Simon King

Atli Thor Sigurgeirsson, Simon King

Some recent models for Text-to-Speech synthesis aim to transfer the prosody of a reference utterance to the generated target synthetic speech. This is done by using a learned embedding of the reference utterance, which is used to condition speech generation. During training, the reference utterance is identical to the target utterance. Yet, during synthesis, these models are often used to transfer prosody from a reference that differs from the text or speaker being synthesized. To address this inconsistency, we propose to use a different, but prosodically-related, utterance during training too. We believe this should encourage the model to learn to transfer only those characteristics that the reference and target have in common. If prosody transfer methods do indeed transfer prosody they should be able to be trained in the way we propose. However, results show that a model trained under these conditions performs significantly worse than one trained using the target utterance as a reference. To explain this, we hypothesize that prosody transfer models do not learn a transferable representation of prosody, but rather an utterance-level representation which is highly dependent on both the reference speaker and reference text.

Dan Lyth, Simon King

Text-to-speech models trained on large-scale datasets have demonstrated impressive in-context learning capabilities and naturalness. However, control of speaker identity and style in these models typically requires conditioning on reference speech recordings, limiting creative applications. Alternatively, natural language prompting of speaker identity and style has demonstrated promising results and provides an intuitive method of control. However, reliance on human-labeled descriptions prevents scaling to large datasets. Our work bridges the gap between these two approaches. We propose a scalable method for labeling various aspects of speaker identity, style, and recording conditions. We then apply this method to a 45k hour dataset, which we use to train a speech language model. Furthermore, we propose simple methods for increasing audio fidelity, significantly outperforming recent work despite relying entirely on found data. Our results demonstrate high-fidelity speech generation in a diverse range of accents, prosodic styles, channel conditions, and acoustic conditions, all accomplished with a single model and intuitive natural language conditioning. Audio samples can be heard at https://text-description-to-speech.com/.

Opeyemi Osakuade, Simon King

Discrete speech units (DSUs) are derived by quantising representations from models trained using self-supervised learning (SSL). They are a popular representation for a wide variety of spoken language tasks, including those where prosody matters. DSUs are especially convenient for tasks where text and speech are jointly modelled, such as text-to-speech and multimodal dialogue systems. But we have found that DSUs encode suprasegmental information less reliably than segmental structure, which we demonstrate in this work using lexical tone, though this limitation likely extends to other suprasegmental features such as prosody. Our investigations using the tone languages Mandarin and Yorùbá show that the SSL latent representations themselves do encode tone, yet DSUs obtained using quantisation tend to prioritise phonetic structure, which makes lexical tone less reliably encoded. This remains true for a variety of quantisation methods, not only the most common, K-means. We conclude that current DSU quantisation strategies have limitations for suprasegmental features, which suggests a need for new, tone-aware (or prosody-aware) techniques in speech representation learning. We point towards a potential form of the solution by performing K-means clustering once to encode phonetic information, then again on the residual representation, which better encodes lexical tone.

Victor Zheleznov, Stefan Bilbao, Alec Wright et al.

Modal methods are a long-standing approach to physical modelling synthesis. Extensions to nonlinear problems are possible, leading to coupled nonlinear systems of ordinary differential equations. Recent work in scalar auxiliary variable techniques has enabled construction of explicit and stable numerical solvers for such systems. On the other hand, neural ordinary differential equations have been successful in modelling nonlinear systems from data. In this work, we examine how scalar auxiliary variable techniques can be combined with neural ordinary differential equations to yield a stable differentiable model capable of learning nonlinear dynamics. The proposed approach leverages the analytical solution for linear vibration of the system's modes so that physical parameters of a system remain easily accessible after the training without the need for a parameter encoder in the model architecture. Compared to our previous work that used multilayer perceptrons to parametrise nonlinear dynamics, we employ gradient networks that allow an interpretation in terms of a closed-form and non-negative potential required by scalar auxiliary variable techniques. As a proof of concept, we generate synthetic data for the nonlinear transverse vibration of a string and show that the model can be trained to reproduce the nonlinear dynamics of the system. Sound examples are presented.

