End-to-end automatic speech recognition system implemented in TensorFlow. If you want to replace feed dict operation with Tensorflow multi-thread and fifoqueue input pipeline, you can refer to my repo TensorFlow-Input-Pipeline for more example codes. My own practices prove that fifoqueue input pipeline would improve the training speed in some time. If you want to look the history of speech recognition, I have collected the significant papers since 1981 in the ASR field. I will update it every week to add new papers, including speech recognition, speech synthesis and language modelling.

The ubiquitous presence of sequence data across fields such as the web, healthcare, bioinformatics, and text mining has made sequence mining a vital research area. However, sequence mining is particularly challenging because of difficulty in finding (dis)similarity/distance between sequences. This is because a distance measure between sequences is not obvious due to their unstructuredness---arbitrary strings of arbitrary length. Feature representations, such as n-grams, are often used but they either compromise on extracting both short- and long-term sequence patterns or have a high computation. We propose a new function, Sequence Graph Transform (SGT), that extracts the short- and long-term sequence features and embeds them in a finite-dimensional feature space. Importantly, SGT has low computation and can extract any amount of short- to long-term patterns without any increase in the computation, also proved theoretically in this paper. Due to this, SGT yields superior result with significantly higher accuracy and lower computation compared to the existing methods. We show it via several experimentation and SGT's real world application for clustering, classification, search and visualization as examples.

Estimation, recognition, and near-future prediction of 3D trajectories based on their two dimensional projections available from one camera source is an exceptionally difficult problem due to uncertainty in the trajectories and environment, high dimensionality of the specific trajectory states, lack of enough labeled data and so on. In this article, we propose a solution to solve this problem based on a novel deep learning model dubbed Disjunctive Factored Four-Way Conditional Restricted Boltzmann Machine (DFFW-CRBM). Our method improves state-of-the-art deep learning techniques for high dimensional time-series modeling by introducing a novel tensor factorization capable of driving forth order Boltzmann machines to considerably lower energy levels, at no computational costs. DFFW-CRBMs are capable of accurately estimating, recognizing, and performing near-future prediction of three-dimensional trajectories from their 2D projections while requiring limited amount of labeled data. We evaluate our method on both simulated and real-world data, showing its effectiveness in predicting and classifying complex ball trajectories and human activities.

Teaching machines to accomplish tasks by conversing naturally with humans is challenging. Currently, developing task-oriented dialogue systems requires creating multiple components and typically this involves either a large amount of handcrafting, or acquiring costly labelled datasets to solve a statistical learning problem for each component. In this work we introduce a neural network-based text-in, text-out end-to-end trainable goal-oriented dialogue system along with a new way of collecting dialogue data based on a novel pipe-lined Wizard-of-Oz framework. This approach allows us to develop dialogue systems easily and without making too many assumptions about the task at hand. The results show that the model can converse with human subjects naturally whilst helping them to accomplish tasks in a restaurant search domain.

In recent years there is a growing interest in using deep representations for reinforcement learning. In this paper, we present a methodology and tools to analyze Deep Q-networks (DQNs) in a non-blind matter. Moreover, we propose a new model, the Semi Aggregated Markov Decision Process (SAMDP), and an algorithm that learns it automatically. The SAMDP model allows us to identify spatio-temporal abstractions directly from features and may be used as a sub-goal detector in future work. Using our tools we reveal that the features learned by DQNs aggregate the state space in a hierarchical fashion, explaining its success. Moreover, we are able to understand and describe the policies learned by DQNs for three different Atari2600 games and suggest ways to interpret, debug and optimize deep neural networks in reinforcement learning.

One of the core components of modern spoken dialogue systems is the belief tracker, which estimates the user's goal at every step of the dialogue. However, most current approaches have difficulty scaling to larger, more complex dialogue domains. This is due to their dependency on either: a) Spoken Language Understanding models that require large amounts of annotated training data; or b) hand-crafted lexicons for capturing some of the linguistic variation in users' language. We propose a novel Neural Belief Tracking (NBT) framework which overcomes these problems by building on recent advances in representation learning. NBT models reason over pre-trained word vectors, learning to compose them into distributed representations of user utterances and dialogue context. Our evaluation on two datasets shows that this approach surpasses past limitations, matching the performance of state-of-the-art models which rely on hand-crafted semantic lexicons and outperforming them when such lexicons are not provided.

