Improved Feature Parameter Extraction From Speech Signals Using Machine Learning Algorithm(2)

3.7. Dataset Dataset collection and generation processes were performed as follows. In this study, the dataset with 120 h of audio was used for the model training. The dataset includes speech audio recordings which consist of sentences with a maximum of 15 words with a total length of approximately 120 h.

Desert ginseng

In addition, the dataset includes a large amount of different text for use in developing the language model. Over 90,650 utterances, 415,780 words, and 65,810 unique words that were included in the text corpus were collected, resulting in around 120 h of transcribed speech data. We split the dataset into training, validation, and test sets. The dataset statistics

are reported in Table 3.

Table 3. The dataset specifications.

We split the dataset into three folders corresponding to the training, validation, and test sets. Each folder contains audio recordings and transcripts. The audio and corresponding transcription filenames are the same, except that the audio recordings are stored as WAV files, whereas the transcriptions are stored as TXT files using the UTF-8 encoding. All the transcriptions are represented using the Latin alphabet consisting of 29 letters and the apostrophe symbol. To prevent overfitting, we applied data augmentation techniques based on velocity perturbation and spectral enhancement.

Cistanche deserticola

4. Proposed Method

An important task for programmers when developing speech recognition systems is the creation of an optimal method for the parametric expression of speech signals [45]. This method enables excellent separation of sounds and spoken words while ensuring that speakers are insensitive to pronunciation patterns and changes in the acoustic environment. Most errors in word recognition are caused by a change in the pitch of the signal owing to a shift in the microphone or a difference in the pitch of the pronunciation [46]. Another common cause of errors is random nonlinear deformations of the spectrum shape, which are always present in the speech signal of a speaker [47,48]. Therefore, one of the most important tasks in creating effective speech recognition systems is the selection of a representation that is sufficient for the content of the analyzed signal as well as insensitive to the voices of speakers and various acoustic environments.

Cistanche supplement near me-Improving Memory

The system that is used for extracting feature parameters typically has the following requirements. The information content, that is, the set of feature parameters, must ensure the reliable identification of recognizable speech elements. Furthermore, the loudness, that is, the maximum compression of the audio signal, and the non-statistical correlation of the parameters must be minimized. Independence from the speaker must also be achieved, that is, the maximal removal of information relating to the characteristics of the speaker from the vector of characters. Finally, homogeneity, which refers to the parameters having the same average variance and the ability to use simple metrics to determine the affinity between character sets, must be provided [49]. However, it is not always possible to satisfy all requirements simultaneously because such requirements are contradictory. The parametric description of the speech elements should be sufficiently detailed to distinguish them reliably and should be as laconic as possible.

Superman herbs cistanche

In practice, the speech signal that is received from a microphone is digitized at a sampling rate of 8 to 22 kHz. Serial numerical values are divided into speech fragments (frames) with a duration of 10 to 30 ms, which correspond to quasi-stationary speech parts. A vector of features is computed from each frame, which is subsequently used at the acoustic level of speech recognition. At present, a wide range of methods is available for the parametric representation of signals based on autocorrelation analysis, hardware linear filtering, spectral analysis, and LPC. The most common approach for speech parameterization is the spectral analysis of signal fragments and the calculation of their cepstral coefficients.

MFCCs have been used as informative features for the speech signal [41]. These characteristics are used extensively in speech recognition and are based on two main concepts: the cepstral and Mel scales. The main advantages of the algorithm are its high level of familiarity and ease of speech. The MFCC features are separated from the recorded speech signals. The MFCC algorithm uses the results of the phonogram and spectrum switching algorithms. The classical algorithm that is used to calculate the MFCCs is depicted in Figure 5.

Figure 5. Classical scheme for calculating MFCCs.

This study presents a rapid method for extracting the function parameters from a speech signal. The proposed algorithm for the rapid calculation of the MFCCs is shown in Figure 6.

Figure 6. Proposed framework for calculating MFCCs.

We consider the execution sequence of the proposed algorithm for the rapid extraction of the function parameters from a speech signal

4.1. Division into Frames

Following preliminary filtering, the speech signal is divided into 16 ms frames. Every frame (except for the first) contains the final 10 ms of the previous frame. This process continues until the end of the signal. In this study, because the sampling rate of the speech signal is 16 kHz, the frame length is N = 256, and the offset length is M = 160. The overlap is 62.5% of the frame length. A coverage of 50% to 75% of the frame length is generally recommended.

