Search

Speech technology has developed to levels equivalent with human parity through the use of deep neural networks. However, it is unclear how the learned dependencies within these networks can be attributed to metrics such as recognition performance. This research focuses on strategies to interpret and exploit these learned context dependencies to improve speech recognition models. Context dependency analysis had not yet been explored for speech recognition networks. In order to highlight and observe dependent representations within speech recognition models, a novel analysis framework is proposed. This analysis framework uses statistical correlation indexes to compute the coefficiency between neural representations. By comparing the coefficiency of neural representations between models using different approaches, it is possible to observe specific context dependencies within network layers. By providing insights on context dependencies it is then possible to adapt modelling approaches to become more computationally efficient and improve recognition performance. Here the performance of End-to-End speech recognition models are analysed, providing insights on the acoustic and language modelling context dependencies. The modelling approach for a speaker recognition task is adapted to exploit acoustic context dependencies and reach comparable performance with the state-of-the-art methods, reaching 2.89% equal error rate using the Voxceleb1 training and test sets with 50% of the parameters. Furthermore, empirical analysis of the role of acoustic context for speech emotion recognition modelling revealed that emotion cues are presented as a distributed event. These analyses and results for speech recognition applications aim to provide objective direction for future development of automatic speech recognition systems.

In pronunciation assessment (PA) of second language (L2) speech, it is known that similarity to a native accent is desired, yet not crucial. There are certain variations in pronunciation which do not interfere with communication. It is up to the listener to decide whether a pronunciation differs from the one of so-called canonical reference. The subjectivity in pronunciation assessment can be referred to as the assessor bias. A computer-assisted pronunciation assessment is subject to the effects of assessor bias. The disagreement between assessors causes inconsistencies in the data used to build models for the assessment task. A model for the bias itself, however, would help build a general reference for a proficient L2 speaker as well as an impartial PA. This thesis proposes a model for the assessor bias to be included as part of a model for a pronunciation assessor. The assessor model consists of an ideal assessor-independent scoring function for PA, modified by an additive term specific to the assessor. The latter term is referred to as bias. The research for the model resulted in four original contributions. All contributions were tested on data from L2 speech from young learners of English in the Netherlands. Each recording was annotated for mispronunciation at the phoneme level by three trained phoneticians. Overlapping annotation made the data the best fit for a consistent model of inter-assessor disagreement. A first contribution is a novel approach for detecting mispronunciations without the need for a precise phoneme alignment, which outperformed a baseline of pronunciation correctness scores based on phoneme alignments. The second contribution is a study of the effect of speaker metadata on learning a pronunciation reference. Models trained on different assessors were proven to be sensitive to different speaker information. The third contribution was the proposal and implementation of the assessor model. Two deep networks combine a bidirectional long short-term memory module with self-attention and a feed-forward classifier to estimate the probabilities of phonemes being pronounced correctly. Both networks were trained jointly to estimate the observed pronunciation labels. Only one network was modelled on the assessor's identity. The fourth contribution consists of methods for increasing the specialisation of the bias networks by reducing its cosine similarity and co-dependence with respect to the assessor-independent network. Using cosine similarity and a contrastive log-ratio upper bound for mutual information, it was possible to both reduce the correlation and dependency between the two networks. The bias network managed to increase its dependence on assessor identity and speaker factors. The mutual information between the assessor and the bias output was useful to illustrate disagreement, as well as which assessors and phonemes were the most prone to the bias.

