ASR

Abstrak Kondisi geologi lokal merupakan salah satu parameter yang mempengaruhi tingkat kerusakan akibat gempabumi. Kondisi geologi kota Bengkulu telah digambarkan oleh nilai Ground Shear Strain (GSS) yang diperoleh dari hasil analisis enam puluh tujuh data mikrotremor yang dipasang di setiap jenis formasi geologi. Berdasarkan hasil analisis, nilai GSS kota Bengkulu relatif heterogen meskipun pada jenis formasi geologi yang sama. GSS kota Bengkulu diperoleh dan dibagi dalam tiga kelompok, diantarany 10-6 (25.37%), 10-5 (40.30%) dan 10-4 (32.84%). Variasi ini mengindikasikan adanya perbedaan ketebalan lapisan sedimen dan tingkat kekerasan batuan untuk setiap titik amat. Wilayah yang memiliki nilai GSS tinggi memiliki kesesuaian dengan kondisi lapangan yang sebenarya (berdasarkan hasil ground survey merupakan daerah bekas rawa dengan mofologi landai) dan kesesuaian dengan sejarah kerusakan akibat gempabumi besar di kota Bengkulu. Berdasarkan peta sebaran nilai GSS, potensi deformasi paling tinggi terindentfikasi di bagian barat Kota Bengkulu yaitu wilayah dengan formasi geologi undak alluvium, alluvium dan sebagian formasi andesit, sedangkan formasi geologi lainya relatif lebih aman. Dinamika lapisan tanah permukaan kota Bengkulu umumnya bersifat elastis dan berpotensi mengalami rekahan dansettlement. Struktur geologi yang relatif lunak dan lapisan sedimen yang tipis memungkinkan terjadinya penguatan efek gelombang dan resonansi tinggi ketika gempabumi terjadi. Kata-kata kunci: Kondisi geologi lokal kota Bengkulu, Ground Shear Strain, deformasi. Abstract Local geology condition has become one of the most parameters which affected damage level of earthquake. Geology condition of Bengkulu city has been described by Ground Shear Strain (GSS) value that obtained by sixty seven microtremor data analysis result what instaled on each geology formation types. According analysis result, GSS value of Bengkulu city was heterogeneous relatively in spite of similar geology formation type. GSS in Bengkulu City has obtained and be divided to three category, including 10-6 (25.37%), 10-5 (40.30%) and 10-4 (32.84%). These variations indicate differences in the thickness of sediment layer and rock solidity level of sites. The Areas of high value of GSS have compatibility with actual field conditions (Based on ground survey, it was swamp area and morphology ramps relatively) and have compatibility with damaged earthquake effect history of Bengkulu city. According map of GSS value distributions, high potency of deformation was indicated in western of Bengkulu city area, it was alluvium terraces formation, alluvium formation and partial of andesite formation, while others were relatively safer. The dynamic characteristic of surface soil layer on Bengkulu city was elastic and potential crack and settlement.The soft geology structure relatively and thin of sediments layer would affected amplification of wave and high resonance when earthquake occur.

One of the biggest challenges in the current ticketing facility is "QUEUE" in buying our suburban railway tickets. In this fast growing world of technology we still stand in the queue or buy with ATVM smart cards and coupons for our suburban tickets, which is more frustrating at times to stand in the queue or if we forget our cards. This paper Android Suburban Railway (ASR) ticketing is mainly to buy the suburban tickets which is the most challenging when compared to booking the long journey tickets through 'M­ ticket' which fails with suburban(local travel) tickets. Our ASR ticket can be bought with just a smart phone application, where you can carry your suburban railway tickets in your smart phone as a QR (Quick Response) code. It uses the smart phones "GPS" facility to validate your ticket. User's ticket information is stored in a CLOUD database for security purpose which is missing in the present suburban system. Also the ticket checker is provided with a checker application to search for the user's ticket with the ticket number in the cloud database for checking purposes.