4.2. Hanning Window and Decreasing Values

A Hanning window size of 1D was used. The Hanning window is also called the raised cosine window. The Hanning window can be thought of as the sum of the frequency spectrum of three rectangular time windows. It can use the side lobes to cancel each other out, eliminating high-frequency interference and energy leakage. Hanning windows are very useful window functions. 

Cistanche supplement near me-Improving Memory

Click here to view  Cistanche Improving Memory and Prevent Alzheimer's Disease products

【Ask for more】 Email:cindy.xue@wecistanche.com /  Whats App:  0086 18599088692 /  Wechat:  18599088692 

A weight box is used to reduce distortion and smooth the individual frames. The floating signal that is examined in this study consists of a dense tone. The magnitude of the flat tone is determined by the magnitude of the pure tone at frequency f, which is filtered through the Hanning window. 

A critical aspect of this window is that it sets the borders of the frames to zero. In this case, the short energies can be calculated while passing through the window, and they can be transferred from the sequence that is used to calculate the lowest amplitude energies. The aim is to remove low-energy signals from the signal by calculating the signal energy while smoothing the signal from this window. This process requires the following signal energy equation:

where En is the energy of the input signal fragment, and xi is the signal value.

In addition to the window process using (11), the signal is processed in the next step, which signifificantly reduces the number of values that enter the processor. Figure 7 presents the parallel processing algorithm.

The window size represents a number of samples and a duration. It is the main parameter of the analysis. The window size depends on the fundamental frequency, intensity and changes in the signal.

Figure 7. Hanning window algorithm for removing silent parts.
4.3. Short-Time Fourier Transform (STFT) Switches An intuitive understanding exists of the meaning of a high or low height. STFT
is a Fourierrelated transform that is used to determine the sinusoidal frequency and phase content of local sections of a signal as it changes over time. In practice, the STFT computation involves the division of a longer time signal into shorter segments of equal length, ollowed by a separate calculation of the Fourier transform on each shorter segment.This reveals the Fourier spectrum of each shorter segment.Discrete-time signals are used in practice. The corresponding time-frequency conver-sion is a discrete Fourier conversion, which describes the length of signal Xn as representative of the complex value frequency domain of the N coefficients. STFT, which describesthe evolution of the frequency components over time, is one of the most widely used toolsfor speech analysis and processing [50]. Similar to the spectrum itself, one advantage ofSTFT is that its parameters have physical and intuitive interpretations. STFT is typicallyvisualized using the log spectra 20log10 (X(h, k)). Such 2D log spectra can then be viewedusing a thermal map known as a spectrogram.
In the third stage of the algorithm, the STFT spectral switching procedure is applied tothe frames that are passed through the weight window. The STFT of the signal is obtainedby opening the windows and determining the DFT of each window. In particular, the
transformation for the Xn and Wn windows of the input signal is determined as follows:

where the k-index corresponds to the frequency values, and wn is the window function, which is commonly a Hanning window or Gaussian window that is centered around zero.

4.4. Mel Transform

In the fourth stage, the signal that is transferred to the frequency band is divided into ranges using triangular fifilters. The fifilter boundaries are calculated using the chalk frequency. The transition to the chalk frequency fifield is based on the following equation:

where f is the frequency range.

The reverse switch is determined as follows:

Consider NN as the number of fifilters (26 fifilters are generally used) and flflows, as high as the frequency range under study. This range is transferred to the Mel scale and divided into NN evenly distributed intersecting ranges. The linear frequency-appropriate boundaries are determined within the fifield, whereas the weighting coeffificients that are obtained based on fifiltration are denoted by H. Subsequently, the fifilters are applied to the square modulus of the coeffificients obtained from the Fourier transform. The values that are obtained are logarithmic owing to the following expression:

The singular value decomposition algorithm is implemented in the fifinal stage of calculating the MFCCs.

5. Experimental Results

We implemented and tested the proposed method in Visual Studio 2019 C++ on a PC with a 4.90 GHz CPU, 32 GB of RAM, and two Nvidia GeForce 2080Ti GPUs, as shown in Table 4. The system was tested in different device environments to evaluate the performance of the signal feature extraction method. In the experiments, during the operation of the algorithm (Figure 8), when the sequence of surfaces of the signal was n = 15, the number of values was reduced by 40–50%, and the processing time increased 1.2-fold. Consequently, this algorithm exhibited signifificantly higher effificiency. Moreover, this algorithm enabled the separation and elimination of areas of silence while passing through the Hanning window.

