We propose improved Deep Neural Network (DNN) training loss functions for more accurate single keyword spotting on resource-constrained embedded devices. The loss function modifications consist of a combination of multi-task training and weighted cross entropy. In the multi-task architecture, the keyword DNN acoustic model is trained with two tasks in parallel-the main task of predicting the keyword-specific phone states, and an auxiliary task of predicting LVCSR senones. We show that multi-task learning leads to comparable accuracy over a previously proposed transfer learning approach where the keyword DNN training is initialized by an LVCSR DNN of the same input and hidden layer sizes. The combination of LVCSRinitialization and Multi-task training gives improved keyword detection accuracy compared to either technique alone. We also propose modifying the loss function to give a higher weight on input frames corresponding to keyword phone targets, with a motivation to balance the keyword and background training data. We show that weighted cross-entropy results in additional accuracy improvements. Finally, we show that the combination of 3 techniques-LVCSR-initialization, multi-task training and weighted cross-entropy gives the best results, with significantly lower False Alarm Rate than the LVCSR-initialization technique alone, across a wide range of Miss Rates.

In this paper we investigate a time delay neural network (TDNN) for a keyword spotting task that requires low CPU, memory and latency. The TDNN is trained with transfer learning and multi-task learning. Temporal subsampling enabled by the time delay architecture reduces computational complexity. We propose to apply singular value decomposition (SVD) to further reduce TDNN complexity. This allows us to first train a larger full-rank TDNN model which is not limited by-CPU/memory constraints. The larger TDNN usually achieves better performance. Afterwards, its size can be compressed by SVD to meet the budget requirements. Hidden Markov models (HMM) are used in conjunction with the networks to perform keyword detection and performance is measured in terms of area under the curve (AUC) for detection error tradeoff (DET) curves. Our experimental results on a large in-house far-field corpus show that the full-rank TDNN achieves a 19.7% DET AUC reduction compared to a similar-size deep neural network (DNN) baseline. If we train a larger size full-rank TDNN first and then reduce it via SVD to the comparable size of the DNN, we obtain a 37.6% reduction in DET AUC compared to the DNN baseline.

We propose a max-pooling based loss function for training Long Short-Term Memory (LSTM) networks for smallfootprint keyword spotting (KWS), with low CPU, memory, and latency requirements. The max-pooling loss training can be further guided by initializing with a cross-entropy loss trained network. A posterior smoothing based evaluation approach is employed to measure keyword spotting performance. Our experimental results show that LSTM models trained using cross-entropy loss or max-pooling loss outperform a cross-entropy loss trained baseline feed-forward Deep Neural Network (DNN). In addition, max-pooling loss trained LSTM with randomly initialized network performs better compared to cross-entropy loss trained LSTM. Finally, the max-pooling loss trained LSTM initialized with a crossentropy pre-trained network shows the best performance, which yields 67.6% relative reduction compared to baseline feed-forward DNN in Area Under the Curve (AUC) measure.

This paper presents our recent effort that aims at improving our Arabic Broadcast News (BN) recognition system by using thousands of hours of un-transcribed Arabic audio in the way of unsupervised training. Unsupervised training is first carried out on the 1,900-hour English Topic Detection and Tracking (TDT) data and is compared with the lightly-supervised training method that we have used for the DARPA EARS evaluations. The comparison shows that unsupervised training produces a 21.7% relative reduction in word error rate (WER), which is comparable to the gain obtained with light supervision methods. The same unsupervised training strategy carried out on a similar amount of Arabic BN data produces an 11.6% relative gain. The gain, though considerable, is substantially smaller than what is observed on the English data. Our initial work towards understanding the reasons for this difference is also described.

Current statistical machine translation (SMT) systems are trained on sentencealigned and word-aligned parallel text collected from various sources. Translation model parameters are estimated from the word alignments, and the quality of the translations on a given test set depends on the parameter estimates. There are at least two factors affecting the parameter estimation: domain match and training data quality. This paper describes a novel approach for automatically detecting and down-weighing certain parts of the training corpus by assigning a weight to each sentence in the training bitext so as to optimize a discriminative objective function on a designated tuning set. This way, the proposed method can limit the negative effects of low quality training data, and can adapt the translation model to the domain of interest. It is shown that such discriminative corpus weights can provide significant improvements in Arabic-English translation on various conditions, using a state-of-the-art SMT system.

Large scale Natural Language Understanding (NLU) systems are typically trained on large quantities of data, requiring a fast and scalable training strategy. A typical design for NLU systems consists of domain-level NLU modules (domain classifier, intent classifier and named entity recognizer). Hypotheses (NLU interpretations consisting of various intent+slot combinations) from these domain specific modules are typically aggregated with another downstream component. The re-ranker integrates outputs from domain-level recognizers, returning a scored list of cross domain hypotheses. An ideal reranker will exhibit the following two properties: (a) it should prefer the most relevant hypothesis for the given input as the top hypothesis and, (b) the interpretation scores corresponding to each hypothesis produced by the re-ranker should be calibrated. Calibration allows the final NLU interpretation score to be comparable across domains. We propose a novel re-ranker strategy that addresses these aspects, while also maintaining domain specific modularity. We design optimization loss functions for such a modularized re-ranker and present results on decreasing the top hypothesis error rate as well as maintaining the model calibration. We also experiment with an extension involving training the domain specific re-rankers on datasets curated independently by each domain to allow further asynchronization.

In this work, we propose a classifier for distinguishing device-directed queries from background speech in the context of interactions with voice assistants. Applications include rejection of false wake-ups or unintended interactions as well as enabling wake-word free followup queries. Consider the example interaction: "Computer, play music", "Computer, reduce the volume". In this interaction, the user needs to repeat the wake-word (Computer) for the second query. To allow for more natural interactions, the device could immediately re-enter listening state after the first query (without wake-word repetition) and accept or reject a potential follow-up as device-directed or background speech. The proposed model consists of two long short-term memory (LSTM) neural networks trained on acoustic features and automatic speech recognition (ASR) 1-best hypotheses, respectively. A feed-forward deep neural network (DNN) is then trained to combine the acoustic and 1-best embeddings, derived from the LSTMs, with features from the ASR decoder. Experimental results show that ASR decoder, acoustic embeddings, and 1-best embeddings yield an equal-error-rate (EER) of 9.3 %, 10.9 % and 20.1 %, respectively. Combination of the features resulted in a 44 % relative improvement and a final EER of 5.2 %.

We study few-shot acoustic event detection (AED) in this paper. Few-shot learning enables detection of new events with very limited labeled data. Compared to other research areas like computer vision, few-shot learning for audio recognition has been under-studied. We formulate few-shot AED problem and explore different ways of utilizing traditional supervised methods for this setting as well as a variety of meta-learning approaches, which are conventionally used to solve few-shot classification problem. Compared to supervised baselines, meta-learning models achieve superior performance, thus showing its effectiveness on generalization to new audio events. Our analysis including impact of initialization and domain discrepancy further validate the advantage of meta-learning approaches in fewshot AED.