جستجوی مقالات مرتبط با کلیدواژه « speech recognition » در نشریات گروه « فنی و مهندسی »

This article is a review of traditional and modern methods of speech recognition. Speech recognition has a history of several decades and started with methods based on signal processing and dynamic time warping. Statistical methods were noticed and welcomed in the 1980s and the methods based on the hidden Markov models were known as the leading methods. Since the 2000s, statistical methods gradually gave way to models based on neural networks, and with the use of deep neural networks, resulted in higher performances compared to the hidden Markov models. Models based on deep neural networks also were transformed and improved immensely. In the next step, models based on transformers and pre-trained models were proposed and achieved higher accuracies. In this article, after an overview of the methods based on the hidden Markov models, the methods based on deep neural networks and their various structures are discussed, and finally, the methods based on pre-trained models are explained and the latest methods of this kind are surveyed. Finally, the results obtained from the reviewed methods are presented and compared based on the word error rate measure.

Islamic Republic of Iran Broadcasting (IRIB) as one of the biggest broadcasting organizations, produces thousands of hours of media content daily. Accordingly, the IRIBchr('39')s archive is one of the richest archives in Iran containing a huge amount of multimedia data. Monitoring this massive volume of data, and brows and retrieval of this archive is one of the key issues for this broadcasting. The aim of this research is to design a content retrieval engine for the IRIB’s media and production using spoken term detection (STD) or keyword spotting. The goal of an STD system is to search for a set of keywords in a set of speech documents. One of the methods for STD is using a speech recognition system in which speech is recognized and converted into text and then, the text is searched for the keywords. Variety of speech documents and the limitation of speech recognition vocabulary are two challenges of this approach. Large vocabulary continuous speech recognition systems (LVCSR) usually have limited but large vocabulary and these systems canchr('39')t recognize out of vocabulary (OOV) words. Therefore, LVCSR-based STD systems suffer OOV problem and canchr('39')t spotting the OOV keywords. Methods such as the use of sub-word units (e.g., phonemes or syllables) and proxy words have been introduced to overcome the vocabulary limitation and to deal with the out of vocabulary (OOV) keywords. This paper proposes a Persian (Farsi) STD system based on speech recognition and uses the proxy words method to deal with OOV keywords. To improve the performance of this method, we have used Long Short-Term Memory-Connectionist Temporal Classification (LSTM-CTC) network. In our experiments, we have designed and implemented a large vocabulary continuous speech recognition systems for Farsi language. Large FarsDat dataset is used to train the speech recognition system. FarsDat contains 80 hours voices from 100 speakers. Kaldi toolkit is used to implement speech recognition system. Since limited dataset, Subspace Gaussian Mixture Models (SGMM) is used to train acoustic model of the speech recognition. Acoustic model is trained based context tri-phones and language model is probability tri-gram words model. Word Error Rate (WER) of Speech recognition system is 2. 71% on FARSDAT test set and also 28.23% on the Persian news collected from IRIB data. Term detection is designed based on weighted finite-state transducers (WFST). In this method, first a speech document is converted to a lattice by the speech recognizer (the lattice contains the full probability of speech recognition system instead of the most probable one), and then the lattice is converted to WFST. This WFST contains the full probability of words that speech recognition computed. Then, text retrieval is used to index and search over the WFST output. The proxy words method is used to deal with OOV. In this method, OOV words are represented by similarly pronunciation in-vocabulary words. To improve the performance of the proxy words methods, an LSTM-CTC network is proposed. This LSTM-CTC is trained based on charterers of words separately (not a continuous sentence). This LSTM-CTC recomputed the probabilities and re-verified proxy outputs. It improves proxy words methods dues to the fact that proxy words method suffers false alarms. Since LSTM-CTC is an end-to-end network and is trained based on the characters, it doesnchr('39')t need a phonetic lexicon and can support OOV words. As the LSTM-CTC is trained based on the separate words, it reduces the weight of the language model and focuses on acoustic model weight. The proposed STD achieve 0.9206 based Actual Term Weighted Value (ATWV) for in vocabulary keywords and for OOV keywords ATWV is 0.2 using proxy word method. Applying the proposed LSTM-CTC improves the ATWV rate to 0.3058. On Persian news dataset, the proposed method receives ATWV of 0.8008.

