We present a neural vocoder designed with low-powered Alternative and Augmentative Communication devices in mind. By combining elements of successful modern vocoders with established ideas from an older generation of technology, our system is able to produce high quality synthetic speech at 48kHz on devices where neural vocoders are otherwise prohibitively complex. The system is trained adversarially using differentiable pitch synchronous overlap add, and reduces complexity by relying on pitch synchronous Inverse Short-Time Fourier Transform (ISTFT) to generate speech samples. Our system achieves comparable quality with a strong baseline (HiFi-GAN) while using only a fraction of the compute. We present results of a perceptual evaluation as well as an analysis of system complexity.

Automatically predicting the outcome of subjective listening tests is a challenging task. Ratings may vary from person to person even if preferences are consistent across listeners. While previous work has focused on predicting listeners' ratings (mean opinion scores) of individual stimuli, we focus on the simpler task of predicting subjective preference given two speech stimuli for the same text. We propose a model based on anti-symmetric twin neural networks, trained on pairs of waveforms and their corresponding preference scores. We explore both attention and recurrent neural nets to account for the fact that stimuli in a pair are not time aligned. To obtain a large training set we convert listeners' ratings from MUSHRA tests to values that reflect how often one stimulus in the pair was rated higher than the other. Specifically, we evaluate performance on data obtained from twelve MUSHRA evaluations conducted over five years, containing different TTS systems, built from data of different speakers. Our results compare favourably to a state-of-the-art model trained to predict MOS scores.

This paper presents a method for upsampling and transforming a compact representation of acoustics into a corresponding speech waveform. Similar to a conventional vocoder, the proposed system takes a pulse train derived from fundamental frequency and a noise sequence as inputs and shapes them to be consistent with the acoustic features. However, the filters that are used to shape the waveform in the proposed system are learned from data, and take the form of layers in a convolutional neural network. Because the network performs the transformation simultaneously for all waveform samples in a sentence, its synthesis speed is comparable with that of conventional vocoders on CPU, and many times faster on GPU. It is trained directly in a fast and straightforward manner, using a combined time- and frequency-domain objective function. We use publicly available data and provide code to allow our results to be reproduced.

This paper explores whether adding Discourse Relation (DR) features improves the naturalness of neural statistical parametric speech synthesis (SPSS) in English. We hypothesize first - in the light of several previous studies - that DRs have a dedicated prosodic encoding. Secondly, we hypothesize that encoding DRs in a speech synthesizer's input will improve the naturalness of its output. In order to test our hypotheses, we prepare a dataset of DR-annotated transcriptions of audiobooks in English. We then perform an acoustic analysis of the corpus which supports our first hypothesis that DRs are acoustically encoded in speech prosody. The analysis reveals significant correlation between specific DR categories and acoustic features, such as F0 and intensity. Then, we use the corpus to train a neural SPSS system in two configurations: a baseline configuration making use only of conventional linguistic features, and an experimental one where these are supplemented with DRs. Augmenting the inputs with DR features improves objective acoustic scores on a test set and leads to significant preference by listeners in a forced choice AB test for naturalness.

There are many aspects of speech that we might want to control when creating text-to-speech (TTS) systems. We present a general method that enables control of arbitrary aspects of speech, which we demonstrate on the task of emotion control. Current TTS systems use supervised machine learning and are therefore heavily reliant on labelled data. If no labels are available for a desired control dimension, then creating interpretable control becomes challenging. We introduce a method that uses external, labelled data (i.e. not the original data used to train the acoustic model) to enable the control of dimensions that are not labelled in the original data. Adding interpretable control allows the voice to be manually controlled to produce more engaging speech, for applications such as audiobooks. We evaluate our method using a listening test.

