Search

local feature information at shallow layers. To solve this problem, this study proposes a method using multiple encoders to improve the ability of speech feature extraction. An additional convolutional encoder branch is added to strengthen the capture of local feature information, make up for the neglect of local feature information in shallow Transformer, and effectively realize the integration of global and local dependencies of andio feature se quences. In other words, a multi-encoder model hased on Transformer is proposed. Experiments on the open source Chinese Mandarin data set Aishell 1 show that without an external language model, the proposed Transformer-based multi-encoder model has a relative reduction of 4.00% in character error rate when compared with the Transformer model. On the internal non public Shanghainese dialect data set, the performance improve ment of the proposed model is more obvious, and the character error rate is reduced by 48.24% from 19.92% to 10. 31%. [ABSTRACT FROM AUTHOR]

Due to the decrease in accuracy of speech recognition software in noisy environments, a robust enhancement method based on domain knowledge is proposed to ensure the safety of using speech control operations. Taking ship control as the application background, a domain knowledge graph is established for ship control. Ship control commands are extracted from nautical books and classic naval warfare film and television materials, and a Chinese speech dataset for ship control commands is constructed. A domain knowledge-embedded decoding method is proposed to correct the output control commands by calculating the matching degree between the recognition result and the domain knowledge graph. Experimental results show that compared with the current popular connection time sequence classification decoding method and attention mechanism-based decoding method, the proposed decoding method reduces the word error rate by 4.0% and 1.5% when recognizing noisy speech with a signal-tonoise ratio of 10dB and 20dB, respectively, and improves the accuracy of command recognition by 10.3% and 6.3%, respectively, improving the robustness of the speech recognition model in recognizing Chinese commands. [ABSTRACT FROM AUTHOR]

To solve the problems that the Attention's overly flexible alignment was poorly adaptable in complex environments and the language features were not fully utilized by simple end-to-end models, a speech recognition method was investigated based on multi-task loss with additional language model. By analyzing the characteristics of the speech signal, the features containing more information were selected in the training. Based on the Attention-based Conformer end-to-end model, the model was trained using multi-task loss of CTC loss assisted pure Conformer (Attention), and the Conformer-CTC speech recognition model was obtained. Based on the Conformer-CTC model, by analyzing and comparing the characteristics and effects of some language models, the Transformer language model was added to the training of the above model through re-scoring mechanism, and the Conformer-CTC-Transformer speech recognition model was obtained. The experiments on the above model were completed on the AISHELL-1 data set. The results show that compared with the pure Conformer (Attention) model, the character error rate (CER) of the Conformer-CTC model on the test set is reduced by 0. 49%, and the CER of the Conformer-CTC-Transformer model on the test set is reduced by 0.79% compared with the Conformer- CTC model. The adaptability of Attention alignment in complex environments can be improved by CTC loss, and after re-scoring the Transformer-CTC model with the Transformer language model, the recognition accuracy can be increased by 0. 30% again. Compared with some existing end-to-end models, the recognition effect of the Conformer-CTC-Transformer model is better, indicating that the model has certain effectiveness. [ABSTRACT FROM AUTHOR]

Recently, deep neural network model has become a hot research object in the field of speech recognition. However, the deep neural network relies on huge parameters and computational overhead, the excessively large model size also increases the difficulty of its deployment on edge devices. Aiming at the above problems, this paper proposed a lightweight speech recognition model based on Transformer. This method used depthwise separable convolution to obtain the feature information. Secondly, this method constructed a two half-step feed-forward layers, namely Macaron-N et, and introduced the lowrank matrix factorization to realize the model compression. Finally, it used a sparse attention mechanism to improve the training speed and decoding speed of the model. It tested on the Aishell-1 and aidatang_200zh datasets. The experimental results show that compared with Open-Transformer, the word error rate and real time factor of LM-Transformer decrease by 19. 8% and 32.1%, respectively. [ABSTRACT FROM AUTHOR]