Opeyemi Osakuade, Simon King

Discrete representations of speech, obtained from Self-Supervised Learning (SSL) foundation models, are widely used, especially where there are limited data for the downstream task, such as for a low-resource language. Typically, discretization of speech into a sequence of symbols is achieved by unsupervised clustering of the latents from an SSL model. Our study evaluates whether discrete symbols - found using k-means - adequately capture tone in two example languages, Mandarin and Yoruba. We compare latent vectors with discrete symbols, obtained from HuBERT base, MandarinHuBERT, or XLS-R, for vowel and tone classification. We find that using discrete symbols leads to a substantial loss of tone information, even for language-specialised SSL models. We suggest that discretization needs to be task-aware, particularly for tone-dependent downstream tasks.

Ariadna Sanchez, Simon King

Speech disorders can make communication hard or even impossible for those who develop them. Personalised Text-to-Speech is an attractive option as a communication aid. We attempt voice reconstruction using a large speech model, with which we generate an approximation of a dysarthric speaker's voice prior to the onset of their condition. In particular, we investigate whether a state-of-the-art large speech model, Parler TTS, can generate intelligible speech while maintaining speaker identity. We curate a dataset and annotate it with relevant speaker and intelligibility information, and use this to fine-tune the model. Our results show that the model can indeed learn to generate from the distribution of this challenging data, but struggles to control intelligibility and to maintain consistent speaker identity. We propose future directions to improve controllability of this class of model, for the voice reconstruction task.

Atli Sigurgeirsson, Simon King

A Prompt-based Text-To-Speech model allows a user to control different aspects of speech, such as speaking rate and perceived gender, through natural language instruction. Although user-friendly, such approaches are on one hand constrained: control is limited to acoustic features exposed to the model during training, and too flexible on the other: the same inputs yields uncontrollable variation that are reflected in the corpus statistics. We investigate a novel fine-tuning regime to address both of these issues at the same time by exploiting the uncontrollable variance of the model. Through principal component analysis of thousands of synthesised samples, we determine latent features that account for the highest proportion of the output variance and incorporate them as new labels for secondary fine-tuning. We evaluate the proposed methods on two models trained on an expressive Icelandic speech corpus, one with emotional disclosure and one without. In the case of the model without emotional disclosure, the method yields both continuous and discrete features that improve overall controllability of the model.

Nicholas Sanders, Yuanchao Li, Korin Richmond et al.

Quantization in SSL speech models (e.g., HuBERT) improves compression and performance in tasks like language modeling, resynthesis, and text-to-speech but often discards prosodic and paralinguistic information (e.g., emotion, prominence). While increasing codebook size mitigates some loss, it inefficiently raises bitrates. We propose Segmentation-Variant Codebooks (SVCs), which quantize speech at distinct linguistic units (frame, phone, word, utterance), factorizing it into multiple streams of segment-specific discrete features. Our results show that SVCs are significantly more effective at preserving prosodic and paralinguistic information across probing tasks. Additionally, we find that pooling before rather than after discretization better retains segment-level information. Resynthesis experiments further confirm improved style realization and slightly improved quality while preserving intelligibility.

Victor Zheleznov, Stefan Bilbao, Alec Wright et al.

Modal synthesis methods are a long-standing approach for modelling distributed musical systems. In some cases extensions are possible in order to handle geometric nonlinearities. One such case is the high-amplitude vibration of a string, where geometric nonlinear effects lead to perceptually important effects including pitch glides and a dependence of brightness on striking amplitude. A modal decomposition leads to a coupled nonlinear system of ordinary differential equations. Recent work in applied machine learning approaches (in particular neural ordinary differential equations) has been used to model lumped dynamic systems such as electronic circuits automatically from data. In this work, we examine how modal decomposition can be combined with neural ordinary differential equations for modelling distributed musical systems. The proposed model leverages the analytical solution for linear vibration of system's modes and employs a neural network to account for nonlinear dynamic behaviour. Physical parameters of a system remain easily accessible after the training without the need for a parameter encoder in the network architecture. As an initial proof of concept, we generate synthetic data for a nonlinear transverse string and show that the model can be trained to reproduce the nonlinear dynamics of the system. Sound examples are presented.

Alistair Carson, Cassia Valentini-Botinhao, Simon King et al.