Several possible means are available for avoiding the waste of memory bandwidth. We propose a new solution that increases computing performance by matching the size of the signal frames to a block size of cache memory. This type of optimization can signifificantly affect the overall parallel processing performance. However, it can be used in digital signal processing by dividing the signal into frames through implementations on multicore processors. However, in practice, the selection is usually performed on a small scale, corresponding to the width of the data bus that connects the cache memory to the main memory and the size of its block. Our method implements the optimal use of these memories in parallel computing. The organization of the cache memory plays an essential role in parallel processing algorithms when dividing data into streams. In particular, the presence of vector-matrix effects in digital signal processing and the size of their streams should be adjusted according to the size of the cache blocks. This can be achieved using the proposed method, as illustrated in Figure 9.

Table 4. The detailed specififications of the experimental setup.

figure 8. (a) Initial incoming signal and (b) appearance of signal after applying proposed algorithm.

Figure 9. Parallel computing structure using RK3288 processors.

In this section, we discuss a quantitative analysis to compare the performance of different systems. We compared our method with well-known speech recognition algorithms based on deep-learning approaches. Evaluation metrics are essential for computing various strategies for speech recognition and assessing the performance of different approaches. Although we used the results of other studies for comparison, we are not certain whether they are true because the source codes and datasets of these methods are not publicly available to verify the actual performance. Figure 10 depicts the result of removing the silent parts during the passage of a speech signal fragment through the Hanning window based on the proposed rapid algorithm. The speed results obtained from the analysis are presented in Table 5.

Figure 10. k value of KNN algorithm (with feature selection).

Table 5. Experimental results of proposed method.

A range has been determined in order to fifind the degree of the neighborhood that gives the best accuracy value in the KNN algorithm. The specifified range covers 1–25. In Figure 10, the graph of the KNN algorithm with feature selection was applied. When the graph was examined, when the neighborhood value was 1 at the beginning, the training accuracy was much higher than the test accuracy. In the KNN algorithm created using the features selected by correlation, the accuracy of the model was determined to be 99.15% in the training dataset and 97.35% in the test dataset.

The word error rate (WER) or character error rate is typically used to evaluate the accuracy of feature extraction from a speech signal. These are objective matrices that are helpful for a fair comparison of recognition techniques. In our previous studies [51–56], we computed metrics such as the F-measure (FM), precision, and recall. The FM is the weighted average that balances the measurements between the precision and recall rates. The precision is the ratio of the number of correctly predicted positive observations to the total number of predicted positive observations. The recall is the ratio of the number of correctly predicted positive observations to the total number of observations in the actual class, as indicated in (9). The following equations can be used to calculate the average precision and recall rates of feature extraction methods:

where TP denotes the number of true positives, FP denotes the number of false positives, and FN denotes the number of false negatives.

The FM is calculated using (10), considering both precision and recall.

The average FM, recall, and precision of the proposed method was 98.4%. False detection occurred in 1.6% of cases owing to the unwanted noise of signals at the microphone. The range of the model’s accuracy was between 0 and 1, and the metric estimation scores reached their best values at 1. An evaluation of our method and other recently published speech feature extraction methods is presented in Table 6. The same number of features was used for a fair comparison. A total of 325 speech samples from each group were analyzed from subjects with similar feature backgrounds. The effects of different frame lengths according to the number of fifilter banks in the MFCC and different frame lengths in the order of the LPC were also examined for improved accuracy.

Table 6. Quantitative accuracy results of speech feature extraction.

As mentioned previously, the WER is the most common measure for speech recognition performance. It is computed by comparing a reference transcription with the output of the speech recognizer. Based on this comparison, it is possible to calculate the number of errors, which typically belong to three categories: (1) insertions when a word is not present in the reference in the output of the automatic speech recognition (ASR), (2) deletions when a word is missed in the ASR output, and (3) substitutions when a word is confused with another word. The WER can be computed as follows.

where S is the number of substitutions of words that are incorrectly recognized, D is the number of deletions, I is the number of insertions, and N is the number of words in the reference transcription. The main issue in computing this score is the alignment between the two-word sequences. This can be determined through dynamic programming using the Levenshtein distance [67].