Natural Language Processing (NLP), is a branch of artificial intelligence. NLP includes speech synthesis, Speech recognition, and Machine translation. Natural Language Processing has a wide range of applications in the Indian context. Most of the rural Indian community is unable to make use of the information technology revolution due to the dominance of English. Developments in the Natural processing Technology will offer universal access to information and services for more number of people in their mother language. The benefits that are expected to accrue as a result of widespread use of local language computing in India are discussed in this paper. India being a multilingual country with 22 official languages and about 1650 dialects, the research in Natural Language Processing faces great challenges in India. This paper also makes a detailed analysis of the difficulties and challenges facing the research community in the field of Natural Language Processing in India, and suggests practical solutions

Natural Language Processing (NLP), which is a branch of artificial intelligence, includes speech synthesis, Speech recognition, and Machine translation. Natural Language Processing has a wide range of applications in the Indian context. Most of the rural Indian community is unable to make use of the information technology revolution due to the dominance of English. Developments in the Natural processing Technology will offer universal access to information and services for more number of people in their mother language. The benefits that are expected to accrue as a result of widespread use of local language computing in India are discussed in this paper. India being a multilingual country with 22 official languages and about 1650 dialects, the research in Natural Language Processing faces great challenges in India. This paper makes a survey of the progress made in this field and explains the difficulties and challenges faced by the research community in the field of Natural Language Processing in India, and also suggests some practical solutions.

Speech recognition applications are becoming common and useful in this day and age as many of the modern devices are designed and produced user-friendly for the convenience of general public. Speaking/communicating directly with the machine to achieve desired objectives make usage of modern devices easier and convenient. Although many interactive software applications are available, the uses of these applications are limited due to language barriers. Hence development of speech recognition systems in local languages will help anyone to make use of this technological advancement. In India, speech recognition systems have been developed for many indigenous languages. In this paper , we present the survey of major research works in the development of automatic speech recognition in Indian language.

Malayalam is one of the 22 scheduled languages in India with more than 130 million speakers. This paper presents a report on the development of a speaker independent, continuous transcription system for Malayalam. The system employs Hidden Markov Model (HMM) for acoustic modeling and Mel Frequency Cepstral Coefficient (MFCC) for feature extraction. It is trained with 21 male and female speakers in the age group ranging from 20 to 40 years. The system obtained a word recognition accuracy of 87.4% and a sentence recognition accuracy of 84%, when tested with a set of continuous speech data.

Stød is a prosodic feature in Danish spoken language that is able to distinguish lexemes. This distinction can also identify word class and has the potential to improve the performance of automatic speech recognisers for Danish spoken language. Stød manifestation exhibits a large amount of variability and may be perceptual in nature, because stød in some cases can be audibly perceived yet not be visible in a spectrogram. The variability is the primary reason there is currently no agreed upon acoustic or phonetic definition of stød. The working definition of stød is “. . . a kind of creaky voice” (Grønnum, 2005) and “stød is not just creak” (Hansen, 2015). In the present work, we investigate whether stød can be exploited in automatic speech recognition. To exploit stød without an acoustic or phonetic definition, we need to use a (almost) zero-knowledge datadriven approach which is based on a number of assumptions that we investigate prior to conducting ASR experimentation. We assume that stød can be detected in audio input, using acoustic features. To detect stød, we need to identify features that signal stød, which requires annotated data. To select the right features, the stød annotation must be reliable and accurate. We therefore conduct a reliability study of stød annotation with inter-annotator agreement measures, rank acoustic features for stød detection according to feature importance using a forest of randomised decision trees and experiment with stød detection as a binary and multi-class classification task. The experiments identify a set of features important or stød detection and confirms that we can detect stød in audio. Lastly, we model stød in automatic speech recognition and show that significant improvements in word error rate can be gained simply by annotating stød in the phonetic dictionary at the expense of decoding speed. Extending the acoustic feature vectors with pitch-related features and other features of voice quality also give significant performance improvement on both read-aloud speech and spontaneous speech. Decoding speed increases when we extend the acoustic feature vectors and actually improve decoding speed over the baseline where stød is not modelled.