Machine learning approaches to modelling analog audio effects have seen intensive investigation in recent years, particularly in the context of non-linear time-invariant effects such as guitar amplifiers. For modulation effects such as phasers, however, new challenges emerge due to the presence of the low-frequency oscillator which controls the slowly time-varying nature of the effect. Existing approaches have either required foreknowledge of this control signal, or have been non-causal in implementation. This work presents a differentiable digital signal processing approach to modelling phaser effects in which the underlying control signal and time-varying spectral response of the effect are jointly learned. The proposed model processes audio in short frames to implement a time-varying filter in the frequency domain, with a transfer function based on typical analog phaser circuit topology. We show that the model can be trained to emulate an analog reference device, while retaining interpretable and adjustable parameters. The frame duration is an important hyper-parameter of the proposed model, so an investigation was carried out into its effect on model accuracy. The optimal frame length depends on both the rate and transient decay-time of the target effect, but the frame length can be altered at inference time without a significant change in accuracy.

Atli Thor Sigurgeirsson, Simon King

Reference-based Text-to-Speech (TTS) models can generate multiple, prosodically-different renditions of the same target text. Such models jointly learn a latent acoustic space during training, which can be sampled from during inference. Controlling these models during inference typically requires finding an appropriate reference utterance, which is non-trivial. Large generative language models (LLMs) have shown excellent performance in various language-related tasks. Given only a natural language query text (the prompt), such models can be used to solve specific, context-dependent tasks. Recent work in TTS has attempted similar prompt-based control of novel speaking style generation. Those methods do not require a reference utterance and can, under ideal conditions, be controlled with only a prompt. But existing methods typically require a prompt-labelled speech corpus for jointly training a prompt-conditioned encoder. In contrast, we instead employ an LLM to directly suggest prosodic modifications for a controllable TTS model, using contextual information provided in the prompt. The prompt can be designed for a multitude of tasks. Here, we give two demonstrations: control of speaking style; prosody appropriate for a given dialogue context. The proposed method is rated most appropriate in 50% of cases vs. 31% for a baseline model.

Devang S Ram Mohan, Vivian Hu, Tian Huey Teh et al.

Text does not fully specify the spoken form, so text-to-speech models must be able to learn from speech data that vary in ways not explained by the corresponding text. One way to reduce the amount of unexplained variation in training data is to provide acoustic information as an additional learning signal. When generating speech, modifying this acoustic information enables multiple distinct renditions of a text to be produced. Since much of the unexplained variation is in the prosody, we propose a model that generates speech explicitly conditioned on the three primary acoustic correlates of prosody: $F_{0}$, energy and duration. The model is flexible about how the values of these features are specified: they can be externally provided, or predicted from text, or predicted then subsequently modified. Compared to a model that employs a variational auto-encoder to learn unsupervised latent features, our model provides more interpretable, temporally-precise, and disentangled control. When automatically predicting the acoustic features from text, it generates speech that is more natural than that from a Tacotron 2 model with reference encoder. Subsequent human-in-the-loop modification of the predicted acoustic features can significantly further increase naturalness.

Pilar Oplustil-Gallegos, Simon King

Many speech synthesis datasets, especially those derived from audiobooks, naturally comprise sequences of utterances. Nevertheless, such data are commonly treated as individual, unordered utterances both when training a model and at inference time. This discards important prosodic phenomena above the utterance level. In this paper, we leverage the sequential nature of the data using an acoustic context encoder that produces an embedding of the previous utterance audio. This is input to the decoder in a Tacotron 2 model. The embedding is also used for a secondary task, providing additional supervision. We compare two secondary tasks: predicting the ordering of utterance pairs, and predicting the embedding of the current utterance audio. Results show that the relation between consecutive utterances is informative: our proposed model significantly improves naturalness over a Tacotron 2 baseline.

Berrak Sisman, Junichi Yamagishi, Simon King et al.

Speaker identity is one of the important characteristics of human speech. In voice conversion, we change the speaker identity from one to another, while keeping the linguistic content unchanged. Voice conversion involves multiple speech processing techniques, such as speech analysis, spectral conversion, prosody conversion, speaker characterization, and vocoding. With the recent advances in theory and practice, we are now able to produce human-like voice quality with high speaker similarity. In this paper, we provide a comprehensive overview of the state-of-the-art of voice conversion techniques and their performance evaluation methods from the statistical approaches to deep learning, and discuss their promise and limitations. We will also report the recent Voice Conversion Challenges (VCC), the performance of the current state of technology, and provide a summary of the available resources for voice conversion research.

