نشریه مهندسی برق و مهندسی کامپیوتر ایران
سال نوزدهم شماره 1 (پیاپی 61، بهار 1400)

One of the challenges of isolated Persian digit recognition is similar pronunciation of some digits such as "zero and three", "nine and two" and "five, seven and eight". This challenge leads to the high substitution errors and reduces the recognition accuracy. In this paper, a combined solution based on short-term memory (LSTM) and hidden Markov model (HMM) is proposed to solve the mentioned challenge. The proposed approach increases the recognition rate of Persian digits on average 2 percent and in the best case 8 percent in comparison to the HMM-based approach. In the following of this work, due to the intensification of the mentioned challenge in noisy conditions, the robust recognition of Persian digits with similar pronunciation was considered. In order to increase the robustness of the LSTM-based recognizer, robust features extracted from the speech spectrum such as spectral entropy, burst degree, bisector frequency, spectral flatness, first formant and autocorrelation-based zero crossing rate were used. Using these features, while reducing the number of features for recognizing similar Persian digits from 39 coefficients to a maximum of 4 and a minimum of 1 coefficient, on average improved the robustness of the isolated digit recognizer in different noisy conditions (30 different situations resulting from five noise types of white, pink, babble, factory and car noises and six signal-to-noise ratios of -5, 0, 5, 10, 15 and 20 decibels) by 10%, 13%, 15% and 13% compared to the HMM-based, LSTM-based, deep belief network-based recognizers with Mel-Cepstrum coefficients and a convolutional neural network-recognizer with Mel Spectrogram features.

This paper proposes a new high speed low power algorithm for unsigned digital multiplier without pipeline which could be easily expanded to a wider number of bits. The blocks of multiplier works in parallel which significantly increase the speed of multiplier. In proposed algorithm, the input bits of multiplier, are divided into smaller groups of bits which multiplication of these groups are in parallel and simultaneously. This division continues until the minimum number of input bits which is 2×2. In calculating the product of each category, the proposed algorithm is used, which leads to acceleration of the product of each category.The final result will be obtained from the sum of these smaller categories.Modified tree adder have been used to add smaller groups, which can increase the multiplication speed. Multipliers with input bit lengths of 64, 32, 16, 8, 4, and 2 have been implemented using the proposed algorithm in 180 nm and 90 nm technology, which its delay and power consumption with bit length of 32 in 180 nm are 3.05 ns and 40 mW respectively. In 90 nm technology and with the 32 bit length the delay is 1.53 nm and power consumption is 9.7 mW. Also, using the proposed method, it is estimated that the delay of 128×128 bits multiplier in the 180 nm and 90 nm technology are equal to 5.4ns and 2.5ns, respectively. According to the results and in comparison with other works reported in the articles and in the same process, without increasing the power consumption and with a silicon area of 1.5 times, the proposed multiplication speed has increased more than 2 times.

With the advancement of technology and the shrinking dimensions of transistors in CMOS technology, several challenges have arisen. One of the main concerns in using CMOS-based memory is the high power consumption of this type of memory. Therefore, new and non-volatile memories were introduced to address the shortcomings of conventional volatile memory. One of the emerging non-volatile technologies is STT-MRAM memory, an effective and efficient alternative to conventional memory such as SRAMs due to low leakage power, high density, and short access time. The positive features of STT-MRAMs make it possible to use them at different memory hierarchy levels, especially the cache level. However, STT-MRAMs suffer from high write energy. In this paper, we present a new write circuit using the temperature method; in addition to improving the high write energy, write delay is also improved. The proposed circuit lead to 22.5% and 18.62% improvement in energy and writing delay, respectively, compared to the existing methods.

Cloud data centers often fail to cope with the millions of delay-sensitive storage and computational requests due to their long distance from end users. A delay-sensitive request requires a response before its predefined deadline expires, even when the network has a high load of requests. Fog computing architecture, which provides computation, storage and communication services at the edge of the network, has been proposed to solve these problems. One of the fog computing challenges is how to allocate cloud and fog nodes resources to user requests in congestion conditions to achieve a higher acceptance rate of user requests and minimize their response time. Fog nodes have limited storage and computational power, and hence their performance is significantly reduced due to high load of user requests. This paper proposes an efficient resource allocation method in fog computing that decides where (fog or cloud) to process the requests considering the available resources of fog nodes and congestion conditions. According to the experimental results, the performance of the proposed method is better compared with existing methods in terms of average response time and percentage of failed requests.

Wireless sensor networks consist of thousands of sensor nodes with limited energy. Energy efficiency is a fundamental challenge issue for wireless sensor networks. Clustering sensor nodes in separate categories and exchanging information through clusters is one of the ways to improve energy consumption. This paper presents a new cluster-based routing protocol called SFLCFBA. The proposed protocol biologically uses fast and effective search features inspired by the Shuffled Frog Leaping algorithm, which acts based on the Frog food behavior to cluster sensor nodes. The proposed protocol also uses fuzzy logic to calculate the node fitness, based on the two criteria of distance to the sink and the remaining energy of the sensor node or power of battery level. IEEE 802.15.4 Protocol and NODIC Protocol with the proposed methodology and OPNET Simulator were simulation and the results in terms of energy consumption, end to end delay, signal to noise ratio, the success property data and throughput were compared with each other. The results of the simulation showed that the proposed method outperforms the IEEE 802.15.4 Protocol and NODIC Protocol due to the use of the criteria listed.

Classification is one of the most important tasks in data mining and machine learning; and the decision tree, as one of the most widely used classification algorithms, has the advantage of simplicity and the ability to interpret results more easily. But when dealing with huge amounts of data, the obtained decision tree would grow in size and complexity, and therefore require excessive running time. Almost all of the tree-construction algorithms need to store all or part of the training data set; but those algorithms which do not face memory shortages because of selecting a subset of data, can save the extra time for data selection. In order to select the best feature to create a branch in the tree, a lot of calculations are required. In this paper we presents an incremental scalable approach based on fast partitioning and pruning; The proposed algorithm builds the decision tree via using the entire training data set but it doesn't require to store the whole data in the main memory. The pre-pruning method has also been used to reduce the complexity of the tree. The experimental results on the UCI data set show that the proposed algorithm, in addition to preserving the competitive accuracy and construction time, could conquer the mentioned disadvantages of former methods.

This paper presents a new method for blind two-channel speech sources separation without the need for prior knowledge about speech sources. In the proposed method, by weighting the mixture signal spectrum based on the location of the speech sources in terms of distance to the microphone, the speech sources are separated. Therefore, by forming an angular spectrum by generalized cross-correlation function, the speech sources in the mixture signal are localized. First, by creating an angular spectrogram by generalized cross-correlation function, the speech sources in the mixture signal are localized. Then according to the location of the sources, the amplitude of the mixture signal spectrum is weighted. By multiplying the weighted spectrum by the values obtained from the angular spectrograms, a binary mask is constructed for each source. By applying the binary mask to the amplitude of the mixture signal spectrum, the speech sources are separated. This method is evaluated on SiSEC database and the measurement tools and criteria contained in this database are used for evaluation. The results show that the proposed method is comparable in terms of the criteria available in the database to the competing ones, has lower computational complexity.