In speaker recognition, deep neural networks deliver state-of-the-art performance due to their large capacities and powerful feature extraction abilities. However, this performance can be highly affected by interference from background noise and other speakers. This thesis focuses on new neural network architectures that are designed to overcome such interference and thereby improve the robustness of the speaker recognition system. In order to improve the noise robustness of the speaker recognition model, two novel network architectures are proposed. The first is the hierarchical attention network, which is able to capture both local and global features in order to improve the robustness of the network. The experimental results show it can deliver results that are comparable to the published state-of-the-art methods, reaching 4.28% equal error rate using the Voxceleb1 training and test sets. The second approach is the speech enhancement and speaker recognition joint system that consists of two networks; the first integrates speech enhancement and speaker recognition into one framework to better filter out noise, while the other makes further use of speaker embeddings input to a speech enhancement network. This provides prior knowledge for the speech enhancement network which improves its performance. The results show that a joint system with a speaker dependent speech enhancement model can deliver results that are comparable to the published state-of-the-art methods, reaching 4.15% equal error rate using the Voxceleb1 training and test sets. In order to overcome interfering speaker, two novel approaches are proposed. The first is referred to as the embedding de-mixing approach that separates the speaker and content properties from a two-speaker signal in an embedding space, rather than in a signal space. The results show that the de-mixed embeddings are close to the clean embeddings in terms of quality, and the back-end speaker recognition model can make use of the de-mixed embeddings to reach 96.9% speaker identification accuracy, compared to those achieved using clean embeddings (98.5%) on TIMIT dataset. The second approach is the first end-to-end weakly supervised speaker identification approach based on a novel hierarchical transformer network architecture. The results show that the proposed model can capture speaker properties from two speakers in one input utterance. The hierarchical transformer network can reach more than 3% relative improvement compared to the baselines in all of the test conditions.

The Fourier analysis plays a key role in speech signal processing. As a complex quantity, it can be expressed in the polar form using the magnitude and phase spectra. The magnitude spectrum is widely used in almost every corner of speech processing. However, the phase spectrum is not an obviously appealing start point for processing the speech signal. In contrast to the magnitude spectrum whose fine and coarse structures have a clear relation to speech perception, the phase spectrum is difficult to interpret and manipulate. In fact, there is not a meaningful trend or extrema which may facilitate the modelling process. Nonetheless, the speech phase spectrum has recently gained renewed attention. An expanding body of work is showing that it can be usefully employed in a multitude of speech processing applications. Now that the potential for the phase-based speech processing has been established, there is a need for a fundamental model to help understand the way in which phase encodes speech information. In this thesis a novel phase-domain source-filter model is proposed that allows for deconvolution of the speech vocal tract (filter) and excitation (source) components through phase processing. This model utilises the Hilbert transform, shows how the excitation and vocal tract elements mix in the phase domain and provides a framework for efficiently segregating the source and filter components through phase manipulation. To investigate the efficacy of the suggested approach, a set of features is extracted from the phase filter part for automatic speech recognition (ASR) and the source part of the phase is utilised for fundamental frequency estimation. Accuracy and robustness in both cases are illustrated and discussed. In addition, the proposed approach is improved by replacing the log with the generalised logarithmic function in the Hilbert transform and also by computing the group delay via regression filter. Furthermore, statistical distribution of the phase spectrum and its representations along the feature extraction pipeline are studied. It is illustrated that the phase spectrum has a bell-shaped distribution. Some statistical normalisation methods such as mean-variance normalisation, Laplacianisation, Gaussianisation and Histogram equalisation are successfully applied to the phase-based features and lead to a significant robustness improvement. The robustness gain achieved through using statistical normalisation and generalised logarithmic function encouraged the use of more advanced model-based statistical techniques such as vector Taylor Series (VTS). VTS in its original formulation assumes usage of the log function for compression. In order to simultaneously take advantage of the VTS and generalised logarithmic function, a new formulation is first developed to merge both into a unified framework called generalised VTS (gVTS). Also in order to leverage the gVTS framework, a novel channel noise estimation method is developed. The extensions of the gVTS framework and the proposed channel estimation to the group delay domain are then explored. The problems it presents are analysed and discussed, some solutions are proposed and finally the corresponding formulae are derived. Moreover, the effect of additive noise and channel distortion in the phase and group delay domains are scrutinised and the results are utilised in deriving the gVTS equations. Experimental results in the Aurora-4 ASR task in an HMM/GMM set up along with a DNN-based bottleneck system in the clean and multi-style training modes confirmed the efficacy of the proposed approach in dealing with both additive and channel noise.

