Sharif Digital Repository

By increasing the number of registered speakers in Speaker Identification (SI) systems, computation time for identifying an unknown speaker is significantly increased. This problem arises from the simple design of conventional methods. Due to this limitation, we cannot use conventional SI methods in real time applications. In this paper, we propose a two-step method to overcome this limitation. We use different identification methods for each step. In the first step we reduce the search space using Nearest Neighbor method. In the second step we identify the target speaker using the conventional GMM-based SI method. The experimental results show 3.4× speed-ups without any accuracy loss using... 

Reliability assessment of phonemes, syllabi, words, concepts or utterances has become the key feature of Automatic Speech Recognition (ASR) engines in order to make a decision to accept or reject a hypothesis. In this paper, we propose utterance-level confidence annotation based on combination of features extracted from multiple knowledge sources in Persian language. The experiment was conducted first to examine the performance of individual features, then to combine them using statistical data analysis and density estimation methods to assign a confidence score to utterances. Using the data collected from a Persian spoken dialogue system, we show that combining features from independent... 

Automatic speech recognition performance degrades significantly when speech is affected by environmental noise. Nowadays, the major challenge is to achieve good robustness in adverse noisy conditions so that automatic speech recognizers can be used in real situations. Spectral subtraction (SS) is a well-known and effective approach; it was originally designed for improving the quality of speech signal judged by human listeners. SS techniques usually improve the quality and intelligibility of speech signal while speech recognition systems need compensation techniques to reduce mismatch between noisy speech features and clean trained acoustic model. Nevertheless, correlation can be expected... 

This paper has presented a novel discriminative parameters calibration approach based on the Model Distance Maximizing (MDM) to improve the performance of our previous proposed robustness method named spectral subtraction (SS) in likelihoodmaximizing framework. In the previous work, for adjusting the spectral over-subtraction factor of SS, conventional ML approach is used that only utilizes the true model without considering other confused models. This makes it very probably to reach a suboptimal solution. While in MDM, by maximizing the dissimilarities among models, the performance of our speech recognizer-based spectral subtraction method could be further improved. Experimental results... 

This paper has presented a novel discriminative parameter calibration approach based on the model distance maximizing (MDM) framework to improve the performance of our previously-proposed method based on spectral subtraction (SS) in a likelihood-maximizing framework. In the previous work, spectral over-subtraction factors were adjusted based on the conventional maximum-likelihood (ML) approach that utilized only the true model and did not consider other confused models, thus likely reached suboptimal solutions. While in the proposed MDM framework, improved speech recognition performance is obtained by maximizing the dissimilarities among models. Experimental results based on FARSDAT, TIMIT... 

In large population Speaker Identification (SI), computation time has become one of the most important issues in recent real time systems. Test computation time depends on the cost of likelihood computation between test features and registered speaker models. For real time application of speaker identification, system must identify an unknown speaker quickly. Hence the conventional SI methods cannot be used. In this paper, we propose a two-step method that utilizes two different identification methods. In the first step we use Nearest Neighbor method to decrease the search space. In the second step we use GMM-based SI methods to specify the target speaker. We achieved 3.5× speed-ups without... 

Cosine similarity and Probabilistic Linear Discriminant Analysis (PLDA) in i-vector space are two state-of-the-art scoring methods in speaker verification field. While PLDA usually gives better accuracy, Cosine Similarity Scoring (CSS) remains a widely used method due to simplicity and acceptable performance. In this domain, several channel compensation and score normalization methods have been proposed to improve the performance. We investigate non-speaker information in cosine similarity metric and propose a new approach to remove it from the decision making process. I-vectors hold a large amount of non-speaker information such as channel effects, language, and phonetic content. This type... 

This paper presents a solution in order to solve the problem of using HMM-based POS tagger in some languages where a word can be comprised of several tokens. Viterbi algorithm is modified in order to support segment of words within a model state. In the other word, the proposed system has a built-in tokenizer where indicates words boundaries as well as its corresponding tag sequence  

This paper focuses on enhancing MFCC features using a set of MLP NN in order to improve phoneme recognition accuracy under different noise types and SNRs. A NN can be used in different domains (between any pair of MFCC feature extraction blocks). It includes FFT, MEL, LOG, DCT and DELTA domains. Various domains have different complexities and achieve different degrees. A comparative study is done in this paper in order to find the best domain. Furthermore, a set of MLP NNs, instead of one NN, is used to enhance various noise types with different levels of SNRs. In this case, each NN is trained with a special noise type and SNR. The database used in the simulations is created by artificially... 

This study focuses on the mode classification of phones speaker modes using GMM 1. In this regard, speech data in both enabled and disabled speaker modes of cell phones and telephones were collected, processed and classified into two different categories. The different mixture numbers (1 to 4) of GMM and wave files sizes of 10, 20, 40 and 80 kb were tested in order to obtain an optimal condition for classification. The GMM method attained 87.99% correct classification rate on test data. This classification is important for speech enabled IVR 2 systems [1], dialog systems and many systems in speech processing in the sense that it could help to load an optimum model for increasing system... 

In this paper we present a novel approach for adjusting a multi band spectral subtraction filter coefficients based on speech recognition system results. Currently most speech enhancement techniques are designed according to various waveform level criteria such as maximizing SNR or minimizing signal error. However improvement in these criteria does not necessarily result in increasing speech recognition performance. Only if these methods generate sequence of features that maximize or increase the likelihood of the correct transcription relative to other incorrect competing hypotheses, speech recognition performance will increase. Here we use an utterance with a known transcription and... 

In this paper we have reviewed Nevisa Persian speech recognition engine. Nevisa is an HMM-based, large vocabulary speaker-independent continuous speech recognition system. Like most successful recognition systems, MFCC with some modification has been used as speech signal features. It also utilizes a VAD based on signal energy and zero-crossing rate. Maximum likelihood estimation criterion the core of which are the classical segmental k-means and Baum-Welsh algorithms is used for training the acoustic models. The system is based on phoneme modeling and utilizes synchronous beam search based on lexicon tree for decoding the acoustic utterances. Language modeling for Persian has been... 

We propose an efficient and effective nonlinear feature domain noise suppression algorithm, motivated by the minimum mean square error (MMSE) optimization criterion. Multi Layer Perceptron (MLP) neural network in the log spectral domain minimizes the difference between noisy and clean speech. By using this method as a pre-processing stage of a speech recognition system, the recognition rate in noisy environments is improved. We can extend the application of the system to different environments with different noises without re-training it. We need only to train the preprocessing stage with a small portion ofnoisy data which is created by artificially adding different types of noises from the... 

In this paper, we have proposed an efficient and effective nonlinear feature domain noise suppression algorithm, motivated by the minimum mean square error (MMSE) optimization criterion. A Multi Layer Perceptron (MLP) neural network in the log spectral domain has been employed to minimize the difference between noisy and clean speech. By using this method, as a pre-processing stage of a speech recognition system, the recognition rate in noisy environments has been improved. We extended the application ofthe system to different environments with different noises without retraining HMMmodel. We trained the feature extraction stage with a small portion of noisy data which was created by...