Deep neural networks have been widely used in speech recognition systems in recent years. However, the robustness of these models in the presence of environmental noise has been less discussed. In this paper, we propose two approaches for the robustness of deep neural networks models against environmental additive noise. In the first approach, we propose to increase the robustness of long short-term memory (LSTM) networks in the presence of noise based on their abilities in learning long-term noise behavior. For this purpose, we propose to use noisy speech for training models. In this way, LSTMs are trained in a noise-aware manner. The results on the noisy TIMIT dataset show that if the models are trained with noisy speech rather than clean speech, recognition accuracy will be improved up to 18%. In the second approach, we propose to reduce noise effects on the extracted features using a denoised autoencoder network and to use the bottleneck features to compress the feature vector and represent the higher level of input features. This method increases the accuracy of the proposed recognition system in the first approach by 4% in the presence of noise.

Although, speech recognition systems are widely used and their accuracies are continuously increased, there is a considerable performance gap between their accuracies and human recognition ability. This is partially due to high speaker variations in speech signal. Deep neural networks are among the best tools for acoustic modeling. Recently, using hybrid deep neural network and hidden Markov model (HMM) leads to considerable performance achievement in speech recognition problem because deep networks model complex correlations between features. The main aim of this paper is to achieve a better acoustic modeling by changing the structure of deep Convolutional Neural Network (CNN) in order to adapt speaking variations. In this way, existing models and corresponding inference task have been improved and extended.
Here, we propose adaptive windows convolutional neural network (AWCNN) to analyze joint temporal-spectral features variation. AWCNN changes the structure of CNN and estimates the probabilities of HMM states. We propose adaptive windows convolutional neural network in order to make the model more robust against the speech signal variations for both single speaker and among various speakers. This model can better model speech signals. The AWCNN method applies to the speech spectrogram and models time-frequency varieties.
This network handles speaker feature variations, speech signal varieties, and variations in phone duration. The obtained results and analysis on FARSDAT and TIMIT datasets show that, for phone recognition task, the proposed structure achieves 1.2%, 1.1% absolute error reduction with respect to CNN models respectively, which is a considerable improvement in this problem. Based on the results obtained by the conducted experiments, we conclude that the use of speaker information is very beneficial for recognition accuracy.

Today, maps are commonly used for navigation purposes in driving and bicycling. In such situations, the interaction of a user with the map is impossible or, at least, dangerous. Moreover, individuals who have difficulties with their vision, may not see the icons and signs properly. As conversation is a natural way of interacting among people, audio inputs can reduce the existing complexities within general-purpose geospatial information systems (GIS). Therefore, the development of a user-oriented GIS which is able to interact with visually impaired individuals and drivers is important. In this paper, we first review the existing navigation systems, focusing on their user interface (UI). Then, the prototype of an audio-enabled open-source web-based GIS is presented. The users of this WebGIS can utilize its capabilties only using audio commands, without the need to use their hands. In order to establish a connection with the browser's speech recognition engine, the annyang library has been used. An audio user interface has some advantages over casual interfaces, namely being hands-free, the compatibility with up-to-date web technologies, not requiring additional hardware, and consequently being low-cost. Beside the audio user interface, a graphical user interface (GUI) has also been designed. Hence, the user can interact with both user interfaces simultaneously.

Speech recognition as one of the important branches of speech processing has been attractive for researchers and scientist, from long time ago. Speech recognition is a kind of technology able to determine the pronounced word (s) shown by acoustic signal. The complexity of speech recognition systems depends on the extracted features, their dimensions and the applied classifier. In this paper, we propose a new classifier which is able to compute two matrices “winner” and “minimum distance” in a knowledge extraction phase, as a suitable model for any reference word using synergy clustering and frequency of observations. In the recognition phase, the proposed method is able to determine the similarity between inputted unknown speech and word reference models based on a penalty-reward mechanism. In order to evaluate the proposed method, the FARSDAT data set is used. The results of several experiments on clean and noisy signals show more resistant against noise, higher accuracy and less time complexity for the proposed method, in comparison to the HMM-based speech recognition system.

