Evaluating retraining rules for semi-supervised learning in neural network based cursive word recognition

Training a system to recognize handwritten words is a task that requires a large amount of data with their correct transcription. However, the creation of such a training set, including the generation of the ground truth, is tedious and costly. One way of reducing the high cost of labeled training data acquisition is to exploit unlabeled data, which can be gathered easily. Making use of both labeled and unlabeled data is known as semi-supervised learning. One of the most general versions of semi-supervised learning is self-training, where a recognizer iteratively retrains itself on its own output on new, unlabeled data. In this paper we propose to apply semi-supervised learning, and in particular self-training, to the problem of cursive, handwritten word recognition. The special focus of the paper is on retraining rules that define what data are actually being used in the retraining phase. In a series of experiments it is shown that the performance of a neural network based recognizer can be significantly improved through the use of unlabeled data and self-training if appropriate retraining rules are applied.

Automaticity in L2 learning: Correlation between vocabulary proficiency and response time in word recognition

Automaticity (in this essay defined as short response time) and fluency in language use are closely connected to each other and some research has been conducted regarding some of the aspects involved. In fact, the notion of automaticity is still debated and many definitions and opinions on what automaticity is have been suggested (Andersson,1987, 1992, 1993, Logan, 1988, Segalowitz, 2010). One aspect that still needs more research is the correlation between vocabulary proficiency (a person’s knowledge about words and ability to use them correctly) and response time in word recognition. Therefore, the aim of this study has been to investigate this correlation using two different tests; one vocabulary size test (Paul Nation) and one lexical decision task (SuperLab) that measures both response time and accuracy. 23 Swedish students partaking in the English 7 course in upper secondary Swedish school were tested. The data were analyzed using a quantitative method where the average values and correlations from the test were used to compare the results. The correlations were calculated using Pearson’s Coefficient Correlations Calculator. The empirical study indicates that vocabulary proficiency is not strongly correlated with shorter response times in word recognition. Rather, the data indicate that L2 learners instead are sensitive to the frequency levels of the vocabulary. The accuracy (number of correct recognized words) and response times correlate with the frequency level of the tested words. This indicates that factors other than vocabulary proficiency are important for the ability to recognize words quickly.

Attentional shifting and the role of the dorsal pathway in visual word recognition

A substantial amount of evidence has been collected to propose an exclusive role for the dorsal visual pathway in the control of guided visual search mechanisms, specifically in the preattentive direction of spatial selection [Vidyasagar, T. R. (1999). A neuronal model of attentional spotlight: Parietal guiding the temporal. Brain Research and Reviews, 30, 66-76; Vidyasagar, T. R. (2001). From attentional gating in macaque primary visual cortex to dyslexia in humans. Progress in Brain Research, 134, 297-312]. Moreover, it has been suggested recently that the dorsal visual pathway is specifically involved in the spatial selection and sequencing required for orthographic processing in visual word recognition. In this experiment we manipulate the demands for spatial processing in a word recognition, lexical decision task by presenting target words in a normal spatial configuration, or where the constituent letters of each word are spatially shifted relative to each other. Accurate word recognition in the Shifted-words condition should demand higher spatial encoding requirements, thereby making greater demands on the dorsal visual stream. Magnetoencephalographic (MEG) neuroimaging revealed a high frequency (35-40 Hz) right posterior parietal activation consistent with dorsal stream involvement occurring between 100 and 300 ms post-stimulus onset, and then again at 200-400 ms. Moreover, this signal was stronger in the shifted word condition, compared to the normal word condition. This result provides neurophysiological evidence that the dorsal visual stream may play an important role in visual word recognition and reading. These results further provide a plausible link between early stage theories of reading, and the magnocellular-deficit theory of dyslexia, which characterises many types of reading difficulty. © 2006 Elsevier Ltd. All rights reserved.