This paper presents a simple but effective method for generating speech waveforms by selecting small units of stored speech to match a low-dimensional target representation. The method is designed as a drop-in replacement for the vocoder in a deep neural network-based text-to-speech system. Most previous work on hybrid unit selection waveform generation relies on phonetic annotation for determining unit boundaries, or for specifying target cost, or for candidate preselection. In contrast, our waveform generator requires no phonetic information, annotation, or alignment. Unit boundaries are determined by epochs, and spectral analysis provides representations which are compared directly with target features at runtime. As in unit selection, we minimise a combination of target cost and join cost, but find that greedy left-to-right nearest-neighbour search gives similar results to dynamic programming. The method is fast and can generate the waveform incrementally. We use publicly available data and provide a permissively-licensed open source toolkit for reproducing our results.

This paper presents a hybrid text-to-speech framework that uses a waveform generation method based on examplars of natural speech waveform. These examplars are selected at synthesis time given a sequence of acoustic features generated from text by a statistical parametric speech synthesis model. In order to match the expected degradation of these target synthesis features, the database of units is constructed such that the units' target representations are generated from the same parametric model. We evaluate two variants of this framework by modifying the size of the examplar: a small unit variant (where unit boundaries are determined by pitch mark location) and a halfphone variant (where unit boundaries are determined by subphone state forced alignment). We found that for a larger dataset (around four hours of training data) the examplar-based waveform generation variants are rated higher than the vocoder-based system.

When a text-to-speech (TTS) system is required to speak world news, a large fraction of the words to be spoken will be proper names originating in a wide variety of languages. Phonetization of these names based on target language letter-to-sound rules will typically be inadequate. This is detrimental not only during synthesis, when inappropriate phone sequences are produced, but also during training, if the system is trained on data from the same domain. This is because poor phonetization during forced alignment based on hidden Markov models can pollute the whole model set, resulting in degraded alignment even of normal target-language words. This paper presents four techniques designed to address this issue in the context of a Swahili TTS system: automatic transcription of proper names based on a lexicon from a better-resourced language; the addition of a parallel phone set and special part-of-speech tag exclusively dedicated to proper names; a manually-crafted phone mapping which allows substitutions for potentially more accurate phones in proper names during forced alignment; the addition in proper names of a grapheme-derived frame-level feature, supplementing the standard phonetic inputs to the acoustic model. We present results from objective and subjective evaluations of systems built using these four techniques.

This paper presents a simple count-based approach to learning word vector representations by leveraging statistics of cooccurrences between text and speech. This type of representation requires two discrete sequences of units defined across modalities. Two possible methods for the discretization of an acoustic signal are presented, which are then applied to fundamental frequency and energy contours of a transcribed corpus of speech, yielding a sequence of textual objects (e.g. words, syllables) aligned with a sequence of discrete acoustic events. Constructing a matrix recording the co-occurrence of textual objects with acoustic events and reducing its dimensionality with matrix decomposition results in a set of context-independent representations of word types. These are applied to the task of acoustic modelling for speech synthesis; objective and subjective results indicate that these representations are useful for the generation of acoustic parameters in a text-to-speech (TTS) system. In general, we observe that the more discretization approaches, acoustic signals, and levels of linguistic analysis are incorporated into a TTS system via these count-based representations, the better that TTS system performs.

Current approaches to statistical parametric speech synthesis using Neural Networks generally require input at the same temporal resolution as the output, typically a frame every 5ms, or in some cases at waveform sampling rate. It is therefore necessary to fabricate highly-redundant frame-level (or sample level) linguistic features at the input. This paper proposes the use of a hierarchical encoder-decoder model to perform the sequence-to-sequence regression in a way that takes the input linguistic features at their original timescales, and preserves the relationships between words, syllables and phones. The proposed model is designed to make more effective use of suprasegmental features than conventional architectures, as well as being computationally efficient. Experiments were conducted on prosodically-varied audiobook material because the use of supra-segmental features is thought to be particularly important in this case. Both objective measures and results from subjective listening tests, which asked listeners to focus on prosody, show that the proposed method performs significantly better than a conventional architecture that requires the linguistic input to be at the acoustic frame rate. We provide code and a recipe to enable our system to be reproduced using the Merlin toolkit.

