Lexicons are important resources for semantic tagging. However, commonly used lexicons collected from entity databases suffer from multiple problems, such as ambiguity, limited coverage and lack of relative importance. In this work we present a lexicon modeling technique that automatically expands the lexicon and assigns weights to its elements. For lexicon expansion, we use a generative model to extract patterns from query logs using known lexicon seeds, and discover new lexicon elements using the learned patterns. For lexicon weighting, we propose two approaches based on generative and discriminative models to learn the relative importance of lexicon elements from user click statistics. Experiments on text queries in multiple domains show that our lexicon modeling technique can significantly improve semantic tagging performance.

This paper addresses the problem of learning multiple spoken language understanding (SLU) tasks that have overlapping sets of slots. In such a scenario, it is possible to achieve better slot filling performance by learning multiple tasks simultaneously, as opposed to learning them independently. We focus on pre-senting a number of simple multi-task learning algorithms for slot filling systems based on semi-Markov CRFs, assuming the knowledge of shared slots. Furthermore, we discuss an intra-domain clustering method that automatically discovers shared slots from training data. The effectiveness of our proposed ap-proaches is demonstrated in an SLU application that involves three different yet related tasks. Index Terms: spoken language understanding, slot filling, multi-task learning

This paper addresses the problem of discriminative training of language models that does not require any transcribed acoustic data. We propose to minimize the conditional entropy of word sequences given phone sequences, and present two settings in which this criterion can be applied. In an inductive learning setting, the phonetic/acoustic confusability information is given by a general phone error model. A transductive approach, in contrast, obtains that information by running a speech recognizer on test-set acoustics, with the goal of optimizing the test-set performance. Experiments show significant recognition accuracy improvements in both rescoring and first-pass decoding experiments using the transductive approach, and mixed results using the inductive approach.

Voice search technology has been successfully applied to help drivers reply SMS messages in automobiles, in which a predefined SMS message template set is searched with ASR hypotheses to form the reply candidate list. In order to efficiently organize the SMS message template set and improve the quality of the reply candidate list, we proposed to apply n-gram translation model and logistic regression to detect paraphrase SMS messages. Both of the proposed algorithms outperform the edit distance based paraphrase detection baseline, brining 40.9% and 50.5% EER reduction (relative), respectively.

Topical query classification, as one step toward understanding users' search intent, is gaining increasing attention in information retrieval. Previous works on this subject primarily focused on enrichment of query features, for example, by augmenting queries with search engine results. In this work, we investigate a completely orthogonal approach—instead of improving feature representation, we aim at drastically increasing the amount of training data. To this end, we propose two semisupervised learning methods that exploit user click-through data. In one approach, we infer class memberships of unlabeled queries from those of labeled ones according to their proximities in a click graph; and then use these automatically labeled queries to train classifiers using query terms as features. In a second approach, click graph learning and query classifier training are conducted jointly with an integrated objective. Our methods are evaluated in two applications, product intent and job intent classification. In both cases, we expand the training data by over two orders of magnitude, leading to significant improvements in classification performance. An additional finding is that with a large amount of training data obtained in this fashion, a classifier based on simple query term features can outperform those using state-of-the-art, augmented features.

Web query classification is an effective way to understand Web user intents, which can further improve Web search and online advertising relevance. However, Web queries are usually very short which cannot fully reflect their meanings. What is more, it is quite hard to obtain enough training data for training accurate classifiers. Therefore, previous work on query classification has focused on two issues. One is how to represent Web queries through query expansion. The other is how to increase the amount of training data. In this paper, we took product query classification as an example, which is to classify Web queries into a predefined product taxonomy, and systematically studied the impact of query expansion and the size of training data. We proposed two methods of enriching Web queries and three approaches of collecting training data. Thereafter, we conducted a series of experiments to compare the classification performance of using different combinations of training data and query representations over a real data set. The data set consists of hundreds of thousands queries collected from a popular commercial search engine. From the experiments, we found some interesting observations, which were not discussed before. Finally, we proposed an effective and efficient product query classification method based on our observations.

