ASR System Research Articles

Automatic speech recognition system for Tunisian dialect

Although Modern Standard Arabic is taught in schools and used in written communication and TV/radio broadcasts, all informal communication is typically carried out in dialectal Arabic. In this work, we focus on the design of speech tools and resources required for the development of an Automatic Speech Recognition system for the Tunisian dialect. The development of such a system faces the challenges of the lack of annotated resources and tools, apart from the lack of standardization at all linguistic levels (phonological, morphological, syntactic and lexical) together with the mispronunciation dictionary needed for ASR development. In this paper, we present a historical overview of the Tunisian dialect and its linguistic characteristics. We also describe and evaluate our rule-based phonetic tool. Next, we go deeper into the details of Tunisian dialect corpus creation. This corpus is finally approved and used to build the first ASR system for Tunisian dialect with a Word Error Rate of 22.6%.

Web Recognition of Spoken Hindi

Technology has evolved and computers have but still Indian communities are far from the use of computers, only 37% [13] user of Indian society like persons from academics, health, engineering and research make use of the benefits of computer. Although there is a revolution in development of operating systems in the past two years, there are no operating systems that support Indian languages like Hindi, they only support English language and people who know English can work on these. This work shows the development of speaker independent ASR system for Indian speaker of a particular type i.e. spoken Hindi. Objective: An automated system is developed which accept Hindi query from the user and search the keywords from the query on web and shows the result accordingly. A bilingual lexicon of 3145 words was built. The techniques to convert acoustic speech signals into strings of words were also designed. The dictionary maintains the POS of source to target language and CFG grammar based model for Hindi language was developed. Methods: A speech database consisting of speech files is created at the rate of 16 KHz. Interviews were conducted uto record speech files from users for developing the speech database for training the system. The system takes spoken Hindi as input and generates Hindi sentences. Hindi sentences are parsed with the help of Top-Down Paser and identify the keywords and form a exact query with the help of grammar rules which are given as input to Google search engine to search the result as per query and display on screen. The lexical categories of Hindi language in grammar are done on the bases of verb, noun, preposition, adjective, adverb phrases. Result: A total of 101 speakers from Faridabad district, comprising of both males and females took part in data collection process. The system was tested in a sound proof room and achieved 88.05% accuracy. It is observed that utilization of more training data together with detailed modeling of speech signals can improve the system performance to a level adequate for actual deployment. Application/Improvements: This system provides access to a computer system or to digital content over the internet to illiterate, vision-impaired, urban and semi-urban Indian people who are not able to read/write English language. The same speech interface can be enhanced to work in other regional languages of India like Punjabi, Marathi and Telugu, etc. To enhance or to extend it for other languages, the transfer or translation rules of grammar is needed, which can be generated with great ease by using the same dataset with different target languages. Keywords: Corpus, Hindi, Recognition, Speech, Web

The Effect Of Changes In Speech Features On The Recognition Accuracy Of ASR System: A Study On The Malay Speech Impaired Children

Speech impairments refers to disability that causes the human speech production to deviate from the norm. Although there have been several researches undertaken to identify the differences between non-impaired and impaired speech, little is known about their effects on the speech intelligibility and the performance of ASR systems in recognizing impaired speech of children. This study investigates the speech features of impaired speech in relation to intelligibility deficits and degradation in ASR performance; which includes, formant frequencies, intensity, fundamental frequency (F0) and perturbation features such as jitter and shimmer. As there is no existing speech database for performing the evaluation, we have developed a speech database of speech impaired children and have analysed the impaired speech features. We have identified significant differences in the selected features. We also have identified the relationship between the ASR system’s Word Error Rate (WER) of impaired speeches with the speech features. The results show that there are significant differences in F0, jitter and shimmer across the Control Group (CG) and the Speech Impaired Group (SIG). This paper explains the differences between impaired speeches and non-impaired speeches that can be used in developing automated speech recognition system. We have observed that F0 affects the ASR performance and was found to be a significant predictor that influences the accuracy of vowel phonemes /e/ and /u/.