Accurate modelling and prediction of speech-sound durations is an important component in generating more natural synthetic speech. Deep neural networks (DNNs) offer a powerful modelling paradigm, and large, found corpora of natural and expressive speech are easy to acquire for training them. Unfortunately, found datasets are seldom subject to the quality-control that traditional synthesis methods expect. Common issues likely to affect duration modelling include transcription errors, reductions, filled pauses, and forced-alignment inaccuracies. To combat this, we propose to improve modelling and prediction of speech durations using methods from robust statistics, which are able to disregard ill-fitting points in the training material. We describe a robust fitting criterion based on the density power divergence (the ß-divergence) and a robust generation heuristic using mixture density networks (MDNs). Perceptual tests indicate that subjects prefer synthetic speech generated using robust models of duration over the baselines.

We investigate two wavelet-based decomposition strategies of the f0 signal and their usefulness as a secondary task for speech synthesis using multi-task deep neural networks (MTL-DNN). The first decomposition strategy uses a static set of scales for all utterances in the training data. We propose a second strategy, where the scale of the mother wavelet is dynamically adjusted to the rate of each utterance. This approach is able to capture f0 variations related to the syllable, word, clitic-group, and phrase units. This method also constrains the wavelet components to be within the frequency range that previous experiments have shown to be more natural. These two strategies are evaluated as a secondary task in multi-task deep neural networks (MTL-DNNs). Results indicate that on an expressive dataset there is a strong preference for the systems using multi-task learning when compared to the baseline system.

Deep neural networks (DNNs) have recently been the focus of much text-to-speech research as a replacement for decision trees and hidden Markov models (HMMs) in statistical parametric synthesis systems. Performance improvements have been reported; however, the configuration of systems evaluated makes it impossible to judge how much of the improvement is due to the new machine learning methods, and how much is due to other novel aspects of the systems. Specifically, whereas the decision trees in HMM-based systems typically operate at the state-level, and separate trees are used to handle separate acoustic streams, most DNN-based systems are trained to make predictions simultaneously for all streams at the level of the acoustic frame. This paper isolates the influence of three factors (machine learning method; state vs. frame predictions; separate vs. combined stream predictions) by building a continuum of systems along which only a single factor is varied at a time. We find that replacing decision trees with DNNs and moving from state-level to frame-level predictions both significantly improve listeners' naturalness ratings of synthetic speech produced by the systems. No improvement is found to result from switching from separate-stream to combined-stream predictions.

A top-down hierarchical system based on deep neural networks is investigated for the modeling of prosody in speech synthesis. Suprasegmental features are processed separately from segmental features and a compact distributed representation of high-level units is learned at syllable-level. The suprasegmental representation is then integrated into a frame-level network. Objective measures show that balancing segmental and suprasegmental features can be useful for the frame-level network. Additional features incorporated into the hierarchical system are then tested. At the syllable-level, a bag-of-phones representation is proposed and, at the word-level, embeddings learned from text sources are used. It is shown that the hierarchical system is able to leverage new features at higher-levels more efficiently than a system which exploits them directly at the frame-level. A perceptual evaluation of the proposed systems is conducted and followed by a discussion of the results.

A problem when developing and tuning speech synthesis systems is that there is no well-established method of automatically rating the quality of the synthetic speech. This research attempts to obtain a new automated measure which is trained on the result of large-scale subjective evaluations employing many human listeners, i.e., the Blizzard Challenge. To exploit the data, we experiment with linear regression, feed-forward and convolutional neural network models, and combinations of them to regress from synthetic speech to the perceptual scores obtained from listeners. The biggest improvements were seen when combining stimulus- and system-level predictions.