Understanding intents from search queries can improve a user's search experience and boost a site's advertising profits. Query tagging via statistical sequential labeling models has been shown to perform well, but annotating the training set for supervised learning requires substantial human effort. Domain-specific knowledge, such as semantic class lexicons, reduces the amount of needed manual annotations, but much human effort is still required to maintain these as search topics evolve over time.

We introduce a direct model for speech recognition that assumes an unstructured, i.e., flat text output. The flat model allows us to model arbitrary attributes and dependences of the output. This is different from the HMMs typically used for speech recognition. This conventional modeling approach is based on sequential data and makes rigid assumptions on the dependences. HMMs have proven to be convenient and appropriate for large vocabulary continuous speech recognition. Our task under consideration, however, is the Windows Live Search for Mobile (WLS4M) task. This is a cellphone application that allows users to interact with web-based information portals. In particular, the set of valid outputs can be considered discrete and finite (although probably large, i.e., unseen events are an issue). Hence, a flat direct model lends itself to this task, making the adding of different knowledge sources and dependences straightforward and cheap. Using e.g. HMM posterior, m-gram, and spotter features, significant improvements over the conventional HMM system were observed.

A voice search system requires a speech interface that can correctly recognize spoken queries uttered by users. The recognition performance strongly relies on a robust language model. In this work, we present the use of multiple data sources, with the focus on query logs, in improving ASR language models for a voice search application. Our contributions are three folds: (1) the use of text queries from web search and mobile search in language modeling; (2) the use of web click data to predict query forms from business listing forms; and (3) the use of voice query logs in creating a positive feedback loop. Experiments show that by leveraging these resources, we can achieve recognition performance comparable to, or even better than, that of a previously deploy system where a large amount of spoken query transcripts are used in language modeling.

When search is against structured documents, it is beneficial to extract information from user queries in a format that is consistent with the backend data structure. As one step toward this goal, we study the problem of query tagging which is to assign each query term to a pre-defined category. Our problem could be approached by learning a conditional random field (CRF) model (or other statistical models) in a supervised fashion, but this would require substantial human-annotation effort. In this work, we focus on a semi-supervised learning method for CRFs that utilizes two data sources: (1) a small amount of manually-labeled queries, and (2) a large amount of queries in which some word tokens have derived labels, i.e., label information automatically obtained from additional resources. We present two principled ways of encoding derived label information in a CRF model. Such information is viewed as hard evidence in one setting and as soft evidence in the other. In addition to the general methodology of how to use derived labels in semi-supervised CRFs, we also present a practical method on how to obtain them by leveraging user click data and an in-domain database that contains structured documents. Evaluation on product search queries shows the effectiveness of our approach in improving tagging accuracies.

This paper presents a novel approach to language modeling for voice search based on the idea and method of statistical machine translation. We propose an n-gram based translation model that can be used for listing-to-query translation. We then leverage the query forms translated from listings to improve language modeling. The translation model is trained in an unsupervised manner using a set of transcribed voice search queries. Experiments show that the translation approach yielded drastic perplexity reductions compared with a baseline language model where no translation is applied.

Text classification has been widely applied to many practical tasks. Inductive models trained from labeled data are the most commonly used technique. The basic assumption underlying an inductive model is that the training data are drawn from the same distribution as the test data. However, labeling such a training set is often expensive for practical applications. On the other hand, a large amount of labeled data, which have been drawn from a different distribution, is often available in the same application domain. It is thus very desirable to take advantage of these data even though there is a discrepancy between their underlying distribution and that of the test set. This paper compares three text classification algorithms applied in this scenario, including two inductive Maximum Entropy (MaxEnt) models, one flatly initialized and the other initialized with a term-frequency/inverse document frequency (Tf*Idf) weighted vector space model, and an example-based learning algorithm, which assigns a class label to a text by learning from the labels assigned to the training data that are similar to the text. Experiment results show that examplebased learning has achieved more than 5 % improvement in precisions across almost all coverage levels. Index Terms: text classification, inductive models, maximum entropy model, Tf*Idf vector space model, example-based learning. 1.

