Use of lexicon density in evaluating word recognizers

We have developed the notion of lexicon density as a metric to measure the expected accuracy of handwritten word recognizers. Thus far, researchers have used the size of the lexicon as a gauge for the difficulty of the handwritten word recognition task. For example, the literature mentions recognizers with accuracies for lexicons of sizes 10, 100, 1000, and so forth, implying that the difficulty of the task increases (and hence recognition accuracy decreases) with increasing lexicon size across recognizers. Lexicon density is an alternate measure which is quite dependent on the recognizer. There are many applications, such as address interpretation, where such a recognizer-dependent measure can be useful. We have conducted experiments with two different types of recognizers. A segmentation-based and a grapheme-based recognizer have been selected to show how the measure of lexicon density can be developed in general for any recognizer. Experimental results show that the lexicon density measure described is more suitable than lexicon size or a simple string edit distance     

Using topic models for OCR correction

Despite several decades of research in document analysis, recognition of unconstrained handwritten documents is still considered a challenging task. Previous research in this area has shown that word recognizers perform adequately on constrained handwritten documents which typically use a restricted vocabulary (lexicon). But in the case of unconstrained handwritten documents, state-of-the-art word recognition accuracy is still below the acceptable limits. The objective of this research is to improve word recognition accuracy on unconstrained handwritten documents by applying a post-processing or OCR correction technique to the word recognition output. In this paper, we present two different methods for this purpose. First, we describe a lexicon reduction-based method by topic categorization of handwritten documents which is used to generate smaller topic-specific lexicons for improving the recognition accuracy. Second, we describe a method which uses topic-specific language models and a maximum-entropy based topic categorization model to refine the recognition output. We present the relative merits of each of these methods and report results on the publicly available IAM database.     

Lexicon based semantic detection of sentiments using expected likelihood estimate smoothed odds ratio

Sentiment analysis is an active research area in today’s era due to the abundance of opinionated data present on online social networks. Semantic detection is a sub-category of sentiment analysis which deals with the identification of sentiment orientation in any text. Many sentiment applications rely on lexicons to supply features to a model. Various machine learning algorithms and sentiment lexicons have been proposed in research in order to improve sentiment categorization. Supervised machine learning algorithms and domain specific sentiment lexicons generally perform better as compared to the unsupervised or semi-supervised domain independent lexicon based approaches. The core hindrance in the application of supervised algorithms or domain specific sentiment lexicons is the unavailability of sentiment labeled training datasets for every domain. On the other hand, the performance of algorithms based on general purpose sentiment lexicons needs improvement. This research is focused on building a general purpose sentiment lexicon in a semi-supervised manner. The proposed lexicon defines word semantics based on Expected Likelihood Estimate Smoothed Odds Ratio that are then incorporated with supervised machine learning based model selection approach. A comprehensive performance comparison verifies the superiority of our proposed approach.     

Phrase-based correction model for improving handwriting recognition accuracies

We propose a method for increasing word recognition accuracies by correcting the output of a handwriting recognition system. We treat the handwriting recognizer as a black box, such that there is no access to its internals. This enables us to keep our algorithm general and independent of any particular system. We use a novel method for correcting the output based on a “phrase-based” system in contrast to traditional source-channel models. We report the accuracies of two in-house handwritten word recognizers before and after the correction. We achieve highly encouraging results for a large synthetically generated dataset. We also report results for a commercially available OCR on real data.     

Prosody dependent speech recognition on radio news corpus of American English

Does prosody help word recognition? This paper proposes a novel probabilistic framework in which word and phoneme are dependent on prosody in a way that reduces word error rates (WER) relative to a prosody-independent recognizer with comparable parameter count. In the proposed prosody-dependent speech recognizer, word and phoneme models are conditioned on two important prosodic variables: the intonational phrase boundary and the pitch accent. An information-theoretic analysis is provided to show that prosody dependent acoustic and language modeling can increase the mutual information between the true word hypothesis and the acoustic observation by exciting the interaction between prosody dependent acoustic model and prosody dependent language model. Empirically, results indicate that the influence of these prosodic variables on allophonic models are mainly restricted to a small subset of distributions: the duration PDFs (modeled using an explicit duration hidden Markov model or EDHMM) and the acoustic-prosodic observation PDFs (normalized pitch frequency). Influence of prosody on cepstral features is limited to a subset of phonemes: for example, vowels may be influenced by both accent and phrase position, but phrase-initial and phrase-final consonants are independent of accent. Leveraging these results, effective prosody dependent allophonic models are built with minimal increase in parameter count. These prosody dependent speech recognizers are able to reduce word error rates by up to 11% relative to prosody independent recognizers with comparable parameter count, in experiments based on the prosodically-transcribed Boston Radio News corpus.     