The performance of speech recognition systems is known to degrade in mismatched conditions, where the acoustic environment and the speaker population significantly differ between the training and target test data. Performance degradation due to the mismatch is widely reported in the literature, particularly for diverse datasets. This thesis approaches the mismatch problem in diverse datasets with various strategies including data refinement, variability modelling and speech recognition model adaptation. These strategies are realised in six novel contributions. The first contribution is a data subset selection technique using likelihood ratio derived from a target test set quantifying mismatch. The second contribution is a multi-style training method using data augmentation. The existing training data is augmented using a distribution of variabilities learnt from a target dataset, resulting in a matched set. The third contribution is a new approach for genre identification in diverse media data with the aim of reducing the mismatch in an adaptation framework. The fourth contribution is a novel method which performs an unsupervised domain discovery using latent Dirichlet allocation. Since the latent domains have a high correlation with some subjective meta-data tags, such as genre labels of media data, features derived from the latent domains are successfully applied to the genre and broadcast show identification tasks. The fifth contribution extends the latent modelling technique for acoustic model adaptation, where latent-domain specific models are adapted from a base model. As the sixth contribution, an alternative adaptation approach is proposed where subspace adaptation of deep neural network acoustic models is performed using the proposed latent-domain aware training procedure. All of the proposed techniques for mismatch reduction are verified using diverse datasets. Using data selection, data augmentation and latent-domain model adaptation methods the mismatch between training and testing conditions of diverse ASR systems are reduced, resulting in more robust speech recognition systems.

Distant speech recognition (DSR) has gained wide interest recently. While deep networks keep improving ASR overall, the performance gap remains between using close-talking recordings and distant recordings. Therefore the work in this thesis aims at providing some insights for further improvement of DSR performance. The investigation starts with collecting the first multi-microphone and multi-media corpus of natural spontaneous multi-party conversations in native English with the speaker location tracked, i.e. the Sheffield Wargame Corpus (SWC). The state-of-the-art recognition systems with the acoustic models trained standalone and adapted both show word error rates (WERs) above 40% on headset recordings and above 70% on distant recordings. A comparison between SWC and AMI corpus suggests a few unique properties in the real natural spontaneous conversations, e.g. the very short utterances and the emotional speech. Further experimental analysis based on simulated data and real data quantifies the impact of such influence factors on DSR performance, and illustrates the complex interaction among multiple factors which makes the treatment of each influence factor much more difficult. The reverberation factor is studied further. It is shown that the reverberation effect on speech features could be accurately modelled with a temporal convolution in the complex spectrogram domain. Based on that a polynomial reverberation score is proposed to measure the distortion level of short utterances. Compared to existing reverberation metrics like C50, it avoids a rigid early-late-reverberation partition without compromising the performance on ranking the reverberation level of recording environments and channels. Furthermore, the existing reverberation measurement is signal independent thus unable to accurately estimate the reverberation distortion level in short recordings. Inspired by the phonetic analysis on the reverberation distortion via self-masking and overlap-masking, a novel partition of reverberation distortion into the intra-phone smearing and the inter-phone smearing is proposed, so that the reverberation distortion level is first estimated on each part and then combined.

Speaker diarisation answers the question “who spoke when?” in an audio recording. The input may vary, but a system is required to output speaker labelled segments in time. Typical stages are Speech Activity Detection (SAD), speaker segmentation and speaker clustering. Early research focussed on Conversational Telephone Speech (CTS) and Broadcast News (BN) domains before the direction shifted to meetings and, more recently, broadcast media. The British Broadcasting Corporation (BBC) supplied data through the Multi-Genre Broadcast (MGB) Challenge in 2015 which showed the difficulties speaker diarisation systems have on broadcast media data. Diarisation is typically an unsupervised task which does not use auxiliary data or information to enhance a system. However, methods which do involve supplementary data have shown promise. Five semi-supervised methods are investigated which use a combination of inputs: different channel types and transcripts. The methods involve Deep Neural Networks (DNNs) for SAD, DNNs trained for channel detection, transcript alignment, and combinations of these approaches. However, the methods are only applicable when datasets contain the required inputs. Therefore, a method involving a pretrained Speaker Separation Deep Neural Network (ssDNN) is investigated which is applicable to every dataset. This technique performs speaker clustering and speaker segmentation using DNNs successfully for meeting data and with mixed results for broadcast media. The task of diarisation focuses on two aspects: accurate segments and speaker labels. The Diarisation Error Rate (DER) does not evaluate the segmentation quality as it does not measure the number of correctly detected segments. Other metrics exist, such as boundary and purity measures, but these also mask the segmentation quality. An alternative metric is presented based on the F-measure which considers the number of hypothesis segments correctly matched to reference segments. A deeper insight into the segment quality is shown through this metric.

