No abstract

We present an algorithmic framework for supervised classification learning where the set of labels is organized in a predefined hierarchical structure. This structure is encoded by a rooted tree which induces a metric over the label set. Our approach combines ideas from large margin kernel methods and Bayesian analysis. Following the large margin principle, we associate a prototype with each label in the tree and formulate the learning task as an optimization problem with varying margin constraints. In the spirit of Bayesian methods, we impose similarity requirements between the prototypes corresponding to adjacent labels in the hierarchy. We describe new online and batch algorithms for solving the constrained optimization problem. We derive a worst case loss-bound for the online algorithm and provide generalization analysis for its batch counterpart. We demonstrate the merits of our approach with a series of experiments on synthetic, text and speech data.

Generating adversarial examples is a critical step for evaluating and improving the robustness of learning machines. So far, most existing methods only work for classification and are not designed to alter the true performance measure of the problem at hand. We introduce a novel flexible approach named Houdini for generating adversarial examples specifically tailored for the final performance measure of the task considered, be it combinatorial and non-decomposable. We successfully apply Houdini to a range of applications such as speech recognition, pose estimation and semantic segmentation. In all cases, the attacks based on Houdini achieve higher success rate than those based on the traditional surrogates used to train the models while using a less perceptible adversarial perturbation.

This chapter introduces a discriminative method for detecting and spotting keywords in spoken utterances. Given a word represented as a sequence of phonemes and a spoken utterance, the keyword spotter predicts the best time span of the phoneme sequence in the spoken utterance along with a confidence. If the prediction confidence is above certain level the keyword is declared to be spoken in the utterance within the predicted time span, otherwise the keyword is declared as not spoken. The problem of keyword spotting training is formulated as a discriminative task where the model parameters are chosen so the utterance in which the keyword is spoken would have higher confidence than any other spoken utterance in which the keyword is not spoken. It is shown theoretically and empirically that the proposed training method resulted with a high area under the receiver operating characteristic (ROC) curve, the most common measure to evaluate keyword spotters. We present an iterative algorithm to train the keyword spotter efficiently. The proposed approach contrasts with standard spotting strategies based on HMMs, for which the training procedure does not maximize a loss directly related to the spotting performance. Several experiments performed on TIMIT and WSJ corpora show the advantage of our approach over HMM-based alternatives.

No abstract

No abstract

A discriminative large-margin algorithm for automatic measurement of voice onset time (VOT) is described, considered as a case of predicting structured output from speech. Manually labeled data are used to train a function that takes as input a speech segment of an arbitrary length containing a voiceless stop, and outputs its VOT. The function is explicitly trained to minimize the difference between predicted and manually measured VOT; it operates on a set of acoustic feature functions designed based on spectral and temporal cues used by human VOT annotators. The algorithm is applied to initial voiceless stops from four corpora, representing different types of speech. Using several evaluation methods, the algorithm's performance is near human intertranscriber reliability, and compares favorably with previous work. Furthermore, the algorithm's performance is minimally affected by training and testing on different corpora, and remains essentially constant as the amount of training data is reduced to 50-250 manually labeled examples, demonstrating the method's practical applicability to new datasets.

Vowel durations are most often utilized in studies addressing specific issues in phonetics. Thus far this has been hampered by a reliance on subjective, labor-intensive manual annotation. Our goal is to build an algorithm for automatic accurate measurement of vowel duration, where the input to the algorithm is a speech segment contains one vowel preceded and followed by consonants (CVC). Our algorithm is based on a deep neural network trained at the frame level on manually annotated data from a phonetic study. Specifically, we try two deep-network architectures: convolutional neural network (CNN), and deep belief network (DBN), and compare their accuracy to an HMM-based forced aligner. Results suggest that CNN is better than DBN, and both CNN and HMM-based forced aligner are comparable in their results, but neither of them yielded the same predictions as models fit to manually annotated data.