New discriminative training algorithms based on the generalized probabilistic descent method

The goal of this communication is to present a weighted likelihood discriminant for minimum error shape classification. Different from traditional maximum likelihood (ML) methods in which classification is carried out based on probabilities from independent individual class models as is the case for general hidden Markov model (HMM) methods, our proposed method utilizes information from all classes to minimize classification error. Proposed approach uses a hidden Markov model as a curvature feature based 2D shape descriptor. In this contribution we present a generalized probabilistic descent (GPD) method to weight the curvature likelihoods to achieve a discriminant function with minimum classification error. In contrast with other approaches, a weighted likelihood discriminant function is introduced. We believe that our sound theory based implementation reduces classification error by combining hidden Markov model with generalized probabilistic descent theory. We show comparative results obtained with our approach and classic maximum-likelihood calculation for fighter planes in terms of classification accuracies

We propose a method for sound monitoring, which enables one to selectively detect unexpected irregular sounds and ignore the other sounds, i.e., regular sounds. The proposed method is based on the generalized probabilistic descent (GPD) method, which was originally developed as a general concept for the discriminative design of pattern recognizers, and is referred to as minimum detection error (MDE) training. The formulation and implementation of MDE training are described in detail, and its utility is demonstrated in a task of detecting irregular events; more specifically, sounds due to the mis-operation of a tool in a noisy environment.

In this work we compare two parameter optimization techniques for discriminative training using the MMI criterion: the extended Baum-Welch (EBW) algorithm and the generalized probabilistic descent (GPD) method. Using Gaussian emission densities we found special expressions for the step sizes in GPD, leading to reestimation formula very similar to those derived for the EBW algorithm. Results were produced for both the TI digitstring and the SieTill corpus for continuously spoken American English and German digitstrings. The results for both techniques do not show significant differences. This experimental results support the strong link between EBW and GPD as expected from the analytic comparison.

Our earlier frame margin probability discriminative training (FMPDT) algorithm has shown its advantage over maximum likelihood estimation (MLE) on multi-condition speech recognition. This paper reports our recent improvement on FMPDT and compares it with minimum classification error (MCE) and minimum phone error (MPE). We rewrite the re-estimation formulae of MCE, MPE in a unified linear interpolation way which facilitates this comparison. We compare the recognition results for MCE and MPE, FMPDT using the Aurora2 standard database and HTK3.4, and find the proposed method can give more balanced improvements under different conditions than other discriminative training algorithms.

This paper introduces an improved HMM (hidden Markov model) for low altitude acoustic target recognition. To overcome the limitation of the classical CDHMM (continuous density hidden Markov model) training algorithm and the generalization ability deficiency of existing discriminative learning methods, a new discriminative training method for estimating the CDHMM in acoustic target recognition is proposed based on the principle of maximizing the minimum relative separation margin. According to the definition of the relative margin, the new training criterion can be equation as a standard constrained minimax optimization problem. Then, the optimization problem can be solved by a GPD (generalized probabilistic descent) algorithm. The experimental results show that the performance of the algorithm is significantly improved compared with the former training method, which can effectively improve the recognition ability of the acoustic target recognition system.

English translation, also a kind of language conversion, refers to the activities of expressing each other in English to express another language or another language to express English. This paper aimed to evaluate and research methods of English translation based on the improved Gaussian mixture model (GMM). The paper proposed a simple evaluation and analysis of English translation using a discriminative training algorithm. In the experimental part, on the one hand, taking students, teachers, and translators as case study objects, through a questionnaire survey, it can be known that the number of students who scored more than 80 points in the English translation test is increasing every year, from 11.30% in the first year to 21.13% in the fourth year, an increase of nearly 10%. On the other hand, it can be seen from the speech translation experiment that the correct translation rate was 46.18% by using intelligent technology to distinguish translation. The experimental results showed that the discriminative training algorithm under the improved GMM algorithm is effective in the research of English translation evaluation.

In this paper, based on the study of two speech recognition algorithms, two designs of speech recognition system are given to realize this isolated speech recognition mobile robot control system based on ARM9 processor. The speech recognition process includes pretreatment of speech signal, characteristic extrication, pattern matching and post-processing. Mel-Frequency cepstrum coefficients (MFCC) and linear prediction cepstrum coefficients (LPCC) are the two most common parameters. Through analysis and comparison the parameters, MFCC shows more noise immunity than LPCC, so MFCC is selected as the characteristic parameters. Both dynamic time warping (DTW) and hidden markov model (HMM) are commonly used algorithm. For the different characteristics of DTW and HMM recognition algorithm, two different programs were designed for mobile robot control system. The effect and speed of the two speech recognition system were analyzed and compared.

