spoken word recognition in noisy background using ANN [12]. The main objective of the work is to find the better combination of speech feature extraction and ANN for speaker identification combined with spoken word recognition in general noisy environment. Experiments are done by using Mel Frequency Cepstral Coefficients (MFCC), Perceptual Linear Prediction (PLP) Cepstral Coefficients and Gammatone Frequency Cepstral Coefficients (GFCC) in combination with Radial Basis Neural Networks and Learning Vector Quantization. Three different test categories such as Spoken word recognition, Speaker Identification, and the combination of both speaker and spoken word recognition have been experimented for the above mentioned combinations. From the above results it is suggested that Radial basis neural networks with GFCC can be chosen for the combination of both spoken word recognition and speaker identification in general noisy conditions. It was also observed that the Radial basis network models were found to be less time consuming compared to Learning Vector Quantization neural networks.
In our previous experiment (2014), four types of feature extraction techniques namely, MFCC, LPC, PLP and GFCC are implemented with DTW, HMM, MLP, SVM and Decision Tree techniques [13]. Experiments are done with 10 Tamil spoken digits (0-9) and 5 Tamil