Due to speech recognition errors, repetition can be a frequent occurrence in voice-search applications. While a proper treatment of this phenomenon requires the joint modeling of two or more utterances simultaneously, currently deployed systems typically treat the utterances independently. In this paper, we analyze the structure of repetitions and find that in at least one commercial directory assistance application, repetitions follow simple structural transformations more than 70 % of the time. We present preliminary results that suggest that significant gains are possible by explicitly modeling this structure in a joint decoding process. Index Terms: speech recognition, minimum bayes risk, joint decoding, repeated utterances

This work presents the use of click graphs in improving query intent classifiers, which are critical if vertical search and general-purpose search services are to be offered in a unified user interface. Previous works on query classification have primarily focused on improving feature representation of queries, e.g., by augmenting queries with search engine results. In this work, we investigate a completely orthogonal approach --- instead of enriching feature representation, we aim at drastically increasing the amounts of training data by semi-supervised learning with click graphs. Specifically, we infer class memberships of unlabeled queries from those of labeled ones according to their proximities in a click graph. Moreover, we regularize the learning with click graphs by content-based classification to avoid propagating erroneous labels. We demonstrate the effectiveness of our algorithms in two different applications, product intent and job intent classification. In both cases, we expand the training data with automatically labeled queries by over two orders of magnitude, leading to significant improvements in classification performance. An additional finding is that with a large amount of training data obtained in this fashion, classifiers using only query words/phrases as features can work remarkably well.

Due to speech recognition errors, repetitions are a frequent phenomenon in spoken dialog systems. In previous work (G. Zweig et al., 2008) we have proposed a joint decoding model that can leverage structural relationships between repeated utterances for improving recognition performance. In this paper we extend this work in two directions. First, we propose a direct, classification-based model for the same task. The new model can leverage features that were fundamentally hard to capture in the previous framework (e.g. spellings, false-starts, etc.) and leads to an additional performance improvement. Second, we show how both models can be used to perform a combined rescoring of two n-best lists that are part of a repetition pair.

This work investigates the use of acoustic data to improve grapheme-to-phoneme conversion for name recognition. We introduce a joint model of acoustics and graphonemes, and present two approaches, maximum likelihood training and discriminative training, in adapting graphoneme model parameters. Experiments on a large-scale voice-dialing system show that the maximum likelihood approach yields a relative 7% reduction in SER compared to the best baseline result we obtained without leveraging acoustic data, while discriminative training enlarges the SER reduction to 12%.

Mouse control has become a crucial aspect of many modern day computer interactions. This poses a challenge for individuals with motor impairments or those whose use of hands are restricted due to situational constraints. We present a system called the Vocal Joystick which allows the user to continuously control the mouse cursor by varying vocal parameters such as vowel quality, loudness and pitch. A survey of existing cursor control methods is presented to highlight the key characteristics of the Vocal Joystick. Evaluations were conducted to characterize expert performance capability of the Vocal Joystick, and to compare novice user performance and preference for the Vocal Joystick and two other existing speech based cursor control methods. Our results show that Fitts' law is a good predictor of the speedaccuracy tradeoff for the Vocal Joystick, and suggests that the optimal performance of the Vocal Joystick may be comparable to that of a conventional hand-operated joystick. Novice user evaluations show that the Vocal Joystick can be used by people without extensive training, and that it presents a viable alternative to existing speech-based cursor control methods.

The Vocal Joystick is a novel human-computer interface mechanism designed to enable individuals with motor impairments to make use of vocal parameters to control objects on a computer screen (buttons, sliders, etc.) and ultimately electro-mechanical instruments (e.g., robotic arms, wireless home automation devices). We have developed a working prototype of our "VJ-engine" with which individuals can now control computer mouse movement with their voice. The core engine is currently optimized according to a number of criterion. In this paper, we describe the engine system design, engine optimization, and user-interface improvements, and outline some of the signal processing and pattern recognition modules that were successful. Lastly, we present new results comparing the vocal joystick with a state-of-the-art eye tracking pointing device, and show that not only is the Vocal Joystick already competitive, for some tasks it appears to be an improvement.