We propose data and knowledge-driven approaches for multilingual training of the automated speech recognition (ASR) system for a target language by pooling speech data from multiple source languages. Exploiting the acoustic similarities between Indian languages, we implement two approaches. In phone/senone mapping, deep neural network (DNN) learns non-linear functions to map senones or phones from one language to the others, and the transcriptions of the source languages are modified such that they can be used along with the target language data to train and fine-tune the target language ASR system. In the other approach, we model the acoustic information for all the languages simultaneously by training a multitask DNN (MTDNN) to predict the senones of each language in different output layers. The cross-entropy loss function and the weight update procedure are modified such that only the shared layers and the output layer responsible for predicting the senone classes of a language are updated during training, if the feature vector belongs to that particular language. In the low-resourced setting (LRS), 40 hours of transcribed speech data each for Tamil, Telugu and Gujarati languages are used for training. The DNN based senone mapping technique gives relative improvements in word error rates (WERs) of 9.66%, 7.2% and 15.21% over the baseline system for Tamil, Gujarati and Telugu languages, respectively. In medium-resourced setting (MRS), 160, 275 and 135 hours of data for Tamil, Kannada and Hindi languages are used, where, the same technique gives better relative improvements of 13.94%, 10.28% and 27.24% for Tamil, Kannada and Hindi, respectively. The MTDNN with senone mapping based training in LRS, gives higher relative WER improvements of 15.0%, 17.54% and 16.06%, respectively for Tamil, Gujarati and Telugu, whereas in MRS, we see improvements of 21.24% 21.05% and 30.17% for Tamil, Kannada and Hindi languages, respectively. CCS Concepts: • Computing methodologies → Speech recognition.

Aim of the study is to determine the better performance measure during the financial crisis. In order to do so, we selected the pandemic period during the Jan to Jul 2020 and analyzed the different performance ratios for our sample of 1416 Indian equity funds. Then we ranked them based on different performance indicators to show the gross difference while opting for different performance ratios. Finally we observe that adjusted Sharpe ratio is the best performance measure that can be used in the volatile markets.

Learning good representations is of crucial importance in deep learning. Mutual Information (MI) or similar measures of statistical dependence are promising tools for learning these representations in an unsupervised way. Even though the mutual information between two random variables is hard to measure directly in high dimensional spaces, some recent studies have shown that an implicit optimization of MI can be achieved with an encoder-discriminator architecture similar to that of Genera-tive Adversarial Networks (GANs). In this work, we learn representations that capture speaker identities by maximizing the mutual information between the encoded representations of chunks of speech randomly sampled from the same sentence. The proposed encoder relies on the SincNet architecture and transforms raw speech waveform into a compact feature vector. The discriminator is fed by either positive samples (of the joint distribution of encoded chunks) or negative samples (from the product of the marginals) and is trained to separate them. We report experiments showing that this approach effectively learns useful speaker representations, leading to promising results on speaker identification and verification tasks. Our experiments consider both unsupervised and semi-supervised settings and compare the performance achieved with different objective functions.

Research on automatic speech recognition has been approach progressively since 1930 and the major advances are since 1980 with the introduction of the statistical modeling of speech with the key technology Hidden Markov Model (HMM) and the stochastic language model . However, the existing reported research works on Bangla speech recognition didn't yet incorporate the HMM technique and language model. This paper presents two different type of Bangla speech recognition from the implementation, performance and application perspective. We used HMM technique for pattern classification and also incorporate stochastic language model with the system. At the signal preprocessing level we perform adaptive noise elimination and end point detection. Spectral feature vectors such as Mel Frequency Cepstral Coefficients (MFCC) with the addition of first and second order coefficients are extracted from each speech wave signal. HMM is used for pattern classification. The system is implemented using the Cambridge Hidden Markov Modeling Toolkit (HTK) (S. Young, 2001(S. Young, -2005.

