End-to-end spoken language understanding (SLU) systems benefit from pretraining on large corpora, followed by fine-tuning on application-specific data. The resulting models are too large for on-edge applications. For instance, BERT-based systems contain over 110M parameters. Observing the model is overparameterized, we propose lean transformer structure where the dimension of the attention mechanism is automatically reduced using group sparsity. We propose a variant where the learned attention subspace is transferred to an attention bottleneck layer. In a low-resource setting and without pre-training, the resulting compact SLU model achieves accuracies competitive with pre-trained large models.
Learning a set of tasks in sequence remains a challenge for Artificial Neural Networks, which, in such scenarios, tend to suffer from Catastrophic Forgetting (CF). The same applies to End-to-End (E2E) Automatic Speech Recognition (ASR) models, even for monolingual tasks. In this paper, we aim to overcome CF for E2E ASR by inserting adapters, small architectures of few parameters which allow a general model to be fine-tuned to a specific task, into our model. We make these adapters task-specific, while either freezing or regularizing the parameters of the model shared by all tasks, thus stimulating the model to fully exploit the adapter parameters while keeping the shared parameters such that they work well for all tasks. Our best method is able to close the gap between worst case (naively fine-tuning) and best case (jointly training on all tasks) by more than 75% after a set of four monolingual tasks.
Adapting Automatic Speech Recognition (ASR) models to new domains leads to a deterioration of performance on the original domain(s), a phenomenon called Catastrophic Forgetting (CF). Even monolingual ASR models cannot be extended to new accents, dialects, topics, etc. without suffering from CF, making them unable to be continually enhanced without storing all past data. Fortunately, Continual Learning (CL) methods, which aim to enable continual adaptation while overcoming CF, can be used. In this paper, we implement an extensive number of CL methods for End-to-End ASR and test and compare their ability to extend a monolingual Hybrid CTC-Transformer model across four new tasks. We find that the best performing CL method closes the gap between the fine-tuned model (lower bound) and the model trained jointly on all tasks (upper bound) by more than 40%, while requiring access to only 0.6% of the original data.
The choice of an optimal time-frequency resolution is usually a difficult but important step in tasks involving speech signal classification, e.g., speech anti-spoofing. The variations of the performance with different choices of timefrequency resolutions can be as large as those with different model architectures, which makes it difficult to judge what the improvement actually comes from when a new network architecture is invented and introduced as the classifier. In this paper, we propose a multi-resolution front-end for feature extraction in an end-to-end classification framework. Optimal weighted combinations of multiple time-frequency resolutions will be learned automatically given the objective of a classification task. Features extracted with different time-frequency resolutions are weighted and concatenated as inputs to the successive networks, where the weights are predicted by a learnable neural network inspired by the weighting block in squeeze-and-excitation networks (SENet). Furthermore, the refinement of the chosen timefrequency resolutions is investigated by pruning the ones with relatively low importance, which reduces the complexity and size of the model. The proposed method is evaluated on the tasks of speech anti-spoofing in ASVSpoof 2019 and its superiority has been justified by comparing with similar baselines.
Recent research in speech processing exhibits a growing interest in unsupervised and self-supervised representation learning from unlabelled data to alleviate the need for large amounts of annotated data. We investigate several popular pre-training methods and apply them to Flemish Dutch. We compare off-the-shelf English pre-trained models to models trained on an increasing amount of Flemish data. We find that the most important factors for positive transfer to downstream speech recognition tasks include a substantial amount of data and a matching pre-training domain. Ideally, we also finetune on an annotated subset in the target language. All pre-trained models improve linear phone separability in Flemish, but not all methods improve Automatic Speech Recognition. We experience superior performance with wav2vec 2.0 and we obtain a 30% WER improvement by finetuning the multilingually pre-trained XLSR-53 model on Flemish Dutch, after integration into an HMM-DNN acoustic model.
We inspect the long-term learning ability of Long Short-Term Memory language models (LSTM LMs) by evaluating a contextual extension based on the Continuous Bag-of-Words (CBOW) model for both sentence- and discourse-level LSTM LMs and by analyzing its performance. We evaluate on text and speech. Sentence-level models using the long-term contextual module perform comparably to vanilla discourse-level LSTM LMs. On the other hand, the extension does not provide gains for discourse-level models. These findings indicate that discourse-level LSTM LMs already rely on contextual information to perform long-term learning.