Visual word recognition:The first half second

We used magnetoencephalography (MEG) to map the spatiotemporal evolution of cortical activity for visual word recognition. We show that for five-letter words, activity in the left hemisphere (LH) fusiform gyrus expands systematically in both the posterior-anterior and medial-lateral directions over the course of the first 500 ms after stimulus presentation. Contrary to what would be expected from cognitive models and hemodynamic studies, the component of this activity that spatially coincides with the visual word form area (VWFA) is not active until around 200 ms post-stimulus, and critically, this activity is preceded by and co-active with activity in parts of the inferior frontal gyrus (IFG, BA44/6). The spread of activity in the VWFA for words does not appear in isolation but is co-active in parallel with spread of activity in anterior middle temporal gyrus (aMTG, BA 21 and 38), posterior middle temporal gyrus (pMTG, BA37/39), and IFG. © 2004 Elsevier Inc. All rights reserved.

The role of age and emotional valence in word recognition: an ex-Gaussian analysis

The aim of this work is to evaluate the roles of age and emotional valence in word recognition in terms of ex-Gaussian distribution components. In order to do that, a word recognition task was carried out with two age groups, in which emotional valence was manipulated. Older participants did not present a clear trend for reaction times. The younger participants showed significant statistical differences in negative words for target and distracting conditions. Addressing the ex-Gaussian tau parameter, often related to attentional demands in the literature, age-related differences in emotional valence seem not to have an effect for negative words. Focusing on emotional valence for each group, the younger participants only showed an effect on negative distracting words. The older participants showed an effect regarding negative and positive target words, and negative distracting words. This suggests that the attentional demand is higher for emotional words, in particular, for the older participants.

Power-law transformation for enhanced recognition of born-digital word images

In this paper, we discuss the issues related to word recognition in born-digital word images. We introduce a novel method of power-law transformation on the word image for binarization. We show the improvement in image binarization and the consequent increase in the recognition performance of OCR engine on the word image. The optimal value of gamma for a word image is automatically chosen by our algorithm with fixed stroke width threshold. We have exhaustively experimented our algorithm by varying the gamma and stroke width threshold value. By varying the gamma value, we found that our algorithm performed better than the results reported in the literature. On the ICDAR Robust Reading Systems Challenge-1: Word Recognition Task on born digital dataset, as compared to the recognition rate of 61.5% achieved by TH-OCR after suitable pre-processing by Yang et. al. and 63.4% by ABBYY Fine Reader (used as baseline by the competition organizers without any preprocessing), we achieved 82.9% using Omnipage OCR applied on the images after being processed by our algorithm.

Benchmarking recognition results on camera captured word image data sets

We have benchmarked the maximum obtainable recognition accuracy on five publicly available standard word image data sets using semi-automated segmentation and a commercial OCR. These images have been cropped from camera captured scene images, born digital images (BDI) and street view images. Using the Matlab based tool developed by us, we have annotated at the pixel level more than 3600 word images from the five data sets. The word images binarized by the tool, as well as by our own midline analysis and propagation of segmentation (MAPS) algorithm are recognized using the trial version of Nuance Omnipage OCR and these two results are compared with the best reported in the literature. The benchmark word recognition rates obtained on ICDAR 2003, Sign evaluation, Street view, Born-digital and ICDAR 2011 data sets are 83.9%, 89.3%, 79.6%, 88.5% and 86.7%, respectively. The results obtained from MAPS binarized word images without the use of any lexicon are 64.5% and 71.7% for ICDAR 2003 and 2011 respectively, and these values are higher than the best reported values in the literature of 61.1% and 41.2%, respectively. MAPS results of 82.8% for BDI 2011 dataset matches the performance of the state of the art method based on power law transform.

NESP: Nonlinear enhancement and selection of plane for optimal segmentation and recognition of scene word images

In this paper, we report a breakthrough result on the difficult task of segmentation and recognition of coloured text from the word image dataset of ICDAR robust reading competition challenge 2: reading text in scene images. We split the word image into individual colour, gray and lightness planes and enhance the contrast of each of these planes independently by a power-law transform. The discrimination factor of each plane is computed as the maximum between-class variance used in Otsu thresholding. The plane that has maximum discrimination factor is selected for segmentation. The trial version of Omnipage OCR is then used on the binarized words for recognition. Our recognition results on ICDAR 2011 and ICDAR 2003 word datasets are compared with those reported in the literature. As baseline, the images binarized by simple global and local thresholding techniques were also recognized. The word recognition rate obtained by our non-linear enhancement and selection of plance method is 72.8% and 66.2% for ICDAR 2011 and 2003 word datasets, respectively. We have created ground-truth for each image at the pixel level to benchmark these datasets using a toolkit developed by us. The recognition rate of benchmarked images is 86.7% and 83.9% for ICDAR 2011 and 2003 datasets, respectively.