Parallel Speech Recognition

Computer speech recognition has been very successful in limited domains and for isolated word recognition. However, widespread use of large-vocabulary continuous-speech recognizers is limited by the speed of current recognizers, which cannot reach acceptable error rates while running in real time. This paper shows how to harness shared memory multiprocessors, which are becoming increasingly common, to increase the speed significantly, and therefore the accuracy or vocabulary size, of a speech recognizer. To cover the necessary background, we begin with a tutorial on speech recognition. We then describe the parallelization of an existing high-quality speech recognizer, achieving a speedup of a factor of 3, 5, and 6 on 4-, 8-, and 12-processors respectively for the benchmark North American business news (NAB) recognition task.     

A probabilistic framework for segment-based speech recognition

Most current speech recognizers use an observation space based on a temporal sequence of measurements extracted from fixed-length “frames” (e.g., Mel-cepstra). Given a hypothetical word or sub-word sequence, the acoustic likelihood computation always involves all observation frames, though the mapping between individual frames and internal recognizer states will depend on the hypothesized segmentation. There is another type of recognizer whose observation space is better represented as a network, or graph, where each arc in the graph corresponds to a hypothesized variable-length segment that is represented by a fixed-dimensional “feature”. In such feature-based recognizers, each hypothesized segmentation will correspond to a segment sequence, or path, through the overall segment-graph that is associated with a subset of all possible feature vectors in the total observation space. In this work we examine a maximum a posteriori decoding strategy for feature-based recognizers and develop a normalization criterion useful for a segment-based Viterbi or A^* search. Experiments are reported for both phonetic and word recognition tasks.     

Combining diverse on-line and off-line systems for handwritten text line recognition

In this paper we present a multiple classifier system (MCS) for on-line handwriting recognition. The MCS combines several individual recognition systems based on hidden Markov models (HMMs) and bidirectional long short-term memory networks (BLSTM). Beside using two different recognition architectures (HMM and BLSTM), we use various feature sets based on on-line and off-line features to obtain diverse recognizers. Furthermore, we generate a number of different neural network recognizers by changing the initialization parameters. To combine the word sequences output by the recognizers, we incrementally align these sequences using the recognizer output voting error reduction framework (ROVER). For deriving the final decision, different voting strategies are applied. The best combination ensemble has a recognition rate of 84.13%, which is significantly higher than the 83.64% achieved if only one recognition architecture (HMM or BLSTM) is used for the combination, and even remarkably higher than the 81.26% achieved by the best individual classifier. To demonstrate the high performance of the classification system, the results are compared with two widely used commercial recognizers from Microsoft and Vision Objects.     

Efficient Development of Lexical Language Resources and their Representation

Statistical approaches in speech technology, whether used for statistical language models, trees, hidden Markov models or neural networks, represent the driving forces for the creation of language resources (LR), e.g., text corpora, pronunciation and morphology lexicons, and speech databases. This paper presents a system architecture for the rapid construction of morphologic and phonetic lexicons, two of the most important written language resources for the development of ASR (automatic speech recognition) and TTS (text-to-speech) systems. The presented architecture is modular and is particularly suitable for the development of written language resources for inflectional languages. In this paper an implementation is presented for the Slovenian language. The integrated graphic user interface focuses on the morphological and phonetic aspects of language and allows experts to produce good performances during analysis. In multilingual TTS systems, many extensive external written language resources are used, especially in the text processing part. It is very important, therefore, that representation of these resources is time and space efficient. It is also very important that language resources for new languages can be easily incorporated into the system, without modifying the common algorithms developed for multiple languages. In this regard the use of large external language resources (e.g., morphology and phonetic lexicons) represent an important problem because of the required space and slow look-up time. This paper presents a method and its results for compiling large lexicons, using examples for compiling German phonetic and morphology lexicons (CISLEX), and Slovenian phonetic (SIflex) and morphology (SImlex) lexicons, into corresponding finite-state transducers (FSTs). The German lexicons consisted of about 300,000 words, SIflex consisted of about 60,000 and SImlex of about 600,000 words (where 40,000 words were used for representation using finite-state transducers). Representation of large lexicons using finite-state transducers is mainly motivated by considerations of space and time efficiency. A great reduction in size and optimal access time was achieved for all lexicons. The starting size for the German phonetic lexicon was 12.53 MB and 18.49 MB for the morphology lexicon. The starting size for the Slovenian phonetic lexicon was 1.8 MB and 1.4 MB for the morphology lexicon. The final size of the corresponding FSTs was 2.78 MB for the German phonetic lexicon, 6.33 MB for the German morphology lexicon, 253 KB for SIflex and 662 KB for the SImlex lexicon. The achieved look-up time is optimal, since it only depends on the length of the input word and not on the size of the lexicon. Integration of lexicons for new languages into the multilingual TTS system is easy when using such representations and does not require any changes in the algorithms used for such lexicons.     