Zack Hodari, Catherine Lai, Simon King

In English, prosody adds a broad range of information to segment sequences, from information structure (e.g. contrast) to stylistic variation (e.g. expression of emotion). However, when learning to control prosody in text-to-speech voices, it is not clear what exactly the control is modifying. Existing research on discrete representation learning for prosody has demonstrated high naturalness, but no analysis has been performed on what these representations capture, or if they can generate meaningfully-distinct variants of an utterance. We present a phrase-level variational autoencoder with a multi-modal prior, using the mode centres as "intonation codes". Our evaluation establishes which intonation codes are perceptually distinct, finding that the intonation codes from our multi-modal latent model were significantly more distinct than a baseline using k-means clustering. We carry out a follow-up qualitative study to determine what information the codes are carrying. Most commonly, listeners commented on the intonation codes having a statement or question style. However, many other affect-related styles were also reported, including: emotional, uncertain, surprised, sarcastic, passive aggressive, and upset.

Jennifer Williams, Joanna Rownicka, Pilar Oplustil et al.

We aim to characterize how different speakers contribute to the perceived output quality of multi-speaker Text-to-Speech (TTS) synthesis. We automatically rate the quality of TTS using a neural network (NN) trained on human mean opinion score (MOS) ratings. First, we train and evaluate our NN model on 13 different TTS and voice conversion (VC) systems from the ASVSpoof 2019 Logical Access (LA) Dataset. Since it is not known how best to represent speech for this task, we compare 8 different representations alongside MOSNet frame-based features. Our representations include image-based spectrogram features and x-vector embeddings that explicitly model different types of noise such as T60 reverberation time. Our NN predicts MOS with a high correlation to human judgments. We report prediction correlation and error. A key finding is the quality achieved for certain speakers seems consistent, regardless of the TTS or VC system. It is widely accepted that some speakers give higher quality than others for building a TTS system: our method provides an automatic way to identify such speakers. Finally, to see if our quality prediction models generalize, we predict quality scores for synthetic speech using a separate multi-speaker TTS system that was trained on LibriTTS data, and conduct our own MOS listening test to compare human ratings with our NN predictions.

Zack Hodari, Oliver Watts, Simon King

Unlike human speakers, typical text-to-speech (TTS) systems are unable to produce multiple distinct renditions of a given sentence. This has previously been addressed by adding explicit external control. In contrast, generative models are able to capture a distribution over multiple renditions and thus produce varied renditions using sampling. Typical neural TTS models learn the average of the data because they minimise mean squared error. In the context of prosody, taking the average produces flatter, more boring speech: an "average prosody". A generative model that can synthesise multiple prosodies will, by design, not model average prosody. We use variational autoencoders (VAEs) which explicitly place the most "average" data close to the mean of the Gaussian prior. We propose that by moving towards the tails of the prior distribution, the model will transition towards generating more idiosyncratic, varied renditions. Focusing here on intonation, we investigate the trade-off between naturalness and intonation variation and find that typical acoustic models can either be natural, or varied, but not both. However, sampling from the tails of the VAE prior produces much more varied intonation than the traditional approaches, whilst maintaining the same level of naturalness.

Cheng-I Lai, Alberto Abad, Korin Richmond et al.

An attacker may use a variety of techniques to fool an automatic speaker verification system into accepting them as a genuine user. Anti-spoofing methods meanwhile aim to make the system robust against such attacks. The ASVspoof 2017 Challenge focused specifically on replay attacks, with the intention of measuring the limits of replay attack detection as well as developing countermeasures against them. In this work, we propose our replay attacks detection system - Attentive Filtering Network, which is composed of an attention-based filtering mechanism that enhances feature representations in both the frequency and time domains, and a ResNet-based classifier. We show that the network enables us to visualize the automatically acquired feature representations that are helpful for spoofing detection. Attentive Filtering Network attains an evaluation EER of 8.99$\%$ on the ASVspoof 2017 Version 2.0 dataset. With system fusion, our best system further obtains a 30$\%$ relative improvement over the ASVspoof 2017 enhanced baseline system.

Gustav Eje Henter, Simon King, Thomas Merritt et al.

