This study presents a novel approach for knowledge distillation (KD) from a BERT teacher model to an automatic speech recognition (ASR) model using intermediate layers. To distil the teacher’s knowledge, we use an attention decoder that learns from BERT’s token probabilities. Our method shows that language model (LM) information can be more effectively distilled into an ASR model using both the intermediate layers and the final layer. By using the intermediate layers as distillation target, we can more effectively distil LM knowledge into the lower network layers. Using our method, we achieve better recognition accuracy than with shallow fusion of an external LM, allowing us to maintain fast parallel decoding. Experiments on the LibriSpeech dataset demonstrate the effectiveness of our approach in enhancing greedy decoding with connectionist temporal classification (CTC).

Speech samples recorded in both indoor and outdoor environments are often contaminated with secondary audio sources. Most end-to-end monaural speech recognition systems either remove these background sounds using speech enhancement or train noise-robust models. For better model interpretability and holistic understanding, we aim to bring together the growing field of automated audio captioning (AAC) and the thoroughly studied automatic speech recognition (ASR). The goal of AAC is to generate natural language descriptions of contents in audio samples. We propose several approaches for end-to-end joint modeling of ASR and AAC tasks and demonstrate their advantages over traditional approaches, which model these tasks independently. A major hurdle in evaluating our proposed approach is the lack of labeled audio datasets with both speech transcriptions and audio captions. Therefore we also create a multi-task dataset by mixing the clean speech Wall Street Journal corpus with multiple levels of background noises chosen from the AudioCaps dataset. We also perform extensive experimental evaluation and show improvements of our proposed methods as compared to existing state-of-the-art ASR and AAC methods.

A streaming style inference of encoder–decoder automatic speech recognition (ASR) systems is important for reducing latency, which is essential for interactive use cases. To this end, we propose a novel blockwise synchronous decoding algorithm with a hybrid approach that combines endpoint prediction and endpoint post-determination. In the endpoint prediction, we compute the expectation of the number of tokens that are yet to be emitted in the encoder features of the current blocks using the CTC posterior. Based on the expectation value, the decoder predicts the endpoint to realize continuous block synchronization, as a running stitch. Meanwhile, end-point post-determination probabilistically detects backward jump of the source–target attention, which is caused by the misprediction of endpoints. Then it resumes decoding by discarding those hypotheses, as back stitch. We combine these methods into a hybrid approach, namely run-and-back stitch search, which reduces the computational cost and latency. Evaluations of various ASR tasks show the efficiency of our proposed decoding algorithm, which achieves a latency reduction, for instance in the Librispeech test set from 1487 ms to 821 ms at the 90th percentile, while maintaining a high recognition accuracy.

In punctuation restoration, we try to recover the missing punctuation from automatic speech recognition output to improve understandability. Currently, large pre-trained transformers such as BERT set the benchmark on this task but there are two main drawbacks to these models. First, the pre-training data does not match the output data from speech recognition that contains errors. Second, the large number of model parameters increases inference time. To address the former, we use a multi-task learning framework with ELECTRA, a recently proposed improvement on BERT, that has a generator-discriminator structure. The generator allows us to inject errors into the training data and, as our experiments show, this improves robustness against speech recognition errors during inference. To address the latter, we investigate knowledge distillation and parameter pruning of ELECTRA. In our experiments on the IWSLT 2012 benchmark data, a model with less than 11% the size of BERT achieved better performance while having an 82% faster inference time.

An important task for language models is the adaptation of general-domain models to specific target domains. For neural network-based language models, feature-based domain adaptation has been a popular method in previous research. Conventional methods use an adaptation feature providing context information that is calculated from a topic model. However, such a topic model needs to be trained separately from the language model. To unify the language and context model training, we present an approach that combines an extractor network and a domain adaptation layer. The extractor network learns a context representation from a fixed-size window of past words and provides the context information for the adaptation layer. The benefit of our method is that the extractor network can be trained jointly with the language model in a single training step. Our proposed method showed superior performance over conventional domain adaptation with topic features on a dataset of TED talks with respect to perplexity and word error rate after 100-best rescoring.