Variable Attention Masking for Configurable Transformer Transducer Speech Recognition

This work studies the use of attention masking in transformer transducer based speech recognition for building a single configurable model for different deployment scenarios. We present a comprehensive set of experiments comparing fixed masking, where the same attention mask is applied at every frame, with chunked masking, where the attention mask for each frame is determined by chunk boundaries, in terms of recognition accuracy and latency. We then explore the use of variable masking, where the attention masks are sampled from a target distribution at training time, to build models that can work in different configurations. Finally, we investigate how a single configurable model can be used to perform both first pass streaming recognition and second pass acoustic rescoring. Experiments show that chunked masking achieves a better accuracy vs latency trade-off compared to fixed masking, both with and without FastEmit. We also show that variable masking improves the accuracy by up to 8% relative in the acoustic re-scoring scenario.

Online Automatic Speech Recognition with Listen, Attend and Spell Model

The Listen, Attend and Spell (LAS) model and other attention-based automatic speech recognition (ASR) models have known limitations when operated in a fully online mode. In this paper, we analyze the online operation of LAS models to demonstrate that these limitations stem from the handling of silence regions and the reliability of online attention mechanism at the edge of input buffers. We propose a novel and simple technique that can achieve fully online recognition while meeting accuracy and latency targets. For the Mandarin dictation task, our proposed approach can achieve a character error rate in online operation that is within 4% relative to an offline LAS model. The proposed online LAS model operates at 12% lower latency relative to a conventional neural network hidden Markov model hybrid of comparable accuracy. We have validated the proposed method through a production scale deployment, which, to the best of our knowledge, is the first such deployment of a fully online LAS model.

Multimodal Representation Learning using Deep Multiset Canonical Correlation

We propose Deep Multiset Canonical Correlation Analysis (dMCCA) as an extension to representation learning using CCA when the underlying signal is observed across multiple (more than two) modalities. We use deep learning framework to learn non-linear transformations from different modalities to a shared subspace such that the representations maximize the ratio of between- and within-modality covariance of the observations. Unlike linear discriminant analysis, we do not need class information to learn these representations, and we show that this model can be trained for complex data using mini-batches. Using synthetic data experiments, we show that dMCCA can effectively recover the common signal across the different modalities corrupted by multiplicative and additive noise. We also analyze the sensitivity of our model to recover the correlated components with respect to mini-batch size and dimension of the embeddings. Performance evaluation on noisy handwritten datasets shows that our model outperforms other CCA-based approaches and is comparable to deep neural network models trained end-to-end on this dataset.