A Methodology for a Semi-Automatic Evaluation of the Lexicons of Machine Translation Systems

The lexicon is a major part of any Machine Translation (MT) system. If the lexicon of an MT system is not adequate, this will affect the quality of the whole system. Building a comprehensive lexicon, i.e., one with a high lexical coverage, is a major activity in the process of developing a good MT system. As such, the evaluation of the lexicon of an MT system is clearly a pivotal issue for the process of evaluating MT systems. In this paper, we introduce a new methodology that was devised to enable developers and users of MT Systems to evaluate their lexicons semi-automatically. This new methodology is based on the idea of the importance of a specific word or, more precisely, word sense, to a given application domain. This importance, or weight, determines how the presence of such a word in, or its absence from, the lexicon affects the MT system's lexical quality, which in turn will naturally affect the overall output quality. The method, which adopts a black-box approach to evaluation, was implemented and applied to evaluating the lexicons of three commercialEnglish–Arabic MT systems. A specific domain was chosen in which the various word-sense weights were determined by feeding sample texts from the domain into a system developed specifically for that purpose. Once this database of word senses and weights was built, test suites were presented to each of the MT systems under evaluation and their output rated by a human operator as either correct or incorrect. Based on this rating, an overall automated evaluation of the lexicons of the systems was deduced.     

Large vocabulary recognition of on-line handwritten cursive words

This paper presents a writer independent system for large vocabulary recognition of on-line handwritten cursive words. The system first uses a filtering module, based on simple letter features, to quickly reduce a large reference dictionary (lexicon) to a more manageable size; the reduced lexicon is subsequently fed to a recognition module. The recognition module uses a temporal representation of the input, instead of a static two-dimensional image, thereby preserving the sequential nature of the data and enabling the use of a Time-Delay Neural Network (TDNN); such networks have been previously successful in the continuous speech recognition domain. Explicit segmentation of the input words into characters is avoided by sequentially presenting the input word representation to the neural network-based recognizer. The outputs of the recognition module are collected and converted into a string of characters that is matched against the reduced lexicon using an extended Damerau-Levenshtein function. Trained on 2,443 unconstrained word images (11 k characters) from 55 writers and using a 21 k lexicon we reached a 97.9% and 82.4% top-5 word recognition rate on a writer-dependent and writer-independent test, respectively     

Towards speech recognizer assessment using a human reference standard

The measurement of the word error rate (WER) of a speech recognizer is valuable for the development of new algorithms but provides only the most limited information about the performance of the recognizer. We propose the use of a human reference standard to assess the performance of speech recognizers, so that the performance of a recognizer could be quoted as being equivalent to the performance of a human hearing speech which is subject to X dB of degradation. This approach should have the major advantage of being independent of the database and speakers used for testing. Furthermore, it would allow factors beyond the word error rate to be measured, such as the performance within an interactive speech system. In this paper, we report on preliminary work to explore the viability of this approach. This has consisted of recording a suitable database for experimentation, devising a method of degrading the speech in a controlled way and conducting two set of experiments on listeners to measure their responses to degraded speech to establish a reference. Results from these experiments raise several questions about the technique but encourage us to experiment with comparisons with automatic recognizers.     

Effect of acoustic and linguistic contexts on human and machine speech recognition

We compared the performance of an automatic speech recognition system using n-gram language models, HMM acoustic models, as well as combinations of the two, with the word recognition performance of human subjects who either had access to only acoustic information, had information only about local linguistic context, or had access to a combination of both. All speech recordings used were taken from Japanese narration and spontaneous speech corpora.Humans have difficulty recognizing isolated words taken out of context, especially when taken from spontaneous speech, partly due to word-boundary coarticulation. Our recognition performance improves dramatically when one or two preceding words are added. Short words in Japanese mainly consist of post-positional particles (i.e. wa, ga, wo, ni, etc.), which are function words located just after content words such as nouns and verbs. So the predictability of short words is very high within the context of the one or two preceding words, and thus recognition of short words is drastically improved. Providing even more context further improves human prediction performance under text-only conditions (without acoustic signals). It also improves speech recognition, but the improvement is relatively small.Recognition experiments using an automatic speech recognizer were conducted under conditions almost identical to the experiments with humans. The performance of the acoustic models without any language model, or with only a unigram language model, were greatly inferior to human recognition performance with no context. In contrast, prediction performance using a trigram language model was superior or comparable to human performance when given a preceding and a succeeding word. These results suggest that we must improve our acoustic models rather than our language models to make automatic speech recognizers comparable to humans in recognition performance under conditions where the recognizer has limited linguistic context.     