This paper proposes a new approach to duration modelling for statistical parametric speech synthesis in which a recurrent statistical model is trained to output a phone transition probability at each timestep (acoustic frame). Unlike conventional approaches to duration modelling - which assume that duration distributions have a particular form (e.g., a Gaussian) and use the mean of that distribution for synthesis - our approach can in principle model any distribution supported on the non-negative integers. Generation from this model can be performed in many ways; here we consider output generation based on the median predicted duration. The median is more typical (more probable) than the conventional mean duration, is robust to training-data irregularities, and enables incremental generation. Furthermore, a frame-level approach to duration prediction is consistent with a longer-term goal of modelling durations and acoustic features together. Results indicate that the proposed method is competitive with baseline approaches in approximating the median duration of held-out natural speech.

Deep neural networks (DNNs) use a cascade of hidden representations to enable the learning of complex mappings from input to output features. They are able to learn the complex mapping from text-based linguistic features to speech acoustic features, and so perform text-to-speech synthesis. Recent results suggest that DNNs can produce more natural synthetic speech than conventional HMM-based statistical parametric systems. In this paper, we show that the hidden representation used within a DNN can be improved through the use of Multi-Task Learning, and that stacking multiple frames of hidden layer activations (stacked bottleneck features) also leads to improvements. Experimental results confirmed the effectiveness of the proposed methods, and in listening tests we find that stacked bottleneck features in particular offer a significant improvement over both a baseline DNN and a benchmark HMM system.

Grapheme-based models have been proposed for both ASR and TTS as a way of circumventing the lack of expert-compiled pronunciation lexicons in under-resourced languages. It is a common observation that this should work well in languages employing orthographies with a transparent letter-to-phoneme relationship,such as Spanish. Our experience has shown, however,that there is still a significant difference in intelligibility between grapheme-based systems and conventional ones for this language. This paper explores the contribution of different levels of linguistic annotation to system intelligibility, and the trade-off between those levels and the quantity of data used for training. Ten systems spaced across these two continua of knowledge and data were subjectively evaluated for intelligibility.

This paper introduces a novel form of parametric synthesis that uses context embeddings produced by the bottleneck layer of a deep neural network to guide the selection of models in a rich-context HMM-based synthesiser. Rich-context synthesis – in which Gaussian distributions estimated from single lin-guistic contexts seen in the training data are used for synthesis, rather than more conventional decision tree-tied models – was originally proposed to address over-smoothing due to averag-ing across contexts. Our previous investigations have confirmed experimentally that averaging across different contexts is in-deed one of the largest factors contributing to the limited quality of statistical parametric speech synthesis. However, a possible weakness of the rich context approach as previously formulated is that a conventional tied model is still used to guide selection of Gaussians at synthesis time. Our proposed approach replaces this with context embeddings derived from a neural network. Index Terms: speech synthesis, hidden Markov model, deep neural networks, rich context, embedding

This paper describes the use of a low-dimensional vector representation of sentence acoustics to control the output of a feed-forward deep neural network text-to-speech system on a sentence-by-sentence basis. Vector representations for sentences in the training corpus are learned during network training along with other parameters of the model. Although the network is trained on a frame-by-frame basis, the standard frame-level inputs representing linguistic features are supplemented by features from a projection layer which outputs a learned representation of sentence-level acoustic characteristics. The projection layer contains dedicated parameters for each sentence in the training data which are optimised jointly with the standard network weights. Sentence-specific parameters are optimised on all frames of the relevant sentence -- these parameters therefore allow the network to account for sentence-level variation in the data which is not predictable from the standard linguistic inputs. Results show that the global prosodic characteristics of synthetic speech can be controlled simply and robustly at run time by supplementing basic linguistic features with sentence-level control vectors which are novel but designed to be consistent with those observed in the training corpus.

The use of shared projection neural nets of the sort used in language modelling is proposed as a way of sharing parameters between multiple text-to-speech system components. We experiment with pretraining the weights of such a shared projection on an auxiliary language modelling task and then apply the resulting word representations to the task of phrase-break prediction. Doing so allows us to build phrase-break predictors that rival conventional systems without any reliance on conventional knowledge-based resources such as part of speech taggers.