Bayesian Unsupervised Batch and Online Speaker Adaptation of Activation Function Parameters in Deep Models for Automatic Speech Recognition

We present a Bayesian framework to obtain maximum a posteriori MAP estimation of a small set of hidden activation function parameters in context-dependent-deep neural network-hidden markov model CD-DNN-HMM-based automatic speech recognition ASR systems. When applied to speaker adaptation, we aim at transfer learning from a well-trained deep model for a “general” usage to a “personalized” model geared toward a particular talker by using a collection of speaker-specific data. To make the framework applicable to practical situations, we perform adaptation in an unsupervised manner assuming that the transcriptions of the adaptation utterances are not readily available to the ASR system. We conduct a series of comprehensive batch adaptation experiments on the Switchboard ASR task and show that the proposed approach is effective even with CD-DNN-HMM built with discriminative sequential training. Indeed, MAP speaker adaptation reduces the word error rate WER to 20.1% from an initial 21.9% on the full NIST 2000 Hub5 benchmark test set. Moreover, MAP speaker adaptation compares favorably with other techniques evaluated on the same speech tasks. We also demonstrate its complementarity to other approaches by applying MAP adaptation to CD-DNN-HMM trained with speaker adaptive features generated through constrained maximum likelihood linear regression and further reduces the WER to 18.6%. Leveraging upon the intrinsic recursive nature in Bayesian adaptation and mitigating possible system constraints on batch learning, we also proposed an incremental approach to unsupervised online speaker adaptation by simultaneously updating the hyperparameters of the approximate posterior densities and the DNN parameters sequentially. The advantage of such a sequential learning algorithm over a batch method is not necessarily in the final performance, but in computational efficiency and reduced storage needs, without having to wait for all the data to be processed. So far, the experimental results are promising.

Speech enhancement for robust automatic speech recognition: Evaluation using a baseline system and instrumental measures

1Present address: Cirrus Logic, Marble Arch House, 66 Seymour St., 1st Floor, London W1H 5BT, United Kingdom.Automatic speech recognition in everyday environments must be robust to significant levels of reverberation and noise. One strategy to achieve such robustness is multi-microphone speech enhancement. In this study, we present results of an evaluation of different speech enhancement pipelines using a state-of-the-art ASR system for a wide range of reverberation and noise conditions. The evaluation exploits the recently released ACE Challenge database which includes measured multichannel acoustic impulse responses from 7 different rooms with reverberation times ranging from 0.33 to 1.34 s. The reverberant speech is mixed with ambient, fan and babble noise recordings made with the same microphone setups in each of the rooms. In the first experiment, performance of the ASR without speech processing is evaluated. Results clearly indicate the deleterious effect of both noise and reverberation. In the second experiment, different speech enhancement pipelines are evaluated with relative word error rate reductions of up to 82%. Finally, the ability of selected instrumental metrics to predict ASR performance improvement is assessed. The best performing metric, Short-Time Objective Intelligibility Measure, is shown to have a Pearson correlation coefficient of 0.79, suggesting that it is a useful predictor of algorithm performance in these tests.

A Review of Challenges in Automatic Speech Recognition

is the nature's gift to the human being which contributes towards the intelligence and discrimination from rest of the animal kingdom. Taking into consideration technological aspects, speech recognition is the buzzword today, as communication and hands free computing evolving day by day. Speech is a very important mode of the communication and interaction with the digital computer. Speech recognition along with the wide range of applicability in domain of computer science, medical science, psychology, sports, neurology has many challenges while developing. Developing real time speech recognizer may hurdle from adverse environment to anatomy of the human body. It also involves linguistic aspects too. This paper explores various challenges in developing a robust ASR system. General Terms Signal Processing and linguistics