We introduce a novel method for adapting discriminative classifiers (multi-layer perceptrons (MLPs) and support vector machines (SVMs)). Ourmethod is based on the idea of regularization, whereby an optimization cost criterion to be minimized includes a penalty in accordance to how “complex” the system is. Specifically, our regularization term penalizes depending on how different an adapted system is from an unadapted system, thus avoiding the problem of overtraining when only a small amount of adaptation data is available. We justify this approach using a max-margin argument. We apply this technique to MLPs and produce a working real-time System for rapid adaptation of vowel classifiers in the context of the Vocal Joystick project. Overall, we find that our method outperforms all other MLP-based adaptation methods we are aware of. Our technique, however, is quite general and can be used whenever rapid adaptation of MLP or SVM classifiers are needed (e.g., from a speaker-independent to a speaker-dependent classifier in a hybrid MLP/HMM or SVM/HMM speech-recognition system).

Vocal Joystick is a mechanism that enables individuals with motor impairments to make use of vocal parameters to control objects on a computer screen (buttons, sliders, etc.) and ultimately will be used to control electro-mechanical instruments (e.g., robotic arms, wireless home automation devices). In an effort to train the VJ-system, speech data from the TIMIT speech corpus was initially used. However, due to problematic issues with co-articulation, we began a large data collection effort in a controlled environment that would not only address the problematic issues, but also yield a new vowel corpus that was representative of the utterances a user of the VJ-system would use. The data collection process evolved over the course of the effort as new parameters were added and as factors relating to the quality of the collected data in terms of the specified parameters were considered. The result of the data collection effort is a vowel corpus of approximately 11 hours of recorded data comprised of approximately 23500 sound files of the monophthongs and vowel combinations (e.g. diphthongs) chosen for the Vocal Joystick project varying along the parameters of duration, intensity and amplitude. This paper discusses how the data collection has evolved since its initiation and provides a brief summary of the resulting corpus. Index Terms: Speech corpora, data collection procedures, speech recognition, Speech HCI for individuals with impairments, Speech/voice-based human-computer interfaces 1.

This paper introduces a novel adaptive direction filtering algorithm in the Vocal Joystick (VJ) setting that utilizes context information and applies real-time inference in a continuous space. The VJ system using this algorithm is endowed with the ability to produce movements in arbitrary directions and the ability to draw smooth curves. This is in contrast to previous VJ settings whereby vowel quality was used to determine mouse movement in only a finite discrete set of directions [1]. Index Terms: computer interface, voice control, adaptive filtering 1.

With the skyrocketing popularity of mobile devices, new processing methods tailored to a specific application have become necessary for low-resource systems. This work presents a high-speed, low-resource speech recognition system using custom arithmetic units, where all system variables are represented by integer indices and all arithmetic operations are replaced by hardware-based table lookups. To this end, several reordering and rescaling techniques, including two accumulation structures for Gaussian evaluation and a novel method for the normalization of Viterbi search scores, are proposed to ensure low entropy for all variables. Furthermore, a discriminatively inspired distortion measure is investigated for scalar quantization of forward probabilities to maximize the recognition rate. Finally, heuristic algorithms are explored to optimize system-wide resource allocation. Our best bit-width allocation scheme only requires 59 kB of ROMs to hold the lookup tables, and its recognition performance with various vocabulary sizes in both clean and noisy conditions is nearly as good as that of a system using a 32-bit floating-point unit. Simulations on various architectures show that, on most modern processor designs, we can expect a cycle-count speedup of at least three times over systems with floating-point units. Additionally, the memory bandwidth is reduced by over 70% and the offline storage for model parameters is reduced by 80%.

Unforeseen user intents can account for a significant portion of unsuccessful calls in an automatic voice response system. Discovering these unforeseen semantic intents usually requires expensive manual transcriptions. We propose a method to cluster the acoustics from logged calls by their estimated semantic intents. This is achieved through training a mixture of language models in an unsupervised manner. Each cluster is presented to the application developer with a suggested language model to cover the semantic intent of the data in that cluster. The application developer validates the cluster and its suggested language model, and then updates the application. A quantitative evaluation on a corporate voice-dialer application shows that updating the application in this manner yields a relative 13.4% reduction in semantic error rate.