Audiobooks have been focused on as promising data for training Text-to-Speech (TTS) systems. However, they usually do not have a correspondence between audio and text data. Moreover, they are usually divided only into chapter units. In practice, we have to make a correspondence of audio and text data before we use them for building TTS synthesisers. However aligning audio and text data is time-consuming and involves manual labor. It also requires persons skilled in speech processing. Previously, we have proposed to use graphemes for automatically aligning speech and text data. This paper further integrates a lightly supervised voice activity detection (VAD) technique to detect sentence boundaries as a pre-processing step before the grapheme approach. This lightly supervised technique requires time stamps of speech and silence only for the first fifty sentences. Combining those, we can semi-automatically build TTS systems from audiobooks with minimum manual intervention. From subjective evaluations we analyse how the grapheme-based aligner and/or the proposed VAD technique impact the quality of HMM-based speech synthesisers trained on audiobooks.

Simple4All Tundra (version 1.0) is the first release of a standardised multilingual corpus designed for text-to-speech re-search with imperfect or found data. The corpus consists of approximately 60 hours of speech data from audiobooks in 14 languages, as well as utterance-level alignments obtained with a lightly-supervised process. Future versions of the corpus will include finer-grained alignment and prosodic annotation, all of which will be made freely available. This paper gives a gen-eral outline of the data collected so far, as well as a detailed description of how this has been done, emphasizing the mini-mal language-specific knowledge and manual intervention used to compile the corpus. To demonstrate its potential use, text-to-speech systems have been built for all languages using unsu-pervised or lightly supervised methods, also briefly presented in the paper. Index Terms: multilingual corpus, light supervision, imperfect data, found data, text-to-speech, audiobook data

Part of speech (POS) tags are foremost among the features conventionally used to predict intonational phrase-breaks for text to speech (TTS) conversion. The construction of such systems therefore presupposes the availability of a POS tagger for the relevant language, or of a corpus manually tagged with POS. However, such tools and resources are not available in the majority of the world’s languages, and manually labelling text with POS tags is an expensive and time-consuming process. We therefore propose the use of continuous-valued features that summarise the distributional characteristics of word types as surrogates for POS features. Importantly, such features are obtained in an unsupervised manner from an untagged text corpus. We present results on the phrase-break prediction task, where use of the features closes the gap in performance between a baseline system (using only basic punctuation-related features) and a topline system (incorporating a state-of-the-art POS tagger).

We explore the use of linguistic features for text to speech (TTS) conversion in the context of a speech-to-speech translation system that can be extracted from unannotated text in an unsupervised, language-independent fashion. The features are intended to act as surrogates for conventional part of speech (POS) features. Unlike POS features, the experimental features assume only the availability of tools and data that must already be in place for the construction of other components of the translation system, and can therefore be used for the TTS module without incurring additional TTS-specific costs. We here describe the use of the experimental features in a speech synthesiser, using six different configurations of the system to allow the comparison of the proposed features with conventional, knowledge-based POS features. We present results of objective and subjective evaluations of the usefulness of the new features.

We analyse the contribution of higher-level elements of the linguistic specification of a data-driven speech synthesiser to the naturalness of the synthetic speech which it generates. The system is trained using various subsets of the full feature-set, in which features relating to syntactic category, intonational phrase boundary, pitch accent and boundary tones are selectively removed. Utterances synthesised by the different configurations of the system are then compared in a subjective evaluation of their naturalness. The work presented forms background analysis for an ongoing set of experiments in performing text-to-speech (TTS) conversion based on shallow features: features that can be trivially extracted from text. By building a range of systems, each assuming the availability of a different level of linguistic annotation, we obtain benchmarks for our on-going work.