Human-inspired modulation frequency features for noise-robust ASR

This paper investigates a computational model that combines a frontend based on an auditory model with an exemplar-based sparse coding procedure for estimating the posterior probabilities of sub-word units when processing noisified speech. Envelope modulation spectrogram (EMS) features are extracted using an auditory model which decomposes the envelopes of the outputs of a bank of gammatone filters into one lowpass and multiple bandpass components. Through a systematic analysis of the configuration of the modulation filterbank, we investigate how and why different configurations affect the posterior probabilities of sub-word units by measuring the recognition accuracy on a semantics-free speech recognition task. Our main finding is that representing speech signal dynamics by means of multiple bandpass filters typically improves recognition accuracy. This effect is particularly noticeable in very noisy conditions. In addition we find that to have maximum noise robustness, the bandpass filters should focus on low modulation frequencies. This reenforces our intuition that noise robustness can be increased by exploiting redundancy in those frequency channels which have long enough integration time not to suffer from envelope modulations that are solely due to noise. The ASR system we design based on these findings behaves more similar to human recognition of noisified digit strings than conventional ASR systems. Thanks to the relation between the modulation filterbank and procedures for computing dynamic acoustic features in conventional ASR systems, the finding can be used for improving the frontends in those systems.

The Sliding Mode Control about ASR of Vehicle with Four Independently Driven In-Wheel Motors Based on the Exponent Approach Law

Acceleration slip regulation control system is a new active safety technology. In this paper, through the research of the four-wheel independent drive electric vehicle, a sliding mode variable structure control algorithm based on exponent approach law is proposed, which is applied to the ASR system. This paper establishes a seven DOFs vehicle dynamics model, tests whether the ASR control strategy is efficient on the poor condition road. The simulation results show that the vehicle acceleration performance improvement rate increases by 43.5% and 58.5% with the control strategy. During the two simulation processes, the results indicate that the sliding mode variable structure control algorithm applied to ASR system has a good adaptation to good and slippery roads. The algorithm can greatly improve the four-wheel independent drive electric vehicle's acceleration performance.

Combination of multiple acoustic models with unsupervised adaptation for lecture speech transcription

Automatic speech recognition systems (ASR) have achieved considerable progress in real applications because of skilled design of the architecture with advanced techniques and algorithms. However, how to design a system efficiently integrating these various techniques to obtain advanced performance is still a challenging task. In this paper, we introduced an ensemble model combination and adaptation based ASR system with two characteristics: (1) large-scale combination of multiple ASR systems based on a Recognizer Output Voting Error Reduction (ROVER) system, and (2) multi-pass unsupervised speaker adaptation for deep neural network acoustic models and topic adaptation on language model. The multiple acoustic models were trained with different acoustic features and model architectures which helped to provide complementary and discriminative information in the ROVER process. With these multiple acoustic models, a better estimation of word confidence could be obtained from ROVER process which helped in selecting data for unsupervised adaptation on the previously trained acoustic models. The final recognition result was obtained using multi-pass decoding, ROVER, and adaptation processes. We tested the system on lecture speeches with topics related to Technology, Entertainment and Design (TED) that were used in the international workshop on spoken language translation (IWSLT) evaluation campaign, and obtained 6.5%, 7.0%, 10.6%, and 8.4% word error rates for test sets in 2011, 2012, 2013, and 2014, which to our knowledge are the best results for these evaluation sets.

A robust BFCC feature extraction for ASR system

An auditory-based feature extraction algorithm naming the Basilar-membrane Frequency-band Cepstral Coefficient (BFCC) is proposed to increase the robustness for automatic speech recognition. Compared to Fourier spectrogram based of the Mel-Frequency Cepstral Coefficient (MFCC) method, the proposed BFCC method engages an auditory spectrogram based on agammachirp wavelet transform to simulate the auditory response of human inner ear to improve the noise immunity. In addition, the Hidden Markov Model (HMM) is used for evaluating the proposed BFCC in phases of training and testing purposes conducted by AURORA-2 corpus with different Signal-to-Noise Ratios (SNRs) degrees of datasets. The experimental results indicate the proposed BFCC, compared with MFCC, Gammatone Wavelet Cepstral Coefficient (GWCC), and Gammatone Frequency Cepstral Coefficient (GFCC), improves the speech recognition rate by 13%, 17%, and 0.5% respectively, on average given speech samples with SNRs ranging from -5 to 20 dB.

