Stefan Braun

Pawel Swietojanski, Stefan Braun, Dogan Can et al.

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.

Maurits Bleeker, Pawel Swietojanski, Stefan Braun et al.

This paper presents an extension to train end-to-end Context-Aware Transformer Transducer ( CATT ) models by using a simple, yet efficient method of mining hard negative phrases from the latent space of the context encoder. During training, given a reference query, we mine a number of similar phrases using approximate nearest neighbour search. These sampled phrases are then used as negative examples in the context list alongside random and ground truth contextual information. By including approximate nearest neighbour phrases (ANN-P) in the context list, we encourage the learned representation to disambiguate between similar, but not identical, biasing phrases. This improves biasing accuracy when there are several similar phrases in the biasing inventory. We carry out experiments in a large-scale data regime obtaining up to 7% relative word error rate reductions for the contextual portion of test data. We also extend and evaluate CATT approach in streaming applications.

Thien Nguyen, Nathalie Tran, Liuhui Deng et al.

Code-switching describes the practice of using more than one language in the same sentence. In this study, we investigate how to optimize a neural transducer based bilingual automatic speech recognition (ASR) model for code-switching speech. Focusing on the scenario where the ASR model is trained without supervised code-switching data, we found that semi-supervised training and synthetic code-switched data can improve the bilingual ASR system on code-switching speech. We analyze how each of the neural transducer's encoders contributes towards code-switching performance by measuring encoder-specific recall values, and evaluate our English/Mandarin system on the ASCEND data set. Our final system achieves 25% mixed error rate (MER) on the ASCEND English/Mandarin code-switching test set -- reducing the MER by 2.1% absolute compared to the previous literature -- while maintaining good accuracy on the monolingual test sets.

Joscha Grüger, Tobias Geyer, Martin Kuhn et al.

Conformance checking is a process mining technique that allows verifying the conformance of process instances to a given model. Thus, this technique is predestined to be used in the medical context for the comparison of treatment cases with clinical guidelines. However, medical processes are highly variable, highly dynamic, and complex. This makes the use of imperative conformance checking approaches in the medical domain difficult. Studies show that declarative approaches can better address these characteristics. However, none of the approaches has yet gained practical acceptance. Another challenge are alignments, which usually do not add any value from a medical point of view. For this reason, we investigate in a case study the usability of the HL7 standard Arden Syntax for declarative, rule-based conformance checking and the use of manually modeled alignments. Using the approach, it was possible to check the conformance of treatment cases and create medically meaningful alignments for large parts of a medical guideline.

Stefan Braun, Erik McDermott, Roger Hsiao

The neural transducer is an end-to-end model for automatic speech recognition (ASR). While the model is well-suited for streaming ASR, the training process remains challenging. During training, the memory requirements may quickly exceed the capacity of state-of-the-art GPUs, limiting batch size and sequence lengths. In this work, we analyze the time and space complexity of a typical transducer training setup. We propose a memory-efficient training method that computes the transducer loss and gradients sample by sample. We present optimizations to increase the efficiency and parallelism of the sample-wise method. In a set of thorough benchmarks, we show that our sample-wise method significantly reduces memory usage, and performs at competitive speed when compared to the default batched computation. As a highlight, we manage to compute the transducer loss and gradients for a batch size of 1024, and audio length of 40 seconds, using only 6 GB of memory.

Stefan Braun

This study provides benchmarks for different implementations of LSTM units between the deep learning frameworks PyTorch, TensorFlow, Lasagne and Keras. The comparison includes cuDNN LSTMs, fused LSTM variants and less optimized, but more flexible LSTM implementations. The benchmarks reflect two typical scenarios for automatic speech recognition, notably continuous speech recognition and isolated digit recognition. These scenarios cover input sequences of fixed and variable length as well as the loss functions CTC and cross entropy. Additionally, a comparison between four different PyTorch versions is included. The code is available online https://github.com/stefbraun/rnn_benchmarks.

Hyung Gun Chi, Florian Pesce, Wonil Chang et al.

Device-directed speech detection (DDSD) is a binary classification task that separates the user's queries to a voice assistant (VA) from background speech or side conversations. This is important for achieving naturalistic user experience. To this end, we propose knowledge distillation (KD) to enhance DDSD accuracy while ensuring efficient deployment. Specifically, we introduce a novel adaptive KD method that transfers knowledge from general representations of an ASR large pre-trained acoustic encoder (teacher). We apply task-specific adapters, on top of the (frozen) teacher encoder, trained jointly with the student model on DDSD. We demonstrate that the proposed adaptive KD outperforms the student model without distillation in the keyword and keyword-free (follow-up) invocations, with an improvement of +26% and +19% in terms of Equal Error Rate, respectively. We also show that this approach generalizes across the transformer and conformer-based model architectures.