Development of Malayalam speech recognition system is in its infancy stage; although many works have been done in other Indian languages. In this paper we present the first work on speaker independent Malayalam continuous speech recognizer based on PLP (Perceptual Linear Predictive) Cepstral Coefficient. The performance of the developed system has been evaluated with different number of states of HMM (Hidden Markov Model), Different number of Gaussian mixtures, and tied states. We have also evaluated the performance of the system with bigram and trigram language models. Moreover this paper compares the recognition accuracy of context independent and context dependent tied state models. The system employs Hidden Markov Model (HMM) for pattern recognition. The system is trained with 21 male and female speakers in the age group ranging from 19 to 41 years. The system obtained a word recognition accuracy of 89% and a sentence recognition accuracy of 83%, when tested with continuous speech data from unseen speakers

In many workloads, most write operations performed on a file system modify only a small number of blocks. The logstructured file system was designed for such a workload, additionally with the aim of fast crash recovery and system snapshots. Surprisingly, although implemented for Berkeley Sprite and BSD systems, there was no complete implementation for the current Linux kernel. In this paper, we present a complete implementation of the log-structured file system for the Linux kernel, which includes a user-space garbage collector and additional tools. We evaluate the measurements obtained in several test cases and compare the results with widely-used ext3.

In this work the random forest language modeling approach is applied with the aim of improving the performance of the LIMSI, highly competitive, Mandarin Chinese speech-to-text system. The experimental setup is that of the GALE Phase 4 evaluation. This setup is characterized by a large amount of available language model training data (over 3.2 billion segmented words). A conventional unpruned 4-gram language model with a vocabulary of 56K words serves as a baseline that is challenging to improve upon. However moderate perplexity and CER improvements over this model were obtained with a random forest language model. Different random forest training strategies were explored so as to attain the maximal gain in performance and Forest of Random Forest language modeling scheme is introduced.

This paper presents an optimum speaker independent isolated digit recognizer for Malayalam language. The system employs Perceptual Linear Predictive (PLP) cepstral coefficient for speech parameterization. The powerful and well accepted pattern recognition technique Hidden Markov Model is used for acoustic modeling. The training data base has the utterance of 21 speakers from the age group ranging from 20 to 40 years and the sound is recorded in the normal office environment where each speaker uttered digits zero to nine independently. The system obtained an accuracy of 99.5% with the unseen data.

Granite is one of the most commonly employed materials in the production of aggregates for concrete, and represents 40% of the total volume of aggregates produced in Portugal. This type of rock is traditionally considered as slowly/late reactive or even non-reactive to alkalis. However, a number of cases of damaged concrete structures in Portugal, due to alkali–silica reaction, have been related to granitic aggregates. A research program has been developed in order to define the best test method for evaluating the potential alkali-reactivity of granitic rocks. The present study involved thirteen granites collected from different quarries. The tests carried out included: petrographic examination of the aggregate, as well as mortar and concrete expansion tests. It was concluded that the content of microcrystalline quartz correlates better with the results of concrete prism expansion tests than with the mortar-bar expansion test.

Resumen-En este artículo se presenta una propuesta metodológica para construir un sistema de reconocimiento automático de voz basado en Modelos Ocultos de Markov utilizando Sphinx-4 como herramienta de implementación y técnicas de diseño de experimentos y análisis de datos para la selección del mejor sistema de reconocimiento. Los indicadores de desempeño del sistema utilizados fueron la tasa de error de reconocimiento de palabras (WER por sus siglas en inglés), el tiempo de procesamiento y el uso de memoria durante el reconocimiento.

