Robust learning of acoustic representations from diverse speech data
Automatic speech recognition is increasingly applied to new domains. A key challenge is to robustly learn, update and maintain representations to cope with transient acoustic conditions. A typical example is broadcast media, for which speakers and environments may change rapidly, and available supervision may be poor. The concern of this thesis is to build and investigate methods for acoustic modelling that are robust to the characteristics and transient conditions as embodied by such media. The first contribution of the thesis is a technique to make use of inaccurate transcriptions as supervision for acoustic model training. There is an abundance of audio with approximate labels, but training methods can be sensitive to label errors, and their use is therefore not trivial. State-of-the-art semi-supervised training makes effective use of a lattice of supervision, inherently encoding uncertainty in the labels to avoid overfitting to poor supervision, but does not make use of the transcriptions. Existing approaches that do aim to make use of the transcriptions typically employ an algorithm to filter or combine the transcriptions with the recognition output from a seed model, but the final result does not encode uncertainty. We propose a method to combine the lattice output from a biased recognition pass with the transcripts, crucially preserving uncertainty in the lattice where appropriate. This substantially reduces the word error rate on a broadcast task. The second contribution is a method to factorise representations for speakers and environments so that they may be combined in novel combinations. In realistic scenarios, the speaker or environment transform at test time might be unknown, or there may be insufficient data to learn a joint transform. We show that in such cases, factorised, or independent, representations are required to avoid deteriorating performance. Using i-vectors, we factorise speaker or environment information using multi-condition training with neural networks. Specifically, we extract bottleneck features from networks trained to classify either speakers or environments. The resulting factorised representations prove beneficial when one factor is missing at test time, or when all factors are seen, but not in the desired combination. The third contribution is an investigation of model adaptation in a longitudinal setting. In this scenario, we repeatedly adapt a model to new data, with the constraint that previous data becomes unavailable. We first demonstrate the effect of such a constraint, and show that using a cyclical learning rate may help. We then observe that these successive models lend themselves well to ensembling. Finally, we show that the impact of this constraint in an active learning setting may be detrimental to performance, and suggest to combine active learning with semi-supervised training to avoid biasing the model. The fourth contribution is a method to adapt low-level features in a parameter-efficient and interpretable manner. We propose to adapt the filters in a neural feature extractor, known as SincNet. In contrast to traditional techniques that warp the filterbank frequencies in standard feature extraction, adapting SincNet parameters is more flexible and more readily optimised, whilst maintaining interpretability. On a task adapting from adult to child speech, we show that this layer is well suited for adaptation and is very effective with respect to the small number of adapted parameters.