Speech detection (SD) greatly improves the separation of speech sources from background interferes [1]. Nowadays, SD techniques attract intense attentions in a general speech processing framework, including automatic speech recognition (ASR) [2], speech enhancement [3] and speech coding [1]. Recently, deep neural network (DNN) based 1D SD algorithms show great advantages over conventional voice activity detectors [4], [5]. The obvious benefits of such approaches lie on their easy integration into ASR, robust performance, and feature fusion capability. Zhang and Wu [4] introduced deep belief network and used stacked Bernoulli-Bernoulli restricted Boltzmann machines (RBMs) to conduct the 1D SD. The idea that incorporating temporal context correlation to strengthen the dynamical detection is widely used in network structure design [6], [7]. Other DNN based 1D SD strategies might either focus on improving the front-end acoustic feature inputs (e.g., acoustic models and statistical models) [8], [9], or exploiting the supervised network structure in terms of sample training [10]. These DNN based approaches rely on comprehensive network training, and then are applied to binarily label the speech activities in the time domain. However, 1D SD methods integrate frequency features, and cannot reveal information in the joint time-frequency domain, which are generally more expressive on speech activities, compared with the binary values in 1D SD approaches.

Users form information trails as they browse the web, checkin with a geolocation, rate items, or consume media. A common problem is to predict what a user might do next for the purposes of guidance, recommendation, or prefetching. First-order and higher-order Markov chains have been widely used methods to study such sequences of data. First-order Markov chains are easy to estimate, but lack accuracy when history matters. Higher-order Markov chains, in contrast, have too many parameters and suffer from overfitting the training data. Fitting these parameters with regularization and smoothing only offers mild improvements. In this paper we propose the retrospective higher-order Markov process (RHOMP) as a low-parameter model for such sequences. This model is a special case of a higher-order Markov chain where the transitions depend retrospectively on a single history state instead of an arbitrary combination of history states. There are two immediate computational advantages: the number of parameters is linear in the order of the Markov chain and the model can be fit to large state spaces. Furthermore, by providing a specific structure to the higher-order chain, RHOMPs improve the model accuracy by efficiently utilizing history states without risks of overfitting the data. We demonstrate how to estimate a RHOMP from data and we demonstrate the effectiveness of our method on various real application datasets spanning geolocation data, review sequences, and business locations. The RHOMP model uniformly outperforms higher-order Markov chains, Kneser-Ney regularization, and tensor factorizations in terms of prediction accuracy.

A growing number of threats to Android phones creates challenges for malware detection. Manually labeling the samples into benign or different malicious families requires tremendous human efforts, while it is comparably easy and cheap to obtain a large amount of unlabeled APKs from various sources. Moreover, the fast-paced evolution of Android malware continuously generates derivative malware families. These families often contain new signatures, which can escape detection when using static analysis. These practical challenges can also cause traditional supervised machine learning algorithms to degrade in performance. In this paper, we propose a framework that uses model-based semi-supervised (MBSS) classification scheme on the dynamic Android API call logs. The semi-supervised approach efficiently uses the labeled and unlabeled APKs to estimate a finite mixture model of Gaussian distributions via conditional expectation-maximization and efficiently detects malwares during out-of-sample testing. We compare MBSS with the popular malware detection classifiers such as support vector machine (SVM), $k$-nearest neighbor (kNN) and linear discriminant analysis (LDA). Under the ideal classification setting, MBSS has competitive performance with 98\% accuracy and very low false positive rate for in-sample classification. For out-of-sample testing, the out-of-sample test data exhibit similar behavior of retrieving phone information and sending to the network, compared with in-sample training set. When this similarity is strong, MBSS and SVM with linear kernel maintain 90\% detection rate while $k$NN and LDA suffer great performance degradation. When this similarity is slightly weaker, all classifiers degrade in performance, but MBSS still performs significantly better than other classifiers.

Information systems experience an ever-growing volume of unstructured data, particularly in the form of textual materials. This represents a rich source of information from which one can create value for people, organizations and businesses. For instance, recommender systems can benefit from automatically understanding preferences based on user reviews or social media. However, it is difficult for computer programs to correctly infer meaning from narrative content. One major challenge is negations that invert the interpretation of words and sentences. As a remedy, this paper proposes a novel learning strategy to detect negations: we apply reinforcement learning to find a policy that replicates the human perception of negations based on an exogenous response, such as a user rating for reviews. Our method yields several benefits, as it eliminates the former need for expensive and subjective manual labeling in an intermediate stage. Moreover, the inferred policy can be used to derive statistical inferences and implications regarding how humans process and act on negations.