Output from statistical parametric speech synthesis (SPSS) remains noticeably worse than natural speech recordings in terms of quality, naturalness, speaker similarity, and intelligibility in noise. There are many hypotheses regarding the origins of these shortcomings, but these hypotheses are often kept vague and presented without empirical evidence that could confirm and quantify how a specific shortcoming contributes to imperfections in the synthesised speech. Throughout speech synthesis literature, surprisingly little work is dedicated towards identifying the perceptually most important problems in speech synthesis, even though such knowledge would be of great value for creating better SPSS systems. In this book chapter, we analyse some of the shortcomings of SPSS. In particular, we discuss issues with vocoding and present a general methodology for quantifying the effect of any of the many assumptions and design choices that hold SPSS back. The methodology is accompanied by an example that carefully measures and compares the severity of perceptual limitations imposed by vocoding as well as other factors such as the statistical model and its use.

José Novoa, Juan Pablo Escudero, Jorge Wuth et al.

This paper evaluates the robustness of a DNN-HMM-based speech recognition system in highly-reverberant real environments using the HRRE database. The performance of locally-normalized filter bank (LNFB) and Mel filter bank (MelFB) features in combination with Non-negative Matrix Factorization (NMF), Suppression of Slowly-varying components and the Falling edge (SSF) and Weighted Prediction Error (WPE) enhancement methods are discussed and evaluated. Two training conditions were considered: clean and reverberated (Reverb). With Reverb training the use of WPE and LNFB provides WERs that are 3% and 20% lower in average than SSF and NMF, respectively. WPE and MelFB provides WERs that are 11% and 24% lower in average than SSF and NMF, respectively. With clean training, which represents a significant mismatch between testing and training conditions, LNFB features clearly outperform MelFB features. The results show that different types of training, parametrization, and enhancement techniques may work better for a specific combination of speaker-microphone distance and reverberation time. This suggests that there could be some degree of complementarity between systems trained with different enhancement and parametrization methods.

Srikanth Ronanki, Oliver Watts, Simon King et al.

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.

Srikanth Ronanki, Siva Reddy, Bajibabu Bollepalli et al.

Text-to-Speech synthesis in Indian languages has a seen lot of progress over the decade partly due to the annual Blizzard challenges. These systems assume the text to be written in Devanagari or Dravidian scripts which are nearly phonemic orthography scripts. However, the most common form of computer interaction among Indians is ASCII written transliterated text. Such text is generally noisy with many variations in spelling for the same word. In this paper we evaluate three approaches to synthesize speech from such noisy ASCII text: a naive Uni-Grapheme approach, a Multi-Grapheme approach, and a supervised Grapheme-to-Phoneme (G2P) approach. These methods first convert the ASCII text to a phonetic script, and then learn a Deep Neural Network to synthesize speech from that. We train and test our models on Blizzard Challenge datasets that were transliterated to ASCII using crowdsourcing. Our experiments on Hindi, Tamil and Telugu demonstrate that our models generate speech of competetive quality from ASCII text compared to the speech synthesized from the native scripts. All the accompanying transliterated datasets are released for public access.

Zhizheng Wu, Simon King

We propose two novel techniques --- stacking bottleneck features and minimum generation error training criterion --- to improve the performance of deep neural network (DNN)-based speech synthesis. The techniques address the related issues of frame-by-frame independence and ignorance of the relationship between static and dynamic features, within current typical DNN-based synthesis frameworks. Stacking bottleneck features, which are an acoustically--informed linguistic representation, provides an efficient way to include more detailed linguistic context at the input. The minimum generation error training criterion minimises overall output trajectory error across an utterance, rather than minimising the error per frame independently, and thus takes into account the interaction between static and dynamic features. The two techniques can be easily combined to further improve performance. We present both objective and subjective results that demonstrate the effectiveness of the proposed techniques. The subjective results show that combining the two techniques leads to significantly more natural synthetic speech than from conventional DNN or long short-term memory (LSTM) recurrent neural network (RNN) systems.

Zhizheng Wu, Simon King

Recently, recurrent neural networks (RNNs) as powerful sequence models have re-emerged as a potential acoustic model for statistical parametric speech synthesis (SPSS). The long short-term memory (LSTM) architecture is particularly attractive because it addresses the vanishing gradient problem in standard RNNs, making them easier to train. Although recent studies have demonstrated that LSTMs can achieve significantly better performance on SPSS than deep feed-forward neural networks, little is known about why. Here we attempt to answer two questions: a) why do LSTMs work well as a sequence model for SPSS; b) which component (e.g., input gate, output gate, forget gate) is most important. We present a visual analysis alongside a series of experiments, resulting in a proposal for a simplified architecture. The simplified architecture has significantly fewer parameters than an LSTM, thus reducing generation complexity considerably without degrading quality.