Speech is the natural voice and primary means of speech. Communication. Communication. Speech is easy, hand-free, quick and it doesn't require anything. Any technical know-how. Communicating with your computer using Speech is a simple and comfortable way for human beings. Speech to the This was made possible by the recognition system. Language and acoustics There is a model for this method, but mostly in English language. There are so many groups in India that can't Comprehend English or speak it. So the device of speech recognition in To these people, the English language is of little use. Here we have implemented Isolated method of recognition of Hindi words, which is part of System for Automated Speech Recognition (ASR)]. The primary purpose of The ASR system recognizes a voice through a device or microphone and To perform the necessary process, it is translated into text. In this article, As a feature extraction technique, we used Mel frequency cepstral coefficients (MFCC), Gaussian Mixture Model(GMM) Vector quantization (VQ) for Recognition of words separated from Hindi. We have practical research for The Hindi word speech dataset of various males and men was prepared.

The effect of palatalization on the articulatory and acoustic realization of Bulgarian plosives is well described. It has been shown in cross-splicing experiments that human listeners are more strongly influenced by the transition than by the burst. In two sets of experiments, palatalization in Bulgarian plosives is investigated in an automatic speech recognition setting. The confusions between phonemes in a plosive identification experiment show that palatalized consonants can be identified well on the basis of closure and burst alone. The addition of vowel transitions leads to a greater improvement in the identification rate for non-palatalized than for palatalized consonants. Follow-up experiments using quasi-cross-splicing indicate that the information carried by the transitions is more important than the information in the closure plus burst for both non-palatalized and palatalized plosives.

Performance of any continuous speech recognition system is dependent on the accuracy of its acoustic model. Hence, preparation of a robust and accurate acoustic model lead to satisfactory recognition performance for a speech recognizer. In acoustic modeling of phonetic unit, context information is of prime importance as the phonemes are found to vary according to the place of occurrence in a word. In this paper we compare and evaluate the effect of context dependent tied (CD tied) models, context dependent (CD) and context independent (CI) models in the perspective of continuous speech recognition of Malayalam language. The database for the speech recognition system has utterance from 21 speakers including 11 female and 10 males. Our evaluation results show that CD tied models outperforms CI models over 21%.

Building an automatic speech recognition (ASR) system from scratch requires a large amount of annotated speech data, which is difficult to collect in many languages. However, there are cases where the low-resource language shares a common acoustic space with a high-resource language having enough annotated data to build an ASR. In such cases, we show that the domain-independent acoustic models learned from the high-resource language through unsupervised domain adaptation (UDA) schemes can enhance the performance of the ASR in the low-resource language. We use the specific example of Hindi in the source domain and Sanskrit in the target domain. We explore two architectures: i) domain adversarial training using gradient reversal layer (GRL) and ii) domain separation networks (DSN). The GRL and DSN architectures give absolute improvements of 6.71% and 7.32%, respectively, in word error rate over the baseline deep neural network model when trained on just 5.5 hours of data in the target domain. We also show that choosing a proper language (Telugu) in the source domain can bring further improvement. The results suggest that UDA schemes can be helpful in the development of ASR systems for low-resource languages, mitigating the hassle of collecting large amounts of annotated speech data.

Speech recognition applications are becoming common and useful in this day and age as many of the modern devices are designed and produced user-friendly for the convenience of general public. Speaking/communicating directly with the machine to achieve desired objectives make usage of modern devices easier and convenient. Although many interactive software applications are available, the uses of these applications are limited due to language barriers. Hence development of speech recognition systems in local languages will help anyone to make use of this technological advancement. In India, speech recognition systems have been developed for many indigenous languages. In this paper , we present the survey of major research works in the development of automatic speech recognition in Indian language. Keywords— Speech Recognition, Indian languages

Many of the world’s languages are unstudied and on the verge of disappearance. A few linguists strive to address this problem, but compared to Indo-European languages, there is much more work to do, with only a fraction of the manpower and under immense time pressure. Over the last two decades the field of ‘documentary linguistics’ has attempted to address this situation, but with successes that can be said to be partial at best.