Dynamic time warping in phoneme modeling for fast pronunciation error detection

It's difficult to build precise mathematical models for complex engineering systems because of the complexity of the structure and dynamics characteristics. Intelligent fault diagnosis introduces artificial intelligence and works in a different way without building the analytical mathematical model of a diagnostic object, so it's a practical approach to solve diagnostic problems of complex systems. This paper presents an intelligent fault diagnosis method, an integrated fault-pattern classifier based on Hidden Markov Model (HMM). This classifier consists of dynamic time warping (DTW) algorithm, self-organizing feature mapping (SOFM) network and Hidden Markov Model. First, after dynamic observation vector in measuring space is processed by DTW, the error vector including the fault feature of being tested system is obtained. Then a SOFM network is used as a feature extractor and vector quantization processor. Finally, fault diagnosis is realized by fault patterns classifying with the Hidden Markov Model classifier. The importing of dynamic time warping solves the problem of feature extracting from dynamic process vectors of complex system such as aeroengine, and makes it come true to diagnose complex system by utilizing dynamic process information. Simulating experiments show that the diagnosis model is easy to extend, and the fault pattern classifier is efficient and is convenient to the detecting and diagnosing of new faults.

In this dissertation, a novel approach of integrating template matching with statistical modeling is proposed to improve continuous speech recognition. Hidden Markov Modeling (HMMs) has been the dominant approach in statistical speech recognition since it provides a principled way of jointly modeling speech spectral variations and time dynamics. However, HMMs have the shortcoming of assuming the observations being independent within each state, which makes it ineffective in modeling the details of speech temporal evolutions that are important for characterizing nonstationary speech sounds. Template-based methods make comparisons between a test pattern and the templates derived from training data, and therefore they are able to capture speech dynamics and time correlation of speech frames better than HMM based methods. However, template matching requires large memory space and computational time since feature vectors of training data need to be stored in computer memory for access at the recognition stage, which is difficult in large vocabulary continuous speech recognition (LVCSR). Our proposed approach takes advantages of both statistical modeling and template matching, which overcomes the weakness of conventional template-based method and is feasible for LVCSR. We use multiple Gaussian Mixture Model (GMM) indices to represent each frame of speech templates, and define the template unit to be context-dependent phone segments (triphone context). We also use phonetic decision trees borrowed from those commonly used in HMMs to tie triphone templates and predict triphones unseen in training data. Two local distances, log likelihood ratio (LLR) and Kullback-Leibler (KL) divergence, are proposed for dynamic time warping (DTW) based template matching. In order to reduce computational complexity and storage space, we propose methods of minimum distance template selection (MDTS) and maximum log-likelihood template selection (MLTS), and investigate a template compression method on top of template selection to further improve recognition performance. The template based methods were used to rescore lattices generated by baseline HMMs on the tasks of TIMIT continuous phone recognition and teleheath LVCSR and experimental results demonstrated that the proposed approach of integrating template matching with statistical modeling significantly improved recognition performances over the HMM baselines. The template selection methods also provided significant recognition accuracy improvements over the HMM baseline while largely reducing the computation and storage complexities. When all templates or MDTS were used, using the LLR local distance obtained better recognition performance than the KL divergence local distance. For MLTS and template compression, KL divergence local distance provided better performance than the LLR local distance, and the template compression method made further improvements over KL based MLTS. Since the templates were constructed based on the GMM indices extracted from HMM baselines, we also validated the effectiveness of the proposed template methods based on enhanced HMM baselines. Experimental results showed that LLR based all template method was able to consistently improve TIMIT phone recognition accuracies based on four enhanced HMM baselines. Prosodic features such as duration, energy, and pitch can reflect longer span information of speech than conventional single frame vectors but they have commonly been ignored by HMMs. Template based methods provide possibilities to conveniently integrate prosodic features into speech recognition, which has not been well studied in the past. In this dissertation, we investigate combining template based methods with the speech prosodic features of duration, energy and pitch to further improve speech recognition accuracy. The scores of prosodic information were computed by a GMM based method and a non-parametric method, and the prosodic scores were combined with the acoustic scores in triphone template matching. Experimental results obtained on the telehealth task showed that prosodic information had positive effects on vowel sound recognition.

