In this article, research was conducted on the development of automatic Uzbek speech recognition technology based on integral models. Methods of continuous speech recognition technology in Uzbek were studied at all stages and suitable ones were selected. A 200-hour speech corpus was trained on the DNN-CTC architecture for acoustic modeling. The accuracy of the developed speech recognition system achieved WER = 17.3

This paper presents an improved method for highly accurate identification of the speaker who speaks sentences in the Uzbek language from a limited set of sentences. It is proposed the use of an advanced method for identifying the speaker speaking in the Uzbek language through the use of classification method - Gaussian mixture model (GMM) for decision making. The advanced method is a combination of Gaussian mixture model and MFCC. Current scientific research is aimed at identifying opportunities to improve performance. One of the latest implementations of MFCC are Delta MFCC (DMFCC) and Delta Delta Delta MFCC (DDMFCC), which improve speaker identification. Identification accuracy was obtained from the results of the research. According to the experimental results, speaker identification by conventional MFCC combined with GMM classification algorithm (94.5%) lower than the second (DMFCC+GMM - 96.8%) and third (DDMFCC+GMM-98.1%) feature extraction methods. It is easy to see the performance improvement observed when using DDMFCC in combination with GMM and is 98.1%. first, second, and third level headings (first level heading).

The article deals with the approach based on integrated models of speech recognition in the Uzbek language. As well as, the advantages of integral models in speech recognition, the stages of their formation and the functions of their components are pointed out. Works on training and testing of a collection of words consisting of names of geographical places in the Republic, terms used in urban planning cadaster, names of constituent elements of urban planning documents were carried out on the basis of the architecture of the speech recognition system based on the proposed integrated model. As a result of testing, the evaluation metric for the number of incorrectly recognized characters was CER = 17%.

We present a freely available speech corpus for the Uzbek language and report preliminary automatic speech recognition (ASR) results using both the deep neural network hidden Markov model (DNNHMM) and end-to-end (E2E) architectures. The Uzbek speech corpus (USC) comprises 958 different speakers with a total of 105 hours of transcribed audio recordings. To the best of our knowledge, this is the first open-source Uzbek speech corpus dedicated to the ASR task. To ensure high quality, the USC has been manually checked by native speakers. We first describe the design and development procedures of the USC, and then explain the conducted ASR experiments in detail. The experimental results demonstrate promising results for the applicability of the USC for ASR. Specifically, 18.1% and 17.4% word error rates were achieved on the validation and test sets, respectively. To enable experiment reproducibility, we share the USC dataset, pre-trained models, and training recipes in our GitHub repository.

Currently, speech recognition systems are becoming more widespread, especially in those applications where speech dialogue is the most convenient mean of information control and exchange with technical facilities. Obtaining an effective voice control system is currently an important task, requiring the development of methods to obtain high recognition accuracy of voice commands. Under these conditions, along with reducing the noise effect on the quality of recognition, the task is to increase the accuracy in voice control system operation, to increase the likelihood of correct command recognition under stationary interference. Another requirement in the voice control system is the correct recognition speed, since the system must work in real time. In this paper, we propose an algorithm for voice control by technical facilities, based on Uzbek language. Some sounds in Uzbek language (explosive and unvoiced consonants) differ strongly from the sounds in other languages; when creating control algorithms, the recognition accuracy does not meet the requirements. Therefore, to ensure the necessary processing speed and maintain the required accuracy, it is proposed to introduce an additional normalization with a decrease in feature space. The algorithm is based on the principle of primary separation of speech signal spectral characteristics. Further, the signal spectrum is normalized and its resolution increases due to the use of low-frequency conversion and logarithms. The obtained cepstral coefficients are fed to the input of a previously learned neural network.

Handwritten character recognition (HCR) is one of the ongoing research field in Artificial Intelligence. The recognition of handwritten characters is based on pattern recognition and image processing. One of the most common machine learning methods for solving this problem is Convolutional Neural Networks (CNN). To implement such tasks, a multilayer neural network is often used. CNN require as much data as possible to ensure high accuracy, while parallel processing can help us save time when neural network training. Such training time can be reduced by improving training operations using parallel computing technology and multi-core platforms. In this paper, exemplary parallelization of CNN training by dint of OpenMP technology and its libraries has been implemented. English alphabets were chosen as an experiment for the handwritten character recognition system, and CNN was used for its better accuracy. Each character data set contains 26 alphabets. That is, the image size of each character ranges from 16×16 to 256×256 pixels, and these pixels are taken as features for training the CNN. We analyzed how the speedup of CNN training depends on image size. Three multi-core processors with different specifications were chosen for the hardware implementation. The results show that the proposed parallel approach using OpenMP technology gives good acceleration, which reduces training time.