In speaker-adaptive HMM-based speech synthesis, there are a few speakers whose synthetic speech sounds worse than that of other speakers, despite having the same amount of adapta-tion data from within the same corpus. This paper investigates these fluctuations in quality and found that as mel-cepstral dis-tance from the average voice becomes larger, the MOS scores generally become worse. Although the negative correlation ob-tained is not strong enough, this helps us improve the training and adaptation strategies for average voice models. Further-more we remark that this correlation is strongly linked to “vocal attractiveness.” Index Terms: speech synthesis, HMM, average voice, speaker adaptation

The synthesis of child speech presents challenges both in the collection of data and in the building of a synthesizer from that data. We chose to build a statistical parametric synthesizer using the hidden Markov model (HMM)-based system HTS, as this technique has previously been shown to perform well for limited amounts of data, and for data collected under imperfect conditions. Six different configurations of the synthesizer were compared, using both speaker-dependent and speaker-adaptive modeling techniques, and using varying amounts of data. For comparison with HMM adaptation, techniques from voice conversion were used to transform existing synthesizers to the characteristics of the target speaker. Speaker-adaptive voices generally outperformed child speaker-dependent voices in the evaluation. HMM adaptation outperformed voice conversion style techniques when using the full target speaker corpus; with fewer adaptation data, however, no significant listener preference for either HMM adaptation or voice conversion methods was found.

In conventional speech synthesis, large amounts of phonetically balanced speech data recorded in highly controlled recording studio environments are typically required to build a voice. Although using such data is a straightforward solution for high quality synthesis, the number of voices available will always be limited, because recording costs are high. On the other hand, our recent experiments with HMM-based speech synthesis systems have demonstrated that speaker-adaptive HMM-based speech synthesis (which uses an “average voice model” plus model adaptation) is robust to non-ideal speech data that are recorded under various conditions and with varying microphones, that are not perfectly clean, and/or that lack phonetic balance. This enables us to consider building high-quality voices on “non-TTS” corpora such as ASR corpora. Since ASR corpora generally include a large number of speakers, this leads to the possibility of producing an enormous number of voices automatically. In this paper, we demonstrate the thousands of voices for HMM-based speech synthesis that we have made from several popular ASR corpora such as the Wall Street Journal (WSJ0, WSJ1, and WSJCAM0), Resource Management, Globalphone, and SPEECON databases. We also present the results of associated analysis based on perceptual evaluation, and discuss remaining issues.

This study compares two different methodologies for producing data-driven synthesis of child speech from existing systems that have been trained on the speech of adults. On one hand, an existing statistical parametric synthesiser is transformed using model adaptation techniques, informed by linguistic and prosodic knowledge, to the speaker characteristics of a child speaker. This is compared with the application of voice conversion techniques to convert the output of an existing waveform concatenation synthesiser with no explicit linguistic or prosodic knowledge. In a subjective evaluation of the similarity of synthetic speech to natural speech from the target speaker, the HMM-based systems evaluated are generally preferred, although this is at least in part due to the higher dimensional acoustic features supported by these techniques. Index Terms: child speech, statistical parametric speech synthesis, HMM-based speech synthesis, voice conversion, HTS, Average Voice Models, Festival

Our recent experiments with HMM-based speech synthesis systems have demonstrated that speaker-adaptive HMM-based speech synthesis (which uses an 'average voice model' plus model adaptation) is robust to non-ideal speech data that are recorded under various conditions and with varying microphones, that are not perfectly clean, and/or that lack of phonetic balance. This enables us consider building high-quality voices on 'non-TTS' corpora such as ASR corpora. Since ASR corpora generally include a large number of speakers, this leads to the possibility of producing an enormous number of voices automatically. In this paper we show thousands of voices for HMM-based speech synthesis that we have made from several popular ASR corpora such as the Wall Street Journal databases (WSJ0/WSJ1/WSJCAM0), Resource Management, Globalphone and Speecon. We report some perceptual evaluation results and outline the outstanding issues.