After considering some of the obstacles to language documentation, this talk takes a fresh look at the technical side of linguistic fieldwork. It does this from the perspective of recent technological innovations, and examines the workflows and approaches to data management in particular. I conduct two case studies, the methods through which, may hold great promise in helping communities that speak understudied languages, namely autocompletion software for mobile phones and the automatic phonetic transcription of spoken languages using neural networks.

Packet loss due to misrouted or delayed packets in voice over IP leads to huge voice quality degradation. Packet loss concealment algorithms try to enhance the quality of the speech. This paper presents a new packet loss concealment algorithm which relies on one hidden Markov model. For this purpose, we introduce a continuous observation vector well-suited for silence, voiced and unvoiced sounds. We show that having a global HMM is relevant for this application. The proposed system is evaluated using standard PESQ score in a realworld application.

Frequency of Hebrew phonemes and phoneme clusters in a data-driven approach*
This article describes applied research primarily intended to form the basis for the establishment of a linguistic infrastructure for the training of acoustic models and a language model for an Automatic Speech Recognition (ASR) in Hebrew. The stage described in this article refers to the process of extracting the sentences that are to be read and recorded by Hebrew speakers. In this process, 50,000 diverse and phonetically compressed sentences were collected in a data driven approach, from a textual corpus containing more than 130 million word-tokens (about 1 million word-types). All the sentences were transcribed according to a pre-prepared lexicon, meaning that the orthographic writing of the sentence was automatically converted to the sequence of phonemes that represented it and that was decided upon in advance. Findings include the distribution of the 31 phonemes in Hebrew, the frequency of diphones, and the frequency of triphones, which were obtained at the sentence level, i.e., between words and within words. The findings show that the most common phoneme is / a / and the most rare is / ʒ /. It was also found that the phonetically unmarked consonant [p] is only in the 27th position at the level of prevalence, whereas more marked consonants such as [x] and [k] were found in higher positions (22 and 12, respectively). In addition, findings demonstrate how diphones and triphones over the word level produce new phonological environments for all the phonemes.
* In Hebrew

This study presents a psycholinguistically motivated evaluation method for phoneme classifiers by using non-categorical perceptual data elicited in a Japanese sibilant matching 2AFC task. Probability values of a perceptual [s]-[S] boundary, obtained from 42 speakers over a 7-step synthetic [s]-[sh] continuum, were compared to probability estimates of Gaussian mixture models (GMMs) of Japanese [s] and [S]. The GMMs, trained on the Corpus of Contemporary Spontaneous Japanese, differed in feature vectors (MFCC, PLP, acoustic features), covariance matrix types (full, tied, diagonal, spherical), and numbers of mixtures (1-20). Using ten-fold cross validation, it was found that GMMs trained on MFCC features had the best sibilant classification accuracies (87.4-90.4%), but their correlations with human perceptual data were non-conclusive (0.35-0.98). Acoustic feature-based GMMs with tied covariance matrices had near human-like synthetic stimuli perception (0.957-0.996), but their classification performance was poor (71.3-80.4%). Models trained on perceptual linear prediction (PLP) features were on par with the acoustic feature-based models in terms correlation to the perceptual experiment (0.884-0.995), while losing slightly on classification performance (86.1-88.9%) compared to MFCC models. Across the board correlation tests and mixture-effect models confirmed that GMMs with better sibilant classifying performance produced more human-like probability estimations on the synthetic sibilant continuum.