In a continuous speech recognition system, aiming at the complex environments (including the variability of speakers and environmental noise), the training data does not match the test data, which results in a low voice recognition rate. A speech recognition method based on adaptive deep neural network is studied. The improved regularized adaptive criterion and the adaptive deep neural network in the feature space are combined to improve data matching. The fusion of speaker identity vector i-vector and noise perception training are used to overcome speaker and environmental noise changes and improve the classification function of the output layer of the traditional deep neural network, which ensures the characteristics of compactness within the class and separation between classes. The test experiment was carried out by superimposing various background noises under the TIMIT English speech data set and the Microsoft Chinese speech data set. The results show that, compared with the current popular GMMHMM and traditional DNN speech acoustic models, our proposal decreases the recognition word error rate by 5.151% and 3.113% respectively, which improves the generalization performance and robustness of the model to a certain extent. [ABSTRACT FROM AUTHOR]

To improve the performance of multi-accent Mandarin speech recognition task,this paper presented a method for hybrid end-to-end automatic speech recognition(ASR) by combining CTC and multi-head attention by using a multiobjective training and joint decoding.The analysis shows that hybrid model with lower CTC weight and deeper encoder layers performance better learning capacity.And it trained a very deep models with up to 48 layers for encode-decoder architecture,which outperforms all previous end-to-end ASR approaches on Aidatatang 200 h multi-accent dataset,achieves 5.6% character error rate(CER) and 26.2% sentence error rate(SER).The experiment proves that the recognition rate of the end-to-end model proposed exceeds the general end-to-end model,and it has certain advancedness in solving the Mandarin recognition with accents. [ABSTRACT FROM AUTHOR]

In the era of big data, many of traditional machine learning methods of disposing unlabeled raw voice data have become less applicable. At the same time, deep learning models can directly process unlabeled data because of its powerful capability of modeling to deal with the massive data, and has become a hot research in the field of speech recognition. To begin with, this paper analyzed and summarized the state-of-the-art deep learning of models. And then,it discussed the applications to speech recognition with speech features extraction and acoustic modeling. Finally,it concluded the problems faced and development orientation. [ABSTRACT FROM AUTHOR]

Based on the Mandarin pronunciation recognition technology, by the frequency selected from Mandarin speech input cavity resonance curve defined by the pickup, the Chinese word, vowel and consonant Pinyin word sequence can be identified. Combination of Chinese speech consonants, vowels and tones forms Chinese Pinyin sentence sequence. By the syllable by syllable analysis method, the frequency domain analysis technique of amplitude frequency does not need to learn or train for a particular persons voice, words and sentences. The adaptive DFT ( Discrete Fourier Transform) spectral analysis algorithm only occupies a very small amount of calculation. The method can determine the four tones of Mandarin, have fast response speed, small storage space, and can easilybe embedded to other devices. [ABSTRACT FROM AUTHOR]

Under the condition of low resource, little labeled training data is available and the performance of speech recognition system is not ideal. To solve this problem. First, this paper investigated long short term memory recurrent neural network ( LSTM RNN) for acoustic modeling. It was a powerful tool to model long time series and could make full use of the context information. Linear projection layer reduced the number of model parameters. Then, it explored speaker modeling methods in the feature space, and extracted identity vector (i-vector) which contained the speaker and channel information simultaneously. Finally, it presented a novel system, which combined the LSTM RNN model and i-vector feature. Results on the standard Open KWS 2013 data set show that this technology produces a relative improvement of about 10% in TER over the DNN baseline system. [ABSTRACT FROM AUTHOR]

To handle the speaker and noise adaptation problem in deep neural network-based speech recognition system, this paper studied the inherent characters of speaker and noise random factors and proposed a new adaptation method using long term features. Firstly, it built a joint adaptation model based on Gaussian mixture models and estimated and used the parameters of speaker and noise factors as long term features. Then, it used these long term features in deep neural network together with traditional short term features. Experiment results on Aurora4 database show that this method can effectively factorize speaker and noise factors, and improve adaptation performance. [ABSTRACT FROM AUTHOR]

