Current state-of-the-art automatic speech recognition systems are trained to work in specific `domains', defined based on factors like application, sampling rate and codec. When such recognizers are used in conditions that do not match the training domain, performance significantly drops. This work explores the idea of building a single domain-invariant model for varied use-cases by combining large scale training data from multiple application domains. Our final system is trained using 162,000 hours of speech. Additionally, each utterance is artificially distorted during training to simulate effects like background noise, codec distortion, and sampling rates. Our results show that, even at such a scale, a model thus trained works almost as well as those fine-tuned to specific subsets: A single model can be robust to multiple application domains, and variations like codecs and noise. More importantly, such models generalize better to unseen conditions and allow for rapid adaptation - we show that by using as little as 10 hours of data from a new domain, an adapted domain-invariant model can match performance of a domain-specific model trained from scratch using 70 times as much data. We also highlight some of the limitations of such models and areas that need addressing in future work.

While research has often shown that building dialect-specific Automatic Speech Recognizers is the optimal approach to dealing with dialectal variations of the same language, we have observed that dialect-specific recognizers do not always output the best recognitions. Often enough, another dialectal recognizer outputs a better recognition than the dialect-specific one. In this paper, we present two methods to select and combine the best decoded hypothesis from a pool of dialectal recognizers. We follow a Machine Learning approach and extract features from the Speech Recognition output along with Word Embeddings and use Shallow Neural Networks for classification. Our experiments using Dictation and Voice Search data from the main four Arabic dialects show good WER improvements for the hypothesis selection scheme, reducing the WER by 2.1 to 12.1% depending on the test set, and promising results for the hypotheses combination scheme.

Acoustic model performance typically decreases when evaluated on a dialectal variation of the same language that was not used during training. Similarly, models simultaneously trained on a group of dialects tend to underperform dialect-specific models. In this paper, we report on our efforts towards building a unified acoustic model that can serve a multi-dialectal language. Two techniques are presented: Distillation and MultiTask Learning (MTL). In Distillation, we use an ensemble of dialect-specific acoustic models and distill its knowledge in a single model. In MTL, we utilize multitask learning to train a unified acoustic model that learns to distinguish dialects as a side task. We show that both techniques are superior to the jointly-trained model that is trained on all dialectal data, reducing word error rates by 4:2% and 0:6%, respectively. While achieving this improvement, neither technique degrades the performance of the dialect-specific models by more than 3:4%.

Sensor devices are becoming ubiquitous, especially in measurement and monitoring applications. Because of the real-time, append-only and semi-infinite natures of the generated sensor data streams, an online incremental approach is a necessity for mining stream data types. In this paper, we propose STAGGER: a one-pass, online and incremental algorithm for mining periodic patterns in data streams. STAGGER does not require that the user pre-specify the periodicity rate of the data. Instead, STAGGER discovers the potential periodicity rates. STAGGER maintains multiple expanding sliding windows staggered over the stream, where computations are shared among the multiple overlapping windows. Small-length sliding windows are imperative for early and real-time output, yet are limited to discover short periodicity rates. As streamed data arrives continuously, the sliding windows expand in length in order to cover the whole stream. Larger-length sliding windows are able to discover longer periodicity rates. STAGGER incrementally maintains a tree-like data structure for the frequent periodic patterns of each discovered potential periodicity rate. In contrast to the Fourier/Wavelet-based approaches used for discovering periodicity rates, STAGGER not only discovers a wider, more accurate set of periodicities, but also discovers the periodic patterns themselves. In fact, experimental results with real and synthetic data sets show that STAGGER outperforms Fourier/Wavelet-based approaches by an order of magnitude in terms of the accuracy of the discovered periodicity rates. Moreover, real-data experiments demonstrate the practicality of the discovered periodic patterns.

Periodicity mining is used for predicting trends in time series data. Periodicity detection is an essential process in periodicity mining to discover potential periodicity rates. Existing periodicity detection algorithms do not take into account the presence of noise, which is inevitable in almost every real-world time series data. In this paper, we tackle the problem of periodicity detection in the presence of noise. We propose a new periodicity detection algorithm that deals efficiently with all types of noise. Based on time warping, the proposed algorithm warps (extends or shrinks) the time axis at various locations to optimally remove the noise. Experimental results show that the proposed algorithm outperforms the existing periodicity detection algorithms in terms of noise resiliency.

Periodicity mining is used for predicting trends in time series data. Discovering the rate at which the time series is periodic has always been an obstacle for fully automated periodicity mining. Existing periodicity mining algorithms assume that the periodicity, rate (or simply the period) is user-specified. This assumption is a considerable limitation, especially in time series data where the period is not known a priori. In this paper, we address the problem of detecting the periodicity rate of a time series database. Two types of periodicities are defined, and a scalable, computationally efficient algorithm is proposed for each type. The algorithms perform in O(n log n) time for a time series of length n. Moreover, the proposed algorithms are extended in order to discover the periodic patterns of unknown periods at the same time without affecting the time complexity. Experimental results show that the proposed algorithms are highly accurate with respect to the discovered periodicity rates and periodic patterns. Real-data experiments demonstrate the practicality of the discovered periodic patterns.

We present the demonstration of the design of "STEAM", Purdue Boiler Makers' stream database system that allows for the processing of continuous and snap-shot queries over data streams. Specifically, the demonstration focuses on the query processing engine, "Nile". Nile extends the query processor engine of an object-relational database management system, PREDATOR, to process continuous queries over data streams. Nile supports extended SQL operators that handle sliding-window execution as an approach to restrict the size of the stored state in operators such as join.

Mining of periodic patterns in time-series databases is an interesting data mining problem. It can be envisioned as a tool for forecasting and prediction of the future behavior of time-series data. Incremental mining refers to the issue of maintaining the discovered patterns over time in the presence of more items being added into the database. Because of the mostly append only nature of updating time-series data, incremental mining would be very effective and efficient. Several algorithms for incremental mining of partial periodic patterns in time-series databases are proposed and are analyzed empirically. The new algorithms allow for online adaptation of the thresholds in order to produce interactive mining of partial periodic patterns. The storage overhead of the incremental online mining algorithms is analyzed. Results show that the storage overhead for storing the intermediate data structures pays off as the incremental online mining of partial periodic patterns proves to be significantly more efficient than the nonincremental nononline versions. Moreover, a new problem, termed merge mining, is introduced as a generalization of incremental mining. Merge mining can be defined as merging the discovered patterns of two or more databases that are mined independently of each other. An algorithm for merge mining of partial periodic patterns in time-series databases is proposed and analyzed.

Data cleaning is a vital process that ensures the quality of data stored in real-world databases. Data cleaning problems are frequently encountered in many research areas, such as knowledge discovery in databases, data warehousing, system integration and e-services. The process of identifying the record pairs that represent the same entity (duplicate records), commonly known as record linkage, is one of the essential elements of data cleaning. In this paper, we address the record linkage problem by adopting a machine learning approach. Three models are proposed and are analyzed empirically. Since no existing model, including those proposed in this paper, has been proved to be superior, we have developed an interactive record linkage toolbox named TAILOR (backwards acronym for "RecOrd LInkAge Toolbox"). Users of TAILOR can build their own record linkage models by tuning system parameters and by plugging in in-house-developed and public-domain tools. The proposed toolbox serves as a framework for the record linkage process, and is designed in an extensible way to interface with existing and future record linkage models. We have conducted an extensive experimental study to evaluate our proposed models using not only synthetic but also real data. The results show that the proposed machine-learning record linkage models outperform the existing ones both in accuracy and in performance.