End-to-end (E2E) spoken language understanding (SLU) systems avoid an intermediate textual representation by mapping speech directly into intents with slot values. This approach requires considerable domain-specific training data. In low-resource scenarios this is a major concern, e.g., in the present study dealing with SLU for dysarthric speech. Pretraining part of the SLU model for automatic speech recognition targets helps but no research has shown to which extent SLU on dysarthric speech benefits from knowledge transferred from other dysarthric speech tasks. This paper investigates the efficiency of pre-training strategies for SLU tasks on dysarthric speech. The designed SLU system consists of a TDNN acoustic model for feature encoding and a capsule network for intent and slot decoding. The acoustic model is pre-trained in two stages: initialization with a corpus of normal speech and finetuning on a mixture of dysarthric and normal speech. By introducing the intelligibility score as a metric of the impairment severity, this paper quantitatively analyzes the relation between generalization and pathology severity for dysarthric speech.
Objective speech disorder classification for speakers with communication difficulty is desirable for diagnosis and administering therapy. With the current state of speech technology, it is evident to propose neural networks for this application. But neural network model training is hampered by a lack of labeled disordered speech data. In this research, we apply an extended version of Factorized Hierarchical Variational Auto-encoders (FHVAE) for representation learning on disordered speech. The FHVAE model extracts both content-related and sequence-related latent variables from speech data, and we utilize the extracted variables to explore how disorder type information is represented in the latent variables. For better classification performance, the latent variables are aggregated at the word and sentence level. We show that an extension of the FHVAE model succeeds in the better disentanglement of the content-related and sequence-related related representations, but both representations are still required for best results on disorder type classification.
We study the merit of transfer learning for two sound recognition problems, i.e., audio tagging and sound event detection. Employing feature fusion, we adapt a baseline system utilizing only spectral acoustic inputs to also make use of pretrained auditory and visual features, extracted from networks built for different tasks and trained with external data. We perform experiments with these modified models on an audiovisual multi-label data set, of which the training partition contains a large number of unlabeled samples and a smaller amount of clips with weak annotations, indicating the clip-level presence of 10 sound categories without specifying the temporal boundaries of the active auditory events. For clip-based audio tagging, this transfer learning method grants marked improvements. Addition of the visual modality on top of audio also proves to be advantageous in this context. When it comes to generating transcriptions of audio recordings, the benefit of pretrained features depends on the requested temporal resolution: for coarse-grained sound event detection, their utility remains notable. But when more fine-grained predictions are required, performance gains are strongly reduced due to a mismatch between the problem at hand and the goals of the models from which the pretrained vectors were obtained.
Objective: Currently, only behavioral speech understanding tests are available, which require active participation of the person. As this is infeasible for certain populations, an objective measure of speech intelligibility is required. Recently, brain imaging data has been used to establish a relationship between stimulus and brain response. Linear models have been successfully linked to speech intelligibility but require per-subject training. We present a deep-learning-based model incorporating dilated convolutions that can be used to predict speech intelligibility without subject-specific (re)training. Methods: We evaluated the performance of the model as a function of input segment length, EEG frequency band and receptive field size while comparing it to a baseline model. Next, we evaluated performance on held-out data and finetuning. Finally, we established a link between the accuracy of our model and the state-of-the-art behavioral MATRIX test. Results: The model significantly outperformed the baseline for every input segment length (p$\leq10^{-9}$), for all EEG frequency bands except the theta band (p$\leq0.001$) and for receptive field sizes larger than 125 ms (p$\leq0.05$). Additionally, finetuning significantly increased the accuracy (p$\leq0.05$) on a held-out dataset. Finally, a significant correlation (r=0.59, p=0.0154) was found between the speech reception threshold estimated using the behavioral MATRIX test and our objective method. Conclusion: Our proposed dilated convolutional model can be used as a proxy for speech intelligibility. Significance: Our method is the first to predict the speech reception threshold from EEG for unseen subjects, contributing to objective measures of speech intelligibility.