We present a novel approach to estimating the first two formants (F1 and F2) of a speech signal using graphical models. Using a graph that takes advantage of less commonly used features of Bayesian networks, both v-structures and soft evidence, the model presented here shows that it can learn to perform reasonably without large amounts of training data, even with minimal processing on the initial signal. It far outperforms a factorial HMM using the same assumptions and suggests that with further refinement the model may produce high quality formant tracks.

Likelihood evaluation can substantially affect the total computational load for continuous hidden Markov model (HMM)-based speech-recognition systems with small vocabularies. This letter presents feature pruning , a simple yet effective technique to reduce computation and, hence, power consumption of likelihood evaluation. Our technique, under certain conditions, only evaluates the likelihoods of a fraction of feature elements and approximates those of the remaining (pruned) ones by a simple function. The order in which feature elements are evaluated is obtained by a data-driven approach to minimize computation. With this order, feature pruning can speed up the likelihood evaluation by a factor of 1.3-1.8 and reduce its power consumption by 27%-43% for various recognition tasks, including those in noisy environments.

Custom arithmetic is a novel and successful technique to reduce the computation and resource utilization of ASR systems running on mobile devices. It represents all floating-point numbers by integer indices and substitutes a sequence of table lookups for all arithmetic operations. The first and crucial step in custom arithmetic design is to quantize system variables, preferably to low precision. This paper explores several techniques to quantize variables with high entropy, including a reordering of Gaussian computation and a normalization of Viterbi search. Furthermore, a discriminatively inspired distortion measure is investigated for scalar quantization to better maintain recognition accuracy. Experiments on an isolated word recognition show that each system variable can be scalar quantized to less than 8 bits using a standard quantization method, except for the alpha probability in Viterbi search which requires 10 bits. However, using our normalization and discriminative distortion measure, the, forward probability can be quantized to 9 bits, thereby halving the corresponding lookup table size. This greatly reduces the memory bandwidth and enables the implementation of custom arithmetic on ASR systems.

With the skyrocketing popularity of mobile devices, new processing methods, tailored for low-resource systems, have become necessary. We propose the use of custom arithmetic logic tailored to a specific application. In a system with all parameters quantized to low precision, such arithmetic can be implemented through a set of small, fast table lookups. We present here a framework for the design of such a system architecture, and several heuristic algorithms to optimize system performance. In addition, we apply our techniques to an automatic speech recognition (ASR) application. Our simulations on various architectures show that on most modern processor designs, we can expect a cycle-count speedup of at least 3 times while requiring a total of only 59 kB of ROMs to hold the lookup tables.

Many pitch trackers based on dynamic programming require meticulous design of local cost and transition cost functions. The forms of these functions are often empirically determined and their parameters are tuned accordingly. Parameter tuning usually requires great effort without a guarantee of optimal performance. This work presents a graphical model framework to automatically optimize pitch tracking parameters in the maximum likelihood sense. Therein, probabilistic dependencies between pitch, pitch transition and acoustical observations are expressed using the language of graphical models, and probabilistic inference is accomplished using the Graphical Model Toolkit (GMTK). Experiments show that this framework not only expedites the design of a pitch tracker, but also yields remarkably good performance for both pitch estimation and voicing decision.

It is important to produce automatic speech recognition (ASR) systems that use as few computational and memory resources as possible, especially in low-memory/low-power environments such as for personal digital assistants. One way to achieve this is through parameter quantization. In this work, we compare a variety of novel subvector clustering procedures for ASR system parameter quantization. Specifically, we look at systematic data-driven subvector selection techniques based on entropy minimization, and compare performance on a 150-word isolated word speech recognition task. While the optimal entropy-minimizing quantization methods are intractable, we show that although several of our heuristic techniques are elaborate in their attempt to approximate the optimal clustering, a simple scalar quantization scheme using separate codebooks performs remarkably well.