Automatic Speech Recognition for African Languages with Vowel Length Contrast

This paper deals with ASR for two languages: Hausa and Wolof. Their common characteristic is to appear with vowel length contrast. In other words, two versions (short/ long) of a same vowel exist in the phoneme inventory of the language. We expect that taking into account this contrast in ASR models might help and this is what we investigate in this pilot study. The experimental results show that while both approaches (vowel length contrast modeling or not) lead to similar results, their combination allows to slightly improve ASR performance. As a by-product of ASR system design, we also show that the acoustic models obtained allow a large scale analysis of vowel length contrast for phonetic studies.

面向双语教学的维吾尔语发音质量自动评估

Pursuing Chinese and Uygur bilingual teaching in XinJiang Province is an important government policy for improving communication between different language speakers. An automatic assessment system can yield many advantages in such bilingual teaching. However, as an agglutinative language, Uygurs particular way of word-building results in leaving many words out of the lexicon. Therefore, performance of the old pronunciation assessment system, which is based on a traditional ASR system, is poor. For building a high-performance system, we decide to use subwords as basic recognition units after analyzing the rules and habits of Uygur pronunciation. Experimental results indicate that the accuracy of the subword-based system is greatly improved by the implementation of machine segmentation based on double-level morphology analysis of Uygur subwords, calculating the posterior probabilitys denominator with a phoneme decoder, and calculating confidence at the subword level.

A Review on Different Approaches for Speech Recognition System

This paper presents the basic idea of speech recognition, proposed types of speech recognition, issues in speech recognition, different useful approaches for feature extraction of the speech signal with its advantage and disadvantage and various pattern matching approaches for recognizing the speech of the different speaker. Now day’s research in speech recognition system is motivated for ASR system with a large vocabulary that supports speaker independent operations and continuous speech in different language.

在人工岛屿和沿海进行淡水存储和恢复的含水层设计

Artificial recharge of water to an aquifer for later recovery and use, otherwise known as artificial storage and recovery (ASR; Pyne 1995), is nowadays widely applied for seasonal or emergency water storage. Therefore, one would expect a lot of thought to be devoted to what the optimal aquifer conditions for ASR would be, but this does not appear to be the case. In some situations local hydrogeology may impact the selection of ASR sites; however, according to Pyne (1995), this is the exception rather than the rule since ASR wells are usually located where they provide the greatest benefit to the water utility or agency. This location is often in the vicinity of the supply area to reduce the costs and time of transportation. At that specific site, one deals with the existing local hydrogeology through well design and construction (e.g., Maliva and Missimer 2010; Zuurbier et al. 2014) and operation (Ward et al. 2009; Bakker 2010). In line with Pyne’s field experience, Lowry and Anderson (2006) distinguished physical properties of the aquifer and operational factors that control the recovery efficiency of ASR. Properties like porosity, hydraulic conductivity, aquifer thickness and density of native water, as well as quality, are regarded as predefined site-specific conditions, while operational factors such as injected volume, location of injection and recovery wells, recharge and recovery rates and storage duration, can be changed at the wellhead by the operator to optimize the ASR system. The most relevant research determining optimal aquifer conditions for ASR focused on the influence of aquifer properties on the recovery efficiency of ASR wells. These studies are, firstly, comparisons between ASR sites, such as the studies executed by e.g., Merritt (1986), Dillon et al. (2006), Lowry and Anderson (2006) and Misut and Voss (2007). On the other hand, there are theoretical considerations regarding dimensionless parameter groups in analytical solutions, as described by e.g., Esmail and Kimbler (1967), Ward et al. (2007, 2008, 2009) and Bakker (2010). The general outcomes of these studies indicate that porosity, hydraulic conductivity, vertical anisotropy, dispersivity, density of native water versus that of the injected water, and the thickness of the aquifer all influence the recovery efficiency. This makes sense because these aquifer properties determine the degree of the underlying recovery efficiency processes of mixing and density stratification. The question of which aquifer conditions are preferable for ASR is usually not relevant in practice because the costs of the required earth displacement and construction works are so high that they generally outweigh the economic benefits of ASR systems. However, the situation is different in the case of coastal expansions and artificial islands that are currently being constructed worldwide. The most well-known examples are probably the Palms and the islands of the World Archipelago in Dubai (UAE) with a mainly touristic purpose. Artificial islands are also constructed for industrial development such as Maasvlakte II in the Port of Rotterdam, the Netherlands. Recently, there is a tendency in densely populated delta cities for urban expansion toward the ocean; examples are Eko Atlantic in the City of Lagos, Nigeria, and the planned land reclamations in Jakarta Bay, Indonesia. Aquifers are, in fact, created in these projects and their conditions can be optimized for specific ASR applications. Specific sediment types may be chosen and different dredging techniques can be applied to create the Boptimal^ aquifer conditions for recharge and recovery in terms of porosity, conductivity, anisotropy, and dispersivity to control mixing processes and density stratification. Received: 1 September 2014 /Accepted: 4 March 2015 Published online: 21 April 2015