Based on Table 6, we performed a statistical analysis to indicate the average accuracy of the compared methods using the WER evaluation metric, as illustrated in Figure 11. The improved feature extractor yielded an accuracy of approximately 98.4%, whereas the other approaches yielded accuracies between 78% and 96%. We used the results provided in the relevant papers for comparison; however, the accuracy of these values is not easily verifiable because the source codes and datasets of these methods are not publicly available to confirm their real performance. Nevertheless, in the case of standard scenes, the proposed method was experimentally demonstrated to provide excellent speech feature extraction accuracy by reducing the computational time, even when the speech data are noisy or of low quality.

Figure 11. Quantitative results of speech signal feature extraction approaches using vertical graphs.

Moreover, we evaluated the false-positive results of the selected methods. As can be observed from Figure 12, the proposed approach had the fewest errors. Furthermore, the highly effificient parallel computation method signifificantly reduced the sound signal feature selection and extraction errors. Overfifitting was one of the main issues during the training, and almost all machine learning models suffer from it. We tried to reduce overfifitting risk using a feature selection technique that aims instead to rank the importance of the existing features in the dataset and discard less important ones (no new features are created).

Figure 12. Visible results of false-positive speech signal feature extraction experiments.

Table 7 displays the performance results of the methods that were used in speech recognition environments based on different properties. Our proposed approach does not suffer from unwanted and unnecessary background noise and is not affected by low-quality human voices such as hoarse voices, voices produced with a sore throat, or even sounds from humans with complete voice loss. Our method aims to overcome the problems of inadequate recording equipment, background noise, difficult accents and dialects, and various pitches in a voice. In a normal environment, the best results for accurately detecting and extracting speech feature challenges were obtained using the proposed method with a reduced processing time.

Table 7. Review of speech feature detection and extraction performance using various features.

The results of the speech recognition methods were classifified as powerful, normal, or weak for the seven categories. The powerful criterion demonstrates that the algorithm can overcome all types of challenges. In contrast, the normal criterion indicates that the algorithm may fail in certain instances because the word boundaries are not defifined beforehand. Finally, the weak criterion suggests that the algorithm is unreliable under background noise or vibrations.

6. Limitations

It is diffificult to conclude that the methods proposed to date do not exhibit any shortcomings. Our proposed method may also result in errors owing to various noise environments. To overcome this problem, we aimed to reduce the number of features in the dataset by creating new features from existing features [69]. Since overfifitting was one of the main issues for training different models during the competition, enriching the training data by adding data samples from different resources could be a possible solution for improving the results. Irrespective of the aforementioned problems, the experimental results revealed that our method was very robust and effective for speech feature extraction tasks, with an average accuracy of 98.4% and FM of 99.5%.

7. Conclusions

A novel high-performance parallel computing approach using a machine learning method has been proposed for speech recognition systems. Acceleration problems in machines with limited computing resources can be solved using distributed systems. The computation speed in signal recognition systems can be increased, and the performance of multicore platforms can be improved by creating and using effificient and rapid algorithms. The results demonstrate that the proposed model reduces the processing time and improves the feature extraction accuracy by 98.4% by effectively using MFCCs. It has been observed that the features with low correlation values extracted by feature selection are also effective in the success of the model. The statistical analysis was performed on the pre-processed data, and meaningful information was produced from the data using the K-nearest neighbors (KNN) machine learning algorithm.

Future studies will focus on improving the accuracy of our method by using deep learning approaches and optimizing the cache memory of multicore processors to detect and extract speech signals without a signifificant quality loss. Furthermore, we plan to construct a spectral analysis model based on parallel processing with robust analysis performance that will enable the establishment of embedded devices with low computational resources using Taris speech datasets [70] in the 3D CNN and 3D U-Net Environment [71–75]

