Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Speech-to-intent systems detect spoken commands from a person and perform the right action in return. A good way to implement these systems is ASSIST. It maps a spoken command to a semantic frame that captures the intent of an utterance. The system is retrained in every new environment to avoid inconsistencies with problem-specific vocabulary. Four different models are proposed in ASSIST, namely RCCN, PCCN, encoder-decoder and NMF. The purpose of this thesis is to check whether data augmentation can improve the accuracy of each of the four models in ASSIST when few training utterances are available. Two different types of methods are proposed. The first type contains the model-driven methods. New features are created by speeding up/slowing down the original utterances and by adapting the way of how the signal is split up in different frames during the feature calculation. The second type is the FHVAE method. An FHVAE is a neural network model that encodes the input utterance in two latent vectors. One contains the within speaker variation and the other the between speaker variation. By perturbing one or both of the vectors a little bit and by reconstructing the utterance by passing through a decoder, new utterances can be created that differ a little bit from the original. Both model-driven methods give a significant improvement for all models of ASSIST except for RCCN. The best result is achieved with the combination of both techniques. PCCN benefits the most followed by NMF and encoder-decoder. FHVAE works best for PCCN and NMF. Both give a significant improvement compared to the original data when few training samples are available. For NMF, this is even significantly better than the combination of both model-driven methods. When more training data is available, the accuracy decreases to less than the value for the original data. The best results are obtained through perturbing the within speaker latent variance vector. RCCN and encoder-decoder do not give a significant improvement. RCCN does not benefit from data augmentation due to its attention layer. PCCN benefits the most because it was initially overfit. The simplicity and efficacy of the model-driven methods proves their usefulness. The FHVAE techniques are promising but not yet perfect. They suffer from a loss of information while passing an utterance through the FHVAE. Adaptations to the architecture of the FHVAE can further improve on this.