Robust speech recognition using speech enhancement

Automatic Speech Recognition (ASR) has matured into a technology which is becoming more common in our everyday lives, and is emerging as a necessity to minimise driver distraction when operating in-car systems such as navigation and infotainment. In “noise-free” environments, word recognition performance of these systems has been shown to approach 100%, however this performance degrades rapidly as the level of background noise is increased. Speech enhancement is a popular method for making ASR systems more ro- bust. Single-channel spectral subtraction was originally designed to improve hu- man speech intelligibility and many attempts have been made to optimise this algorithm in terms of signal-based metrics such as maximised Signal-to-Noise Ratio (SNR) or minimised speech distortion. Such metrics are used to assess en- hancement performance for intelligibility not speech recognition, therefore mak- ing them sub-optimal ASR applications. This research investigates two methods for closely coupling subtractive-type enhancement algorithms with ASR: (a) a computationally-efficient Mel-filterbank noise subtraction technique based on likelihood-maximisation (LIMA), and (b) in- troducing phase spectrum information to enable spectral subtraction in the com- plex frequency domain. Likelihood-maximisation uses gradient-descent to optimise parameters of the enhancement algorithm to best fit the acoustic speech model given a word se- quence known a priori. Whilst this technique is shown to improve the ASR word accuracy performance, it is also identified to be particularly sensitive to non-noise mismatches between the training and testing data. Phase information has long been ignored in spectral subtraction as it is deemed to have little effect on human intelligibility. In this work it is shown that phase information is important in obtaining highly accurate estimates of clean speech magnitudes which are typically used in ASR feature extraction. Phase Estimation via Delay Projection is proposed based on the stationarity of sinusoidal signals, and demonstrates the potential to produce improvements in ASR word accuracy in a wide range of SNR. Throughout the dissertation, consideration is given to practical implemen- tation in vehicular environments which resulted in two novel contributions – a LIMA framework which takes advantage of the grounding procedure common to speech dialogue systems, and a resource-saving formulation of frequency-domain spectral subtraction for realisation in field-programmable gate array hardware. The techniques proposed in this dissertation were evaluated using the Aus- tralian English In-Car Speech Corpus which was collected as part of this work. This database is the first of its kind within Australia and captures real in-car speech of 50 native Australian speakers in seven driving conditions common to Australian environments.

Towards improved speech recognition for resource poor languages

In recent times, the improved levels of accuracy obtained by Automatic Speech Recognition (ASR) technology has made it viable for use in a number of commercial products. Unfortunately, these types of applications are limited to only a few of the world’s languages, primarily because ASR development is reliant on the availability of large amounts of language specific resources. This motivates the need for techniques which reduce this language-specific, resource dependency. Ideally, these approaches should generalise across languages, thereby providing scope for rapid creation of ASR capabilities for resource poor languages. Cross Lingual ASR emerges as a means for addressing this need. Underpinning this approach is the observation that sound production is largely influenced by the physiological construction of the vocal tract, and accordingly, is human, and not language specific. As a result, a common inventory of sounds exists across languages; a property which is exploitable, as sounds from a resource poor, target language can be recognised using models trained on resource rich, source languages. One of the initial impediments to the commercial uptake of ASR technology was its fragility in more challenging environments, such as conversational telephone speech. Subsequent improvements in these environments has gained consumer confidence. Pragmatically, if cross lingual techniques are to considered a viable alternative when resources are limited, they need to perform under the same types of conditions. Accordingly, this thesis evaluates cross lingual techniques using two speech environments; clean read speech and conversational telephone speech. Languages used in evaluations are German, Mandarin, Japanese and Spanish. Results highlight that previously proposed approaches provide respectable results for simpler environments such as read speech, but degrade significantly when in the more taxing conversational environment. Two separate approaches for addressing this degradation are proposed. The first is based on deriving better target language lexical representation, in terms of the source language model set. The second, and ultimately more successful approach, focuses on improving the classification accuracy of context-dependent (CD) models, by catering for the adverse influence of languages specific phonotactic properties. Whilst the primary research goal in this thesis is directed towards improving cross lingual techniques, the catalyst for investigating its use was based on expressed interest from several organisations for an Indonesian ASR capability. In Indonesia alone, there are over 200 million speakers of some Malay variant, provides further impetus and commercial justification for speech related research on this language. Unfortunately, at the beginning of the candidature, limited research had been conducted on the Indonesian language in the field of speech science, and virtually no resources existed. This thesis details the investigative and development work dedicated towards obtaining an ASR system with a 10000 word recognition vocabulary for the Indonesian language.