References 

1. Meng, Y.J.; Liu, W.J.; Zhang, R.Z.; Du, H.S. Speech Feature Parameter Extraction and Recognition Based on Interpolation. Appl. Mech. Mater. 2014, 602–605, 2118–2123. [CrossRef] 2. Musaev, M.; Rakhimov, M. Accelerated Training for Convolutional Neural Networks. In Proceedings of the 2020 International Conference on Information Science and Communications Technologies (ICISCT), Tashkent, Uzbekistan, 4–6 November 2020; pp. 1–5. [CrossRef] 3. Ye, F.; Yang, J. A Deep Neural Network Model for Speaker Identifification. Appl. Sci. 2021, 11, 3603. [CrossRef] 4. Musaev, M.; Rakhimov, M. A Method of Mapping a Block of Main Memory to Cache in Parallel Processing of the Speech Signal. In Proceedings of the 2019 International Conference on Information Science and Communications Technologies (ICISCT), Karachi, Pakistan, 9–10 March 2019; pp. 1–4. [CrossRef] 5. Jiang, N.; Liu, T. An improved speech segmentation and clustering algorithm based on SOM and k-means. Math. Probl. Eng. 2020, 2020, 3608286. [CrossRef] 6. Hu, W.; Yang, Z.; Chen, C.; Sun, B.; Xie, Q. A vibration segmentation approach for the multi-action system of numerical control turret. Signal Image Video Process. 2021, 16, 489–496. [CrossRef]