This thesis presents a study into the nature and structure of academic lectures, with a special focus on metadiscourse phenomena. Metadiscourse refers to a set of linguistics expressions that signal specific discourse functions such as the Introduction: “Today we will talk about ... ” and Emphasising: “This is an important point”. These functions are important because they are part of lecturers’ strategies in understanding of what happens in a lecture. The knowledge of their presence and identity could serve as initial steps toward downstream applications that will require functional analysis of lecture content such as a browser for lectures archives, summarisation, or an automatic minute-taker for lectures. One challenging aspect for metadiscourse detection and classification is that the set of expressions are semi-fixed, meaning that different phrases can indicate the same function. To that end a four-stage approach is developed to study metadiscourse in academic lectures. Firstly, a corpus of metadiscourse for academic lectures from Physics and Economics courses is built by adapting an existing scheme that describes functional-oriented metadiscourse categories. Second, because producing reference transcripts is a time-consuming task and prone to some errors due to the manual efforts required, an automatic speech recognition (ASR) system is built specifically to produce transcripts of lectures. Since the reference transcripts lack time-stamp information, an alignment system is applied to the reference to be able to evaluate the ASR system. Then, a model is developed using Support Vector Machines (SVMs) to classify metadiscourse tags using both textual and acoustical features. The results show that n-grams are the most inductive features for the task; however, due to data sparsity the model does not generalise for unseen n-grams. This limits its ability to solve the variation issue in metadiscourse expressions. Continuous Bag-of-Words (CBOW) provide a promising solution as this can capture both the syntactic and semantic similarities between words and thus is able to solve the generalisation issue. However, CBOW ignores the word order completely, something which is very important to be retained when classifying metadiscourse tags. The final stage aims to address the issue of sequence modelling by developing a joint CBOW and Convolutional Neural Network (CNN) model. CNNs can work with continuous features such as word embedding in an elegant and robust fashion by producing a fixed-size feature vector that is able to identify indicative local information for the tagging task. The results show that metadiscourse tagging using CNNs outperforms the SVMs model significantly even on ASR outputs, owing to its ability to predict a sequence of words that is more representative for the task regardless of its position in the sentence. In addition, the inclusion of other features such as part-of-speech (POS) tags and prosodic cues improved the results further. These findings are consistent in both disciplines. The final contribution in this thesis is to investigate the suitability of using metadiscourse tags as discourse features in the lecture structure segmentation model, despite the fact that the task is approached as a classification model and most of the state-of-art models are unsupervised. In general, the obtained results show remarkable improvements over the state-of-the-art models in both disciplines.

Colloquial Arabic (CA) is the set of spoken variants of modern Arabic that exist in the form of regional dialects and are considered generally to be mother-tongues in those regions. CA has limited textual resource because it exists only as a spoken language and without a standardised written form. Normally the modern standard Arabic (MSA) writing convention is employed that has limitations in phonetically representing CA. Without phonetic dictionaries the pronunciation of CA words is ambiguous, and can only be obtained through word and/or sentence context. Moreover, CA inherits the MSA complex word structure where words can be created from attaching affixes to a word. In automatic speech recognition (ASR), commonly used approaches to model acoustic, pronunciation and word variability are language independent. However, one can observe significant differences in performance between English and CA, with the latter yielding up to three times higher error rates. This thesis investigates the main issues for the under-performance of CA ASR systems. The work focuses on two directions: first, the impact of limited lexical coverage, and insufficient training data for written CA on language modelling is investigated; second, obtaining better models for the acoustics and pronunciations by learning to transfer between written and spoken forms. Several original contributions result from each direction. Using data-driven classes from decomposed text are shown to reduce out-of-vocabulary rate. A novel colloquialisation system to import additional data is introduced; automatic diacritisation to restore the missing short vowels was found to yield good performance; and a new acoustic set for describing CA was defined. Using the proposed methods improved the ASR performance in terms of word error rate in a CA conversational telephone speech ASR task.