Jan Pennekamp, Erik Buchholz, Markus Dahlmanns et al.

Following the recent Internet of Things-induced trends on digitization in general, industrial applications will further evolve as well. With a focus on the domains of manufacturing and production, the Internet of Production pursues the vision of a digitized, globally interconnected, yet secure environment by establishing a distributed knowledge base. Background. As part of our collaborative research of advancing the scope of industrial applications through cybersecurity and privacy, we identified a set of common challenges and pitfalls that surface in such applied interdisciplinary collaborations. Aim. Our goal with this paper is to support researchers in the emerging field of cybersecurity in industrial settings by formalizing our experiences as reference for other research efforts, in industry and academia alike. Method. Based on our experience, we derived a process cycle of performing such interdisciplinary research, from the initial idea to the eventual dissemination and paper writing. This presented methodology strives to successfully bootstrap further research and to encourage further work in this emerging area. Results. Apart from our newly proposed process cycle, we report on our experiences and conduct a case study applying this methodology, raising awareness for challenges in cybersecurity research for industrial applications. We further detail the interplay between our process cycle and the data lifecycle in applied research data management. Finally, we augment our discussion with an industrial as well as an academic view on this research area and highlight that both areas still have to overcome significant challenges to sustainably and securely advance industrial applications. Conclusions. With our proposed process cycle for interdisciplinary research in the intersection of cybersecurity and industrial application, we provide a foundation for further research.

Ting-Yao Hu, Ashish Shrivastava, Jen-Hao Rick Chang et al.

Data augmentation methods usually apply the same augmentation (or a mix of them) to all the training samples. For example, to perturb data with noise, the noise is sampled from a Normal distribution with a fixed standard deviation, for all samples. We hypothesize that a hard sample with high training loss already provides strong training signal to update the model parameters and should be perturbed with mild or no augmentation. Perturbing a hard sample with a strong augmentation may also make it too hard to learn from. Furthermore, a sample with low training loss should be perturbed by a stronger augmentation to provide more robustness to a variety of conditions. To formalize these intuitions, we propose a novel method to learn a Sample-Adaptive Policy for Augmentation -- SapAugment. Our policy adapts the augmentation parameters based on the training loss of the data samples. In the example of Gaussian noise, a hard sample will be perturbed with a low variance noise and an easy sample with a high variance noise. Furthermore, the proposed method combines multiple augmentation methods into a methodical policy learning framework and obviates hand-crafting augmentation parameters by trial-and-error. We apply our method on an automatic speech recognition (ASR) task, and combine existing and novel augmentations using the proposed framework. We show substantial improvement, up to 21% relative reduction in word error rate on LibriSpeech dataset, over the state-of-the-art speech augmentation method.

Stefan Braun, Daniel Neil, Enea Ceolini et al.

Recent work on encoder-decoder models for sequence-to-sequence mapping has shown that integrating both temporal and spatial attention mechanisms into neural networks increases the performance of the system substantially. In this work, we report on the application of an attentional signal not on temporal and spatial regions of the input, but instead as a method of switching among inputs themselves. We evaluate the particular role of attentional switching in the presence of dynamic noise in the sensors, and demonstrate how the attentional signal responds dynamically to changing noise levels in the environment to achieve increased performance on both audio and visual tasks in three commonly-used datasets: TIDIGITS, Wall Street Journal, and GRID. Moreover, the proposed sensor transformation network architecture naturally introduces a number of advantages that merit exploration, including ease of adding new sensors to existing architectures, attentional interpretability, and increased robustness in a variety of noisy environments not seen during training. Finally, we demonstrate that the sensor selection attention mechanism of a model trained only on the small TIDIGITS dataset can be transferred directly to a pre-existing larger network trained on the Wall Street Journal dataset, maintaining functionality of switching between sensors to yield a dramatic reduction of error in the presence of noise.

Stefan Braun, Daniel Neil, Shih-Chii Liu

The performance of automatic speech recognition systems under noisy environments still leaves room for improvement. Speech enhancement or feature enhancement techniques for increasing noise robustness of these systems usually add components to the recognition system that need careful optimization. In this work, we propose the use of a relatively simple curriculum training strategy called accordion annealing (ACCAN). It uses a multi-stage training schedule where samples at signal-to-noise ratio (SNR) values as low as 0dB are first added and samples at increasing higher SNR values are gradually added up to an SNR value of 50dB. We also use a method called per-epoch noise mixing (PEM) that generates noisy training samples online during training and thus enables dynamically changing the SNR of our training data. Both the ACCAN and the PEM methods are evaluated on a end-to-end speech recognition pipeline on the Wall Street Journal corpus. ACCAN decreases the average word error rate (WER) on the 20dB to -10dB SNR range by up to 31.4% when compared to a conventional multi-condition training method.