In this paper investigates the approach to realization of recognition of Uzbek words on the basis of end-to-end models is considered. Also presented are some theoretical data on the architecture of neural networks used in the integrated model, and the results of preliminary experimental studies conducted on their basis. Deep recurrent neural networks, which combine the multiple levels of representation that have proved so effective in deep networks with the flexible use of long-range context that empowers RNNs. When trained end-to-end with suitable regularization, we find that deep BRNNs achieve a test set error of CER=49.1% on our dataset.

In this paper has been discussed about speech recognition using spectrogram images and deep convolution neural network(CNN) of Uzbek spoken digits. Spectrogram images from speech signal were generated and it were used for deep CNN training. Presented CNN model contains 3 convolution layers and 2 fully connected layers that discriminative features can be divided and estimated of spectrogram images by those layers. In current research period, dataset of Uzbek spoken digits were made and in based on presented CNN model they were trained. Testing results shows that, proposed approach for Uzbek spoken digits classified 100% accuracy.

This paper proposes a new method for optimal using the multi-core processor cache in parallel processing of signals and achieving to maximum speed up. Acceleration problems have not been fully resolved till present. One of the most actual tasks is the creation of fast algorithms using multicore processors, especially in the field of digital signal processing. The proposed method is designed for parallel processing of speech signals based on spectral analysis. Input signals are processed by splitting frames using a new parallel algorithm. Intel’s advanced tools, such as TBB and OpenMP, were used to create fast algorithms. The results of parallel processing acceleration were analyzed. Also, the problem of optimizing the cache memory of multi-core processors was analyzed in parallel processing tasks. For hardware implementations were chosen quad-core and eight-core processors. The results showed that when the signal frame size corresponds to the size of the L3 cache blocks, the acceleration of the calculation of the signal spectrum becomes closer to the number of processor cores.

In such computerized systems, such as voice control units, personal identification, IP-telephony, weapon control commands, accepting applications for reference services, automated stenography, recognition of individual words and phrases plays a major role in enhancing the effectiveness of technical systems. In the preliminary stages of speech processing algorithms for the allocation of phonetic characteristics is implemented, which are subjected to syntactic and semantic analysis in subsequent stages. To isolate the phonetic characteristics acoustic treatment is performed, which includes algorithms of pre-filtering, spectral analysis, segmentation, and a calculation of cepstral features. In this paper solves the problems of accelerated implementation of the acoustic processing algorithms in real-time systems, in which there is recognition of the individual teams. To implement the stream computing is used opportunities of C++ programming language and OpenMP technologies. To perform stream computing algorithms are used dual-core and multi-core processors. Obtained results of acceleration calculations correspond to the number of processor cores.

In this paper, signal compression algorithm of Chebyshev polynomial is implemented. Upper case and lover case paralleled algorithm of compression signals processing suggested. However, audio and speech signal compressing accuracy in ADSPBF-561 dual core processor are analyzed. Index Terms—Discrete cosine transform, Chebyshev polynomial, signal processors, parallel programming, library automation programming. I. INTRODUCTION Today spectral methods of digital signal processing includes several transformations. Among them, Fourier transform, cosine and many similar transformations. These transformations differ each other with basis functions and the calculation processes. Using these transformations, first we calculate spectral values and based on them spectral values signal can be restored. After that using spectral value we can research internal properties of the signal. In these values are reflected analysis of signal, filtration process, signal compression processes. Processors and digital signal processors provides the process of calculation in digital signal processing. Digital signal processing is mainly implemented in real time systems. In the real time systems signals are received, processed and transmitted simultaneously. During signal processing in real-time systems, a signal incoming from the analog-to-digital converter using spectral methods obtains spectral values of signal (1). Using these spectral values we can organize of the signal processing. The one of the main direction of digital signal processing is the signal compression. Today, there are many methods of digital signal compression based on different mathematical methods. By using these digital compression algorithms in real time systems, we can solve kind of problems like saving memory space, time delay, signal processing time, performance of digital signal processors etc. Consequently, working on digital signal compressing and developing algorithms of digital signal compressing of the real time systems becomes important. coefficients compresses signal. These algorithms refers Newton, Lagrange, method of least squares and others. Using classic algorithms to digital signal compressing gives a good effect but it can't give normal results when you use in real time systems for digital signal processing. The main reason is increasing amount of signal values leads to increase number of solvable equation. Consequently, the solving the equations requires time, this leads to a delay which is not desirable for real time systems. Therefore, today many real time systems use digital signal processors. General structure of the process of direct and inverse transformation in a specialized signal processor is given in the illustration below (Fig. 1).

Within the scope of linear theory the transmission of signal or its fragment into the spectral sphere with one of the discrete basis of Fourier-analysis is given, further the signal spectrum of given algorithms transforms into the algebraic polynomial.