Spoken command recognition for robotics
Declaration I hereby declare that except where specific reference is made to the work of others, the contents of this dissertation are original and have not been submitted in whole or in part for consideration for any other degree or qualification in this, or any other university. This dissertation is my own work and contains nothing which is the outcome of work done in collaboration with others, except as specified in the text and Acknowledgements. This dissertation contains fewer than 65,000
... ords including appendices, bibliography, footnotes, tables and equations and has fewer than 150 figures. Abstract In this thesis, I investigate spoken command recognition technology for robotics. While high robustness is expected, the distant and noisy conditions in which the system has to operate make the task very challenging. Unlike commercial systems which all rely on a "wake-up" word to initiate the interaction, the pipeline proposed here directly detect and recognizes commands from the continuous audio stream. In order to keep the task manageable despite low-resource conditions, I propose to focus on a limited set of commands, thus trading off flexibility of the system against robustness. Domain and speaker adaptation strategies based on a multi-task regularization paradigm are first explored. More precisely, two different methods are proposed which rely on a tied loss function which penalizes the distance between the output of several networks. The first method considers each speaker or domain as a task. A canonical task-independent network is jointly trained with task-dependent models, allowing both types of networks to improve by learning from one another. While an improvement of 3.2% on the frame error rate (FER) of the task-independent network is obtained, this only partially carried over to the phone error rate (PER), with 1.5% of improvement. Similarly, a second method explored the parallel training of the canonical network with a privileged model having access to i-vectors. This method proved less effective with only 1.2% of improvement on the FER. In order to make the developed technology more accessible, I also investigated the use of a sequence-to-sequence (S2S) architecture for command classification. The use of an attention-based encoder-decoder model reduced the classification error by 40% relative to a strong convolutional neural network (CNN)-hidden Markov model (HMM) baseline, showing the relevance of S2S architectures in such context. In order to improve the flexibility of the trained system, I also explored strategies for few-shot learning, which allow to extend the set of commands with minimum requirements in terms of data. Retraining a model on the combination of original and new commands, I managed to achieve 40.5% of accuracy on the new commands with only 10 examples for each of them. This scores goes up to 81.5% of accuracy with a larger set of 100 examples per new command. An alternative strategy, based on model adaptation achieved even better scores, with 68.8% and 88.4% of accuracy with 10 v and 100 examples respectively, while being faster to train. This high performance is obtained at the expense of the original categories though, on which the accuracy deteriorated. Those results are very promising as the methods allow to easily extend an existing S2S model with minimal resources. Finally, a full spoken command recognition system (named iCubrec) has been developed for the iCub platform. The pipeline relies on a voice activity detection (VAD) system to propose a fully hand-free experience. By segmenting only regions that are likely to contain commands, the VAD module also allows to reduce greatly the computational cost of the pipeline. Command candidates are then passed to the deep neural network (DNN)-HMM command recognition system for transcription. The VoCub dataset has been specifically gathered to train a DNN-based acoustic model for our task. Through multi-condition training with the CHiME4 dataset, an accuracy of 94.5% is reached on VoCub test set. A filler model, complemented by a rejection mechanism based on a confidence score, is finally added to the system to reject non-command speech in a live demonstration of the system.