Keyword spotting on embedded platforms must balance accuracy and strict resource limits while remaining independent of network connectivity. The aim of the study was to develop and experimentally validate a classical, frugal recognition method that increases feature informativeness without increasing model complexity and is suitable for autonomous use on edge devices that rely only on a central processing unit. A weighted acoustic f ingerprinting mechanism was proposed. Mel-frequency cepstral coefficients, together with their derivatives, were reweighted, aggregated, and serialised into compact discrete “fingerprints”, which were then classified using the Levenshtein edit distance. Experiments were carried out on a Ukrainian-language command corpus from six native speakers (three male, three female), recorded with both headsets and far-field microphones; lexicons of 10, 100, and 200 words were evaluated under speaker-independent splits of 70%/15%/15%. The methodology comprised f ixed parametrisation of mel-frequency cepstral coefficients, construction of a static weighting vector, voice activity detection with spectral subtraction, uniform quantisation and serialisation, and deterministic edit-distance classification; for comparison, equal-weight baselines, hidden Markov models with Gaussian mixture emissions, Dynamic Time Warping, a lightweight convolutional neural network, and a reference depthwise-separable convolutional neural network were considered. The proposed method achieved macro-averaged harmonic means of precision and recall of 0.96/0.92/0.89 for 10/100/200-word lexicons in clean audio, and 0.78 at a signal-to-noise ratio of 5 decibels (100-word lexicon). The implementation required approximately 250 kilobytes of memory and operated with a real-time factor of 0.005 on Raspberry Pi 4 with 4 gigabytes, i.e., faster than real time. Superiority over equal-weight baselines, hidden Markov models with Gaussian mixture emissions, and Dynamic Time Warping was demonstrated, with performance approaching that of a compact convolutional neural network. It is concluded that weighted acoustic fingerprinting provides a robust, efficient, and autonomous keyword-spotting solution for deployments that use only a central processing unit

The authors suggest a linear discriminant function to complete the distance score instead of a conventional average distance. Several discriminative algorithms are proposed to learn the discriminant function. These include one heuristic method, two methods based on the error propagation algorithm, and one method based on the generalized probabilistic descent (GPD) algorithm. The authors study these methods in a speaker-independent speech recognition task involving utterances of the highly confusable English E-set. The results show that the best performance is obtained by using the GPD method, which achieved a 78.1% accuracy, compared to 67.6% with the traditional average method.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Gesture recognition has become a very active research area with the advent of the Kinect sensor. The most common approaches for gesture recognition use temporal information and are based on methods such as Hidden Markov Models (HMM) and Dynamic Time Warping (DTW). In this paper, we present a novel non-temporal alternative for gesture recognition using the Microsoft Kinect device. The proposed approach, Recognition by Characteristic Window (RCW), identifies, using clustering techniques and a sliding window, distinctive portions of individual gestures which have low overlapping information with other gestures. Once a distinctive portion has been identified for each gesture, all these sub-sequences are used to recognize a new instance. The proposed method was compared against HMM and DTW on a benchmark gesture’s dataset showing very competitive performance.

In this paper, we present a novel occlusion detection scheme for hidden Markov modeled shapes. First, hidden Markov model (HMM) is built using multiple examples of the shape. A reference path for the shape is built from the HMM, which is nothing but optimal path followed by the most likely example. The reference path stores temporal information about the entire shape, while the HMM only retains relationship between temporal information. For the shape of interest, its optimal path through HMM is calculated and warped to match the reference path using dynamic time warping (DTW). Occluded part of the shape is detected by identifying imbalance among various components of the matching cost. Detection results obtained for two shape data sets are presented for varying degrees of occlusion.

The paper focuses on processing of long biological signals used during monitoring procedures like in the case of portable Holter device for arrythmia analysis (ECG), intracranial pressure monitoring (ICP) in intensive care unit or overnight electroencephalogram monitoring (EEG) for sleep apnea detection. Two methods taken from speech processing are proposed: dynamic time warping (DTW) and hidden Markov models (HMM). The unsupervised analysis of ECG and ICP beats is carried out using hierarchical clustering approach. In case of EEG, first the estimation of sleep stages is performed and next the different breathing events are detected by HMM by means of Viterbi inference. We show that for the first two problems DTW outperforms HMM while in the third case the HMM inference capability makes HMM suitable for sleep apnea diagnosis.