Multiple topic identification in human/human conversations

The paper deals with the automatic analysis of real-life telephone conversations between an agent and a customer of a customer care service (ccs). The application domain is the public transportation system in Paris and the purpose is to collect statistics about customer problems in order to monitor the service and decide priorities on the intervention for improving user satisfaction.Of primary importance for the analysis is the detection of themes that are the object of customer problems. Themes are defined in the application requirements and are part of the application ontology that is implicit in the ccs documentation.Due to variety of customer population, the structure of conversations with an agent is unpredictable. A conversation may be about one or more themes. Theme mentions can be interleaved with mentions of facts that are irrelevant for the application purpose. Furthermore, in certain conversations theme mentions are localized in specific conversation segments while in other conversations mentions cannot be localized. As a consequence, approaches to feature extraction with and without mention localization are considered.Application domain relevant themes identified by an automatic procedure are expressed by specific sentences whose words are hypothesized by an automatic speech recognition (asr) system. The asr system is error prone. The word error rates can be very high for many reasons. Among them it is worth mentioning unpredictable background noise, speaker accent, and various types of speech disfluencies.As the application task requires the composition of proportions of theme mentions, a sequential decision strategy is introduced in this paper for performing a survey of the large amount of conversations made available in a given time period. The strategy has to sample the conversations to form a survey containing enough data analyzed with high accuracy so that proportions can be estimated with sufficient accuracy.Due to the unpredictable type of theme mentions, it is appropriate to consider methods for theme hypothesization based on global as well as local feature extraction. Two systems based on each type of feature extraction will be considered by the strategy. One of the four methods is novel. It is based on a new definition of density of theme mentions and on the localization of high density zones whose boundaries do not need to be precisely detected.The sequential decision strategy starts by grouping theme hypotheses into sets of different expected accuracy and coverage levels. For those sets for which accuracy can be improved with a consequent increase of coverage a new system with new features is introduced. Its execution is triggered only when specific preconditions are met on the hypotheses generated by the basic four systems.Experimental results are provided on a corpus collected in the call center of the Paris transportation system known as ratp. The results show that surveys with high accuracy and coverage can be composed with the proposed strategy and systems. This makes it possible to apply a previously published proportion estimation approach that takes into account hypothesization errors.

Audio visual isolated Hindi digits recognition using HMM

Automatic speech recognition ASR system performs well under restricted conditions but the performance degrades under noisy environment. Audio-visual features play an important role in ASR systems in presence of noise. In this paper, Hindi isolated digits recognition system is designed using audio visual features. The visual features of the lip region integrated with audio features to get better recognition performance under noisy environments. Colour intensity and pseudo hue methods have been used for lip localisation approach with hidden Markov model HMM as a classifier. Recognition performance using HMM is better than LDA recogniser. For image compression, principal component analysis technique has been utilised.