7. Popescu, T.D.; Aiordachioaie, D. Fault detection of rolling element bearings using optimal segmentation of vibrating signals. Mech. Syst. Signal Process. 2019, 116, 370–391. [CrossRef] 8. Shihab, M.S.H.; Aditya, S.; Setu, J.H.; Imtiaz-Ud-Din, K.M.; Efat, M.I.A. A Hybrid GRU-CNN Feature Extraction Technique for Speaker Identifification. In Proceedings of the 2020 23rd International Conference on Computer and Information Technology (ICCIT), Dhaka, Bangladesh, 19–21 December 2020; pp. 1–6. [CrossRef] 9. Korkmaz, O.; Atasoy, A. Emotion recognition from speech signal using mel-frequency cepstral coefficients. In Proceedings of the 9th International Conference on Electrical and Electronics Engineering (ELECO), Bursa, Turkey, 26–28 November 2015; pp. 1254–1257. 10. Ayvaz, U.; Gürüler, H.; Khan, F.; Ahmed, N.; Whangbo, T.; Abdusalomov, A. Automatic Speaker Recognition Using Mel-Frequency Cepstral Coeffificients Through Machine Learning. CMC-Comput. Mater. Contin. 2022, 71, 5511–5521. 11. Al-Qaderi, M.; Lahamer, E.; Rad, A. A Two-Level Speaker Identifification System via Fusion of Heterogeneous Classififiers and Complementary Feature Cooperation. Sensors 2021, 21, 5097. [CrossRef] 12. Batur Dinler, Ö.; Aydin, N. An Optimal Feature Parameter Set Based on Gated Recurrent Unit Recurrent Neural Networks for Speech Segment Detection. Appl. Sci. 2020, 10, 1273. [CrossRef] 13. Kim, H.; Shin, J.W. Dual-Mic Speech Enhancement Based on TF-GSC with Leakage Suppression and Signal Recovery. Appl. Sci. 2021, 11, 2816. [CrossRef] 14. Lee, S.-J.; Kwon, H.-Y. A Preprocessing Strategy for Denoising of Speech Data Based on Speech Segment Detection. Appl. Sci. 2020, 10, 7385. [CrossRef] 15. Rusnac, A.-L.; Grigore, O. CNN Architectures and Feature Extraction Methods for EEG Imaginary Speech Recognition. Sensors 2022, 22, 4679. [CrossRef] [PubMed] 16. Wafa, R.; Khan, M.Q.; Malik, F.; Abdusalomov, A.B.; Cho, Y.I.; Odarchenko, R. The Impact of Agile Methodology on Project Success, with a Moderating Role of Person’s Job Fit in the IT Industry of Pakistan. Appl. Sci. 2022, 12, 10698. [CrossRef] 17. Aggarwal, A.; Srivastava, A.; Agarwal, A.; Chahal, N.; Singh, D.; Alnuaim, A.A.; Alhadlaq, A.; Lee, H.-N. Two-Way Feature Extraction for Speech Emotion Recognition Using Deep Learning. Sensors 2022, 22, 2378. [CrossRef] [PubMed] 18. Marini, M.; Vanello, N.; Fanucci, L. Optimising Speaker-Dependent Feature Extraction Parameters to Improve Automatic Speech Recognition Performance for People with Dysarthria. Sensors 2021, 21, 6460. [CrossRef] 19. Tiwari, S.; Jain, A.; Sharma, A.K.; Almustafa, K.M. Phonocardiogram Signal Based Multi-Class Cardiac Diagnostic Decision Support Syste. IEEE Access 2021, 9, 110710–110722. [CrossRef] 20. Mohtaj, S.; Schmitt, V.; Möller, S. A Feature Extraction based Model for Hate Speech Identifification. arXiv 2022, arXiv: 2201.04227. 21. Kuldoshbay, A.; Abdusalomov, A.; Mukhiddinov, M.; Baratov, N.; Makhmudov, F.; Cho, Y.I. An improvement for the automatic classifification method for ultrasound images used on CNN. Int. J. Wavelets Multiresolution Inf. Process. 2022, 20, 2150054. 22. Passricha, V.; Aggarwal, R.K. A hybrid of deep CNN and bidirectional LSTM for automatic speech recognition. J. Intell. Syst. 2020, 29, 1261–1274. [CrossRef] 23. Mukhamadiyev, A.; Khujayarov, I.; Djuraev, O.; Cho, J. Automatic Speech Recognition Method Based on Deep Learning Approaches for Uzbek Language. Sensors 2022, 22, 3683. [CrossRef] [PubMed] 24. Li, F.; Liu, M.; Zhao, Y.; Kong, L.; Dong, L.; Liu, X.; Hui, M. Feature extraction and classifification of heart sound using 1D convolutional neural networks. EURASIP J. Adv. Signal Process. 2019, 2019, 59. [CrossRef] 25. Chang, L.-C.; Hung, J.-W. A Preliminary Study of Robust Speech Feature Extraction Based on Maximizing the Probability of States in Deep Acoustic Models. Appl. Syst. Innov. 2022, 5, 71. [CrossRef] 26. Ramírez, J.; Górriz, J.M.; Segura, J.C. Voice Activity Detection. Fundamentals and Speech Recognition System Robustness. In Robust Speech Recognition and Understanding; Grimm, M., Kroschel, K., Eds.; I-TECH Education and Publishing: London, UK, 2007; pp. 1–22. 27. Oh, S. DNN Based Robust Speech Feature Extraction and Signal Noise Removal Method Using Improved Average Prediction LMS Filter for Speech Recognition. J. Converg. Inf. Technol. 2021, 11, 1–6. [CrossRef] 28. Abbaschian, B.J.; Sierra-Sosa, D.; Elmaghraby, A. Deep Learning Techniques for Speech Emotion Recognition, from Databases to Models. Sensors 2021, 21, 1249. [CrossRef] 29. Rakhimov, M.; Mamadjanov, D.; Mukhiddinov, A. A High-Performance Parallel Approach to Image Processing in Distributed Computing. In Proceedings of the 2020 IEEE 14th International Conference on Application of Information and Communication Technologies (AICT), Uzbekistan, Tashkent, 7–9 October 2020; pp. 1–5. [CrossRef] 30. Abdusalomov, A.; Mukhiddinov, M.; Djuraev, O.; Khamdamov, U.; Whangbo, T.K. Automatic Salient Object Extraction Based on Locally Adaptive Thresholding to Generate Tactile Graphics. Appl. Sci. 2020, 10, 3350. [CrossRef] 31. Abdusalomov, A.; Whangbo, T.K. An improvement for the foreground recognition method using shadow removal technique for indoor environments. Int. J. Wavelets Multiresolution Inf. Process. 2017, 15, 1750039. [CrossRef] 32. Abdusalomov, A.; Whangbo, T.K. Detection and Removal of Moving Object Shadows Using Geometry and Color Information for Indoor Video Streams. Appl. Sci. 2019, 9, 5165. [CrossRef] 33. Mery, D. Computer Vision for X-ray Testing; Springer International Publishing: Cham, Switzerland, 2015; p. 271, ISBN 978-3319207469. 34. Mark, S. Speech imagery recalibrates speech-perception boundaries. Atten. Percept. Psychophys. 2016, 78, 1496–1511. [CrossRef] 35. Mudgal, E.; Mukuntharaj, S.; Modak, M.U.; Rao, Y.S. Template Based Real-Time Speech Recognition Using Digital Filters on DSP-TMS320F28335. In Proceedings of the 2018 Fourth International Conference on Computing Communication Control and Automation (ICCUBEA), Pune, India, 16–18 August 2018; pp. 1–6. [CrossRef]