Present Mel Frequency Cepstral Coefficient (MFCC) based Bangla Automatic Speech Recognition (ASR) systems are mostly implemented with delta and acceleration coefficients. With delta and acceleration coefficients of MFCC and the log energy, a vector set of 39 dimensions is obtained per 10ms. In this paper, our objective is to observe the effect of third differential coefficients on the performance of Bangla ASR, which is not explored in this field yet. In doing so, we have appended 13 third differential coefficients along with previous 39 coefficients to make a vector set of 52 coefficients per 10ms frame. We have observed the performance of Bangla ASR system in the presence and absence of third differential coefficients using Hidden Markov Model (HMM) based tied-state triphone model. To make the speech corpus, 100 sentences have been uttered by a different number of speakers at different phases including both male and female of similar ages in between 22–24. Hidden-Markov-Model Toolkit (HTK) has been used here for the comparative analysis. We have considered the Sentence Correction Rate (SCR) as the performance indicator. From the experiments, it has been observed that the MFCC based system of 39 (MFCC39) and 52 (MFCC52) dimensions have average SCR of 98.89% and 98.94% respectively. Therefore, our finding is that slight improvement is possible with the inclusion of third differential coefficients when the sampling data rate is as high as 44.1 KHz.

Among the several options of managed aquifer recharge (MAR) techniques, the aquifer storage and recovery (ASR) is a well-known sub-surface technique to replenish depleted aquifers, which is contingent upon the selection of appropriate sites. This paper explores the potential of ASR for groundwater recharge in the hydrological, hydrogeological, social, and economic context of South Bihar in India. Based on the water samples from more than 137 wells and socio-economic surveys, ASR installations were piloted through seven selected entrepreneurial farmers in two villages of South Bihar. The feasibility of ASR in both hard rock and deep alluvial aquifers was demonstrated for the prominent aquifer types in the marginal alluvial plains of South Bihar and elsewhere. It was postulated through this pilot study that a successful spread of ASR in South Bihar can augment usable water resources for agriculture during the winter cropping season. More importantly, ASR can adapt to local circumstanc...

Language model plays an important role in
automatic speech recognition (ASR) systems. Performance of
this model depends on its adaptation to the linguistic features.
Accordingly, adaptation methods endeavour to apply syntactic
and semantic characteristics of the language for language
modeling. The previous adaptation methods such as family of
Dirichlet class language model (DCLM) extract class of history
words. These methods due to lake of syntactic information are
not suitable for high morphology languages such as Farsi. This
work proposes an idea for using syntactic information such as
part-of-speech (POS) in DCLM for combining with an n-gram
language model. In our proposed approach, word clustering is
based on POS of previous words and history words. The
performance of language models are evaluated on BijanKhan
corpus using a hidden Markov model based ASR system. Our
experiments show that using POS information along with
history words and class of history words improves language
model, and decreases the perplexity on our corpus. Exploiting
POS information along with DCLM, the word error rate of the
ASR system decreases by 1% in comparison to DCLM.

Research on automatic speech recognition has been approach progressively since 1930 and the major advances are since 1980 with the introduction of the statistical modeling of speech with the key technology Hidden Markov Model (HMM) and the stochastic language model . However, the existing reported research works on Bangla speech recognition didn't yet incorporate the HMM technique and language model. This paper presents two different type of Bangla speech recognition from the implementation, performance and application perspective. We used HMM technique for pattern classification and also incorporate stochastic language model with the system. At the signal preprocessing level we perform adaptive noise elimination and end point detection. Spectral feature vectors such as Mel Frequency Cepstral Coefficients (MFCC) with the addition of first and second order coefficients are extracted from each speech wave signal. HMM is used for pattern classification. The system is implemented using the Cambridge Hidden Markov Modeling Toolkit (HTK) (S. Young, 2001(S. Young, -2005.

This paper reports our recent development of a highly reliable call analysis technique that makes novel use of automatic speech recognition (ASR), speech utterance classification and non-speech features. The main ideas include the use the N-Gram filler model to improve the ASR accuracy on important words in a message, and the integration of recognized utterance with non-speech features such as utterance length, and the use of utterance classification technique to interpret the message and extract additional information. Experimental evaluation shows that the use of the utterance length, recognized text, and the classifier's confidence measure reduces the classification error rate to 2.5% of the test sets.