We investigate massive end-to-end automatic speech recognition (ASR) models with efficiency improvements achieved by time reduction. The encoders of our models use the neural architecture of Google’s universal speech model (USM), with additional funnel pooling layers to significantly reduce the frame rate and speed up training and inference. We also explore a few practical methods to mitigate potential accuracy loss due to time reduction, while enjoying most efficiency gain. Our methods are demonstrated to work with both Connectionist Temporal Classification (CTC) and RNN-Transducer (RNN-T), with up to 2B model parameters, and over two domains. For a large-scale voice search recognition task, we perform extensive studies on vocabulary size, time reduction strategy, and its generalization performance on long-form test sets, and show that a 900M RNN-T is very tolerant to severe time reduction, with as low encoder output frame rate as 640ms. We also provide ablation studies on the Librispeech benchmark for important training hyperparameters and architecture designs, in training 600M RNN-T models at the frame rate of 160ms.

We employ the method of fine-tuning wav2vec2.0 for recognition of phonemes in aphasic speech. Our effort focuses on data augmentation, by supplementing data from both in-domain and out-of-domain datasets for training. We found that although a modest amount of out-of-domain data may be helpful, the performance of the model degrades significantly when the amount of out-of-domain data is much larger than in-domain data. Our hypothesis is that fine-tuning wav2vec2.0 with a CTC loss not only learns bottom-up acoustic properties but also top-down constraints. Therefore, out-of-domain data augmentation is likely to degrade performance if there is a language model mismatch between “in” and “out” domains. For in-domain audio without ground truth labels, we found that it is beneficial to exclude samples with less confident pseudo labels. Our final model achieves 16.7% PER (phoneme error rate) on the validation set, without using a language model for decoding. The result represents a relative error reduction of 14% over the baseline model trained without data augmentation. Finally, we found that “canonicalized” phonemes are much easier to recognize than manually transcribed phonemes.

We propose using lexical resources (thesaurus, VAD) to fine-tune pretrained deep nets such as BERT and ERNIE. Then at inference time, these nets can be used to distinguish synonyms from antonyms, as well as VAD distances. The inference method can be applied to words as well as texts such as multiword expressions (MWEs), out of vocabulary words (OOVs), morphological variants and more. Code and data are posted on https://github.com/kwchurch/syn_ant.

Multi-layer multi-head self-attention mechanism is widely applied in modern neural language models. Attention redundancy has been observed among attention heads but has not been deeply studied in the literature. Using BERT-base model as an example, this paper provides a comprehensive study on attention redundancy which is helpful for model interpretation and model compression. We analyze the attention redundancy with Five-Ws and How. (What) We define and focus the study on redundancy matrices generated from pre-trained and fine-tuned BERT-base model for GLUE datasets. (How) We use both token-based and sentence-based distance functions to measure the redundancy. (Where) Clear and similar redundancy patterns (cluster structure) are observed among attention heads. (When) Redundancy patterns are similar in both pre-training and fine-tuning phases. (Who) We discover that redundancy patterns are task-agnostic. Similar redundancy patterns even exist for randomly generated token sequences. (“Why”) We also evaluate influences of the pre-training dropout ratios on attention redundancy. Based on the phase-independent and task-agnostic attention redundancy patterns, we propose a simple zero-shot pruning method as a case study. Experiments on fine-tuning GLUE tasks verify its effectiveness. The comprehensive analyses on attention redundancy make model understanding and zero-shot model pruning promising.

This paper designs a Monolingual Lexicon Induction task and observes that two factors accompany the degraded accuracy of bilingual lexicon induction for rare words. First, a diminishing margin between similarities in low frequency regime, and secondly, exacerbated hubness at low frequency. Based on the observation, we further propose two methods to address these two factors, respectively. The larger issue is hubness. Addressing that improves induction accuracy significantly, especially for low-frequency words.