Simulated effects of aquifer heterogeneity on ASR system performance

All aquifers have some degrees of heterogeneity, which if extreme, can adversely affect the performance of ASR systems. There are many types of aquifer heterogeneity that may be encountered in the field. Two of them are considered here: non-uniform distribution of transmissivity in the storage zone and dual porosity. As indicated by the flowmeter log data, the hydraulic conductivity values may vary by several orders of magnitude within the target storage zone. The heterogeneity introduced by the variable distribution of flow zones may affect the performance of an ASR system. Dual-porosity conditions are often encountered in storage aquifers, which can significantly impact the recovery efficiency of an ASR system. Several cases of heterogeneity and their potential impacts on ASR system recovery efficiency were investigated by simulations of hypothetical ASR systems with reasonable input parameters. Modeling results indicate that aquifer heterogeneity will reduce the performance of ASR systems, compared to systems using a homogeneous aquifer. Existence of dual-porosity conditions has a tendency to lower the recovery efficiency of an ASR system when freshwater is injected into the native brackish water due to mixing and differential solute travel time. Buoyancy effects are more pronounced in the fractured domain than in a homogenous system. This resulted in a wider freshwater plume in the shallower layers within the fractured domain. The simulation results also suggest that recovery efficiency can be improved by setting the extraction point shallower than the injection point.

Drive and Brake Joint Control of Acceleration Slip Regulation Road Test

Based on the independent design front wheel drive vehicle traction control system (TCS), we finished the two kinds of working condition winter low adhesion real vehicle road test, including homogenous pavement and separate pavement straight accelerate, respectively completed the contrastive experiment with TCS and without TCS. Test results show that based on driver (AMR) and brake (BMR) joint control ASR system worked reliably, controlled effectively, being able to control excessive driving wheel slip in time, effectively improved the driving ability and handling stability of vehicle.

Low Complexity On-Line Adaptation Techniques in Context of Assamese Spoken Query System

In this work, we present the development of an Assamese spoken query (SQ) system for accessing the price of agricultural commodities. The developed system intends to make the cultivators aware of the recent market trends. The SQ system enables the user to access the latest price of the commodity by calling the system using a landline/mobile phone. The spoken query input by the user is processed and then current price of the desired commodity in the given district is played back by the system. Features that make the system user friendly are incorporated into the design after taking feedbacks from local farmers. In other words, the system is tuned as per the needs of the users. Furthermore, the issues of adapting such query systems to the end user are also explored in this work. In case of the developed SQ system, the typical user responses are of extremely small duration (1---2 seconds due to isolated word response from the user). Moreover, the employed adaptation approach must keep the system latency low since these systems are meant for real-time applications. Consequently, adapting such systems to the end user becomes an extremely challenging task. In this regard, acoustic model interpolation based adaptation techniques are proposed that employ interpolation weights derived in an approximate fashion. The proposed approaches try to minimize the latency in the system response by avoiding the iterative weight estimation procedure used in the earlier reported works. Even with extremely small amount of adaptation data, the proposed approaches are found to result in a relative improvement of 12 % over the baseline ASR system.

Design and Development of Automatic Speech Recognition of Isolated Marathi Words for Agricultural Purpose

Speech is a natural mode of communication for people. Yet people are so comfortable with speech, they would also interact with the computers via speech, and various interfacing devices such as keyboards and pointing devices. Outstanding work in speech recognition and computing has produced the commercial speech recognition systems for voice driven computing and word-processing systems. The analysis within the space of speech recognition lots of work is done for English language and European language and the opposite hand very little work done for Indian language and really little in Marathi. Only Because of This We Develop Automatic Speech Recognition of Isolates Marathi Words For Agriculture Purpose. For developing ASR system we use hybrid feature extraction technique i.e. MFCC with RASTA and for recognition Dynamic Time Wrapping is used.