Low-cost hardware speech enhancement for improved speech recognition in automotive environments

Voice recognition is one of the key enablers to reduce driver distraction as in-vehicle systems become more and more complex. With the integration of voice recognition in vehicles, safety and usability are improved as the driver’s eyes and hands are not required to operate system controls. Whilst speaker independent voice recognition is well developed, performance in high noise environments (e.g. vehicles) is still limited. La Trobe University and Queensland University of Technology have developed a low-cost hardware-based speech enhancement system for automotive environments based on spectral subtraction and delay–sum beamforming techniques. The enhancement algorithms have been optimised using authentic Australian English collected under typical driving conditions. Performance tests conducted using speech data collected under variety of vehicle noise conditions demonstrate a word recognition rate improvement in the order of 10% or more under the noisiest conditions. Currently developed to a proof of concept stage there is potential for even greater performance improvement.

Multiple cameras for audio-visual speech recognition in an automotive environment

Audio-visualspeechrecognition, or the combination of visual lip-reading with traditional acoustic speechrecognition, has been previously shown to provide a considerable improvement over acoustic-only approaches in noisy environments, such as that present in an automotive cabin. The research presented in this paper will extend upon the established audio-visualspeechrecognition literature to show that further improvements in speechrecognition accuracy can be obtained when multiple frontal or near-frontal views of a speaker's face are available. A series of visualspeechrecognition experiments using a four-stream visual synchronous hidden Markov model (SHMM) are conducted on the four-camera AVICAR automotiveaudio-visualspeech database. We study the relative contribution between the side and central orientated cameras in improving visualspeechrecognition accuracy. Finally combination of the four visual streams with a single audio stream in a five-stream SHMM demonstrates a relative improvement of over 56% in word recognition accuracy when compared to the acoustic-only approach in the noisiest conditions of the AVICAR database.

Multi Pattern Dynamic Time Warping For Automatic Speech Recognition

We are addressing the problem of jointly using multiple noisy speech patterns for automatic speech recognition (ASR), given that they come from the same class. If the user utters a word K times, the ASR system should try to use the information content in all the K patterns of the word simultaneously and improve its speech recognition accuracy compared to that of the single pattern based speech recognition. T address this problem, recently we proposed a Multi Pattern Dynamic Time Warping (MPDTW) algorithm to align the K patterns by finding the least distortion path between them. A Constrained Multi Pattern Viterbi algorithm was used on this aligned path for isolated word recognition (IWR). In this paper, we explore the possibility of using only the MPDTW algorithm for IWR. We also study the properties of the MPDTW algorithm. We show that using only 2 noisy test patterns (10 percent burst noise at -5 dB SNR) reduces the noisy speech recognition error rate by 37.66 percent when compared to the single pattern recognition using the Dynamic Time Warping algorithm.

Joint evaluation of multiple speech patterns for speech recognition and training

We are addressing the novel problem of jointly evaluating multiple speech patterns for automatic speech recognition and training. We propose solutions based on both the non-parametric dynamic time warping (DTW) algorithm, and the parametric hidden Markov model (HMM). We show that a hybrid approach is quite effective for the application of noisy speech recognition. We extend the concept to HMM training wherein some patterns may be noisy or distorted. Utilizing the concept of ``virtual pattern'' developed for joint evaluation, we propose selective iterative training of HMMs. Evaluating these algorithms for burst/transient noisy speech and isolated word recognition, significant improvement in recognition accuracy is obtained using the new algorithms over those which do not utilize the joint evaluation strategy.