Handwritten word recognition with character and inter-characterneural networks

An off-line handwritten word recognition system is described. Images of handwritten words are matched to lexicons of candidate strings. A word image is segmented into primitives. The best match between sequences of unions of primitives and a lexicon string is found using dynamic programming. Neural networks assign match scores between characters and segments. Two particularly unique features are that neural networks assign confidence that pairs of segments are compatible with character confidence assignments and that this confidence is integrated into the dynamic programming. Experimental results are provided on data from the U.S. Postal Service.     

A robust method for coarse classifier construction from a large number of basic recognizers for on-line handwritten Chinese/Japanese character recognition

In this paper, a systematic method is described that constructs an efficient and a robust coarse classifier from a large number of basic recognizers obtained by different parameters of feature extraction, different discriminant methods or functions, etc. The architecture of the coarse classification is a sequential cascade of basic recognizers that reduces the candidates after each basic recognizer. A genetic algorithm determines the best cascade with the best speed and highest performance. The method was applied for on-line handwritten Chinese and Japanese character recognitions. We produced hundreds of basic recognizers with different classification costs and different classification accuracies by changing parameters of feature extraction and discriminant functions. From these basic recognizers, we obtained a rather simple two-stage cascade, resulting in the whole recognition time being reduced largely while maintaining classification and recognition rates.     

Spoken language recognition on a DSP array processor

A new architecture is presented to support the general class of real-time large-vocabulary speaker-independent continuous speech recognizers incorporating language models. Many such recognizers require multiple high-performance central processing units (CPU's) as well as high interprocessor communication bandwidth. This array processor provides a peak CPU performance of 2.56 giga-floating point operations per second (GFLOPS) as well as a high-speed communication network. In order to efficiently utilize these resources, algorithms were devised for partitioning speech models for mapping into the array processor. Also, a novel scheme is presented for a functional partitioning of the speech recognizer computations. The recognizer is functionally partitioned into six stages, namely, the linear predictive coding (LPC) based feature extractor, mixture probability computer, (phone) state probability computer, word probability computer, phrase probability computer, and traceback computer. Each of these stages is further subdivided as many times as necessary to fit the individual processing elements (PE's). The functional stages are pipelined and synchronized with the frame rate of the incoming speech signal. This partitioning also allows a multistage stack decoder to be implemented for reduction of computation     

Hierarchical Singleton-Type Recurrent Neural Fuzzy Networks for Noisy Speech Recognition

This paper proposes noisy speech recognition using hierarchical singleton-type recurrent neural fuzzy networks (HSRNFNs). The proposed HSRNFN is a hierarchical connection of two singleton-type recurrent neural fuzzy networks (SRNFNs), where one is used for noise filtering and the other for recognition. The SRNFN is constructed by recurrent fuzzy if-then rules with fuzzy singletons in the consequences, and their recurrent properties make them suitable for processing speech patterns with temporal characteristics. In n words recognition, n SRNFNs are created for modeling n words, where each SRNFN receives the current frame feature and predicts the next one of its modeling word. The prediction error of each SRNFN is used as recognition criterion. In filtering, one SRNFN is created, and each SRNFN recognizer is connected to the same SRNFN filter, which filters noisy speech patterns in the feature domain before feeding them to the SRNFN recognizer. Experiments with Mandarin word recognition under different types of noise are performed. Other recognizers, including multilayer perceptron (MLP), time-delay neural networks (TDNNs), and hidden Markov models (HMMs), are also tested and compared. These experiments and comparisons demonstrate good results with HSRNFN for noisy speech recognition tasks     

MLN-based Bangla ASR using context sensitive triphone HMM

Building a continuous speech recognizer for the Bangla (widely used as Bengali) language is a challenging task due to the unique inherent features of the language like long and short vowels and many instances of allophones. Stress and accent vary in spoken Bangla language from region to region. But in formal read Bangla speech, stress and accents are ignored. There are three approaches to continuous speech recognition (CSR) based on the sub-word unit viz. word, phoneme and syllable. Pronunciation of words and sentences are strictly governed by set of linguistic rules. Many attempts have been made to build continuous speech recognizers for Bangla for small and restricted tasks. However, medium and large vocabulary CSR for Bangla is relatively new and not explored. In this paper, the authors have attempted for building automatic speech recognition (ASR) method based on context sensitive triphone acoustic models. The method comprises three stages, where the first stage extracts phoneme probabilities from acoustic features using a multilayer neural network (MLN), the second stage designs triphone models to catch context of both sides and the final stage generates word strings based on triphone hidden Markov models (HMMs). The objective of this research is to build a medium vocabulary triphone based continuous speech recognizer for Bangla language. In this experimentation using Bangla speech corpus prepared by us, the recognizer provides higher word accuracy as well as word correct rate for trained and tested sentences with fewer mixture components in HMMs.