In this article a pre-processing method is introduced which is applicable in speech recognized texts retrieval task. We have a text corpus that generated from a speech recognition system and a query as inputs, want to search queries in these documents and find relevant documents. The main problem is that the typical speech recognized texts suffer from some percentage of recognition error. This problem causes terms to have erroneously assign to irrelevant documents.
The idea of our proposed method, is to detect error-prone terms and to find similar words for each term. A parameter is defined which calculate the probability for occurring error in the error-prone words. To recognize similar words for each specific term, based on a criterian which is called average detection rate (ADR) and levenshtein distance criterion, some candidates are chosen as the initial similar words set. Then, a conversion probability is defined based on the conversion rate (CR) and the noisy channel model (NCM) and the words with higher probability based on a threshold level are selected as the final similar words. In the retrieval process, these words are considered in the search step in addition to the base word. Implementation result shows a significant improvement up to 30% in F-measure in information retrieval method with consideration this pre-processing.

The performance of automatic speech recognition (ASR) systems is adversely affected by the variations in speakers, audio channels and environmental conditions. Making these systems robust to these variations is still a big challenge. One of the main sources of variations in the speakers is the differences between their Vocal Tract Length (VTL). Vocal Tract Length Normalization (VTLN) is an effective method introduced to cope with this variation. In this method, the speech spectrum of each speaker is frequency warped according to a specific warping factor of that speaker. In this paper, we first developed the common search-based method to obtain the appropriate warping factor over a HMM-based Persian continuous speech recognition system. Then pointing out the computational cost of search-based method, we proposed a linear regression process for estimating warping factor based on the scores generated by our gender detection system. Experimental results over a Persian conversational speech database shown an improvement about 0.54 percent in word recognition accuracy as well as a significant reduction in computational cost of estimating warping factor, compared to search-based approach.

Temporal Pattern feature of a speech signal could be either extracted from the time domain or via their front-end vectors. This feature includes long-term information of variations in the connected speech units. In this paper، the second approach is followed، i. e. the features which are the cases of temporal computations، consisting of Spectral-based (LFBE) and Cepstrum-based (MFCC) feature vectors، are considered. To extract these features، we use posterior probability-based output of the proposed MTMLP neural networks. The combination of the temporal patterns، which represents the long-term dynamics of the speech signal، together with some traditional features، composed of the MFCC and its first and second derivatives are evaluated in an ASR task. It is shown that the use of such a combined feature vector results in the increase of the phoneme recognition accuracy by more than 1 percent regarding the results of the baseline system، which does not benefit from the long-term temporal patterns. In addition، it is shown that the use of extracted features by the proposed method gives robust recognition under different noise conditions (by 13 percent) and، therefore، the proposed method is a robust feature extraction method.

Recent researches show that nonlinear and chaotic behavior of the speech signal can be studied in the reconstructed phase space (RPS). Delay embedding theorem is a useful tool to study embedded speech trajectories in the RPS. Characteristics of the speech trajectories have rarely used in the practical speech recognition systems. Therefore, in this paper, a new feature extraction (FE) method is proposed based on parameters of vector AR (VAR) analysis over the speech trajectories. In this method, using filter and reflection matrices obtained from applying VAR analysis on static and dynamic information of the speech trajectory in the RPS, a high-dimensional feature vector can be achieved. Then, different transformation methods are utilized to attain final feature vectors with appropriate dimension. Results of discrete and continuous phoneme recognition over FARSDAT speech corpus show that the efficiency of the proposed FE method is better than other time-domain-based FE methods such as LPC and LPREF.

One of the most important challenges in automatic speech recognition is in the case of difference between the training and testing data. To decrease this difference, the conventional methods try to enhance the speech or use the statistical model adaptation. Training the model in different situations is another example of these methods. The success rate in these methods compared to those of cognitive and recognition systems of human beings seems too much primary. In this paper, an inspiration from human being's recognition system helped us in developing and implementing a new connectionist lexical model. Integration of imputation and classification in a single NN for ASR with missing data was investigated. This can be considered as a variant of multi-task learning because we train the imputation and classification tasks in parallel fashion. Cascading of this model and the acoustic model corrects the sequence of the mined phonemes from the acoustic model to the desirable sequence. This approach was implemented on 400 isolated words of TFARSDAT Database (Actual telephone database). In the best case, the phoneme recognition correction increased in 16.9 percent. Incorporating prior knowledge (high level knowledge) in acoustic-phonetic information (lower level) can improve the recognition. By cascading the lexical model and the acoustic model, the feature parameters were corrected based on the inversion techniques in the neural networks. Speech enhancement by this method had a remarkable effect in the mismatch between the training and testing data. Efficiency of the lexical model and speech enhancement was observed by improving the phoneme's recognition correction in 18 percent compared to the acoustic model.