The impact of changes in a speaker's vocal effort on the performance of automatic speech recognition is the research target. Firstly, this paper performed analysis for sound intensity level, duration, frame energy distribution and spectral tilt, and proposed vocal effort detector based on GMM by using these phonetic properties and acoustic properties. Then, it studied impact of vocal effort variability on automatic speech recognition, and presented multiple model framework approach to improve robustness of speech recognition system. Experiments conducted on isolated-word recognizer show that accompanied by a slight decline for normal mode, significantly improvement of recognition accuracy for the remaining four vocal effort modes can be achieved. Thus, potential of the method is demonstrated. [ABSTRACT FROM AUTHOR]

To improve the performance of the conventional speech recognition system, this paper introduced the autoencoder deep learning neural networks which was applied to speech recognition. The neural networks based on deep learning introduced greedy layer-wise learning algorithm by pretraining and fine-tuning. It could extract the essential features of speech signal which was needed to recognition. It could overcome the shortcomings of the conventional multilayer artificial neural networks which easily trapped into local optimum when training the model. And they needed a large number of labeled data. Then the structured alignment networks could align arbitrary frames of features to fixed frames. And it input these features to a classifier to speech recognition. This paper did some experiment with back propagation neural networks and autoencoder neural networks respectively. The results illustrate that the deep learning neural networks can outperform the conventional neural networks by 20.0% in accuracy. It is an excellent speech recognition model. [ABSTRACT FROM AUTHOR]

This paper proposed a framework which attempted to incorporate articulatory information into the stochastic segment model based mandarin speech recognition system. According to the characteristics of the stochastic segment model, it used hierarchical artificial neural network to obtain the posteriors of speech signal belonging to the phonemes. Then, the posteriors were integrated into the stochastic segment model system in the first search process. Experiments conducted on "863-test" set show that about 5.93% relative improvement can be achieved in the recognition accuracy. Thus, potential of the method is demonstrated. [ABSTRACT FROM AUTHOR]

In order to meet the requirement environment in which both speech and speaker's group were simultaneously recognized through embedded system, this paper proposed a new method with low space consumption for simultaneous recognition of both speech and speaker's group. On the basis of research on mechanism of speech recognition and speaker identification in recent years, GMM instead of HMM, was used to model the voice characteristics of speech and speaker and voting mechanism to complete the speaker's group identification to reduce space occupation. SQ(soft quantization) was used to integrate the outputs from multiple speech recognizer. When integrating 6 models, it reduced 20. 88% error rate in speech recognition and achieved 81.57% average recognition rate in speaker's group. The experimental results confirm feasibility of the proposed algorithm. [ABSTRACT FROM AUTHOR]

This paper proposed a framework which attempted to incorporate articulatory information into the stochastic segment model based on Mandarin speech recognition system. According to the characteristics of the stochastic segment model, it used hierarchical artificial neural network to obtain the posterior probability of speech signal belonging to the phonemes. Then, it integrated the posterior probability into the stochastic segment model system in the first search process. Experiments conducted on "863-test" set show that about 5.93% relative improvement could be achieved in the recognition accuracy. Thus, it demonstrates the feasibility of the method. [ABSTRACT FROM AUTHOR]

A speaker adaptation method of Stochastic Segment Model(SSM)is proposed. According to the SSM's characteristics, the theory of Maximum Likelihood Linear Regression(MLLR)method is introduced into the SSM-based systems. Continuous Chinese speech recognition experiment on "863-test" test suite shows that the proposed method makes the error rate of Chinese characters decrease obviously under different decoding speeds. Experiment results indicate that the proposal can also improve the recognition performance on the SSM-based systems. [ABSTRACT FROM AUTHOR]

It is an urgent need to improve the innovation ability of college students in the modern society, and robot practice platform is the best engineering project. Therefore, the course design of MCU takes "robots control with speaker-independent voice"as subject. The students work in group and design the experiment by themselves to go through the prescribed and alternative experiment modules. Purpose is to achieve the robots control with voice in the engineering practice. It is the optimal combination of teaching and scientific research, and matches the idea of teaching innovation. The practice of a group is taken as the example, it is proved that this innovative course design of "robots control with speaker-independent voice " can lay a solid foundation of the theoretical knowledge and enhance the students' practical ability, and can also lay the experimental foundation for innovation practice platform of robots in college. [ABSTRACT FROM AUTHOR]