Jump to Content
Rob Clark

Rob Clark

Rob Clark received his PhD from the University of Edinburgh in 2003. His primary interest is in producing engaging synthetic speech. Before joining Google Rob was at the University of Edinburgh for many years involved in both teaching and research relating to text-to-speech synthesis. Rob was one of the primary developers and maintainers of the open source Festival text-to-speech synthesis system.
Authored Publications
Google Publications
Other Publications
Sort By
  • Title
  • Title, desc
  • Year
  • Year, desc
    Preview abstract The quality of synthetic speech is typically evaluated using subjective listening tests. An underlying assumption is that these tests are reliable, i.e., running the test multiple times gives consistent results. A common approach to study reliability is a replication study. Existing studies focus primarily on Mean Opinion Score (MOS), and few consider the error bounds from the original test. In contrast, we present a replication study of both MOS and AB preference tests to answer two questions: (1) which of the two test types is more reliable for system comparison, and (2) for both test types, how reliable are the results with respect to their estimated standard error? We find that while AB tests are more reliable for system comparison, standard errors are underestimated for both test types. We show that these underestimates are partially due to broken independence assumptions, and suggest alternate methods of standard error estimation that account for dependencies among ratings. View details
    Preview abstract Most text-to-speech acoustic models, such as WaveNet, Tacotron, ClariNet etc., use either a phoneme sequence or a letter sequence as the foundational unit of speech. Although the letter (or grapheme) sequence more closely matches the actual runtime input of the TTS system, it often fails to represent the fine-grained and often plentiful grapheme-to-phoneme relationships of the target language. A purely phonemic input seems to perform better in practice, though is heavily dependent on a scrupulous phonology and lexicon to provide the model with the phoneme sequences. This reliance poses issues (namely with quality and consistency) which can lead to the need for a trade-off between quality and scalability. In order to overcome this, we propose using a mix of the two inputs, namely providing both phonemic and graphemic identities to the model. In this paper, we show that this approach can help the model learn to disambiguate some of the more subtle phonemic variations (such as the realization of reduced vowels), and that this effect improves the fidelity to the accent of the original voice talent. We present a way of generating an unbiased targeted test using phoneme spectral diffs, and using that, show improvement over the baseline approach. Since different types of neural networks build on top of the same input feature space, we show that the improvement scales to multiple voice technologies, and on several languages. View details
    Preview abstract Transfer tasks in text-to-speech (TTS) synthesis — where one or more aspects of the speech of one set of speakers is transferred to another set of speakers that do not feature these aspects originally — remains a challenging task. One of the challenges is that models that have high-quality transfer capabilities can have issues in stability, making them impractical for user-facing critical tasks. This paper demonstrates that transfer can be obtained by training an robust TTS system on data generated by a less robust TTS system designed for a high-quality transfer task; In particular, a CHiVE-BERT monolingual TTS system is trained on the output of a Tacotron model designed for accent transfer. While some quality loss is inevitable with this approach, experimental results show that the models trained on synthetic data this way can produce high quality audio displaying accent transfer, while preserving speaker characteristics such as speaking style. View details
    Preview abstract Recently, WaveNet has become a popular choice of neural network to synthesize speech audio. Autoregressive WaveNet is capable of producing high-fidelity audio, but is too slow for real-time synthesis. As a remedy, Parallel WaveNet was proposed, which can produce audio faster than real time through distillation of an autoregressive teacher into a feedforward student network. A shortcoming of this approach, however, is that a large amount of recorded speech data is required to produce high-quality student models, and this data is not always available. In this paper, we propose StrawNet: a self-training approach to train a Parallel WaveNet. Self-training is performed using the synthetic examples generated by the autoregressive WaveNet teacher. We show that, in low-data regimes, training on high-fidelity synthetic data from an autoregressive teacher model is superior to training the student model on (much fewer) examples of recorded speech. We compare StrawNet to a baseline Parallel WaveNet, using both side-by-side tests and Mean Opinion Score evaluations. To our knowledge, synthetic speech has not been used to train neural text-to-speech before. View details
    Preview abstract The prosody of currently available speech synthesis systems can be unnatural due to the systems only having access to the text, possibly enriched by linguistic information such as part-of-speech tags and parse trees. We show that incorporating a BERT model in an RNN-based speech synthesis model - where the BERT model is pretrained on large amounts of unlabeled data, and fine-tuned to the speech domain - improves prosody. Additionally, we propose a way of handling arbitrarily long sequences with BERT. Our findings indicate that small BERT models work better than big ones, and that fine-tuning the BERT part of the model is pivotal for getting good results. View details
    Preview abstract Intonation is characterized by rises and falls in pitch and energy. In previous work, we explicitly modelled these prosodic features using Clockwork Hierarchical Variational Autoencoders (CHiVE) to show we can generate multiple intonation contours for any text. However, recent advances in text-to-speech synthesis produce spectrograms which are inverted by neural vocoders to produce waveforms. Spectrograms encode intonation in a complex way; there is no simple, explicit representation analogous to pitch (fundamental frequency) and energy. In this paper, we extend CHiVE to model intonation within a spectrogram. Compared to the original model, the spectrogram extension gives better mean opinion scores in subjective listening tests. We show that the intonation in the generated spectrograms match the intonations represented by the generated pitch curves. View details
    Preview abstract The prosodic aspects of speech signals produced by current text-to-speech systems are typically averaged over training material, and as such lack the variety and liveliness found in natural speech. To avoid monotony and averaged prosody contours, it is desirable to have a way of modeling the variation in the prosodic aspects of speech, so audio signals can be synthesized in multiple ways for a given text. We present a new, hierarchically structured conditional variational autoencoder to generate prosodic features (fundamental frequency, energy and duration) suitable for use with a vocoder or a generative model like WaveNet. At inference time, an embedding representing the prosody of a sentence may be sampled from the variational layer to allow for prosodic variation. To efficiently capture the hierarchical nature of the linguistic input (words, syllables and phones), both the encoder and decoder parts of the auto-encoder are hierarchical, in line with the linguistic structure, with layers being clocked dynamically at the respective rates. We show in our experiments that our dynamic hierarchical network outperforms a non-hierarchical state-of-the-art baseline, and, additionally, that prosody transfer across sentences is possible by employing the prosody embedding of one sentence to generate the speech signal of another. View details
    Preview abstract Text-to-speech systems are typically evaluated on single sentences. When long-form content, such as data consisting of full paragraphs or dialogues is considered, evaluating sentences in isolation is not always appropriate as the context in which the sentences are synthesized is missing. In this paper, we investigate three different ways of evaluating the naturalness of long-form text-to-speech synthesis. We compare the results obtained from evaluating sentences in isolation, evaluating whole paragraphs of speech, and presenting a selection of speech or text as context and evaluating the subsequent speech. We find that, even though these three evaluations are based upon the same material, the outcomes differ per setting, and moreover that these outcomes do not necessarily correlate with each other. We show that our findings are consistent between a single speaker setting of read paragraphs and a two-speaker dialogue scenario. We conclude that to evaluate the quality of long-form speech, the traditional way of evaluating sentences in isolation does not suffice, and that multiple evaluations are required. View details
    Preview abstract This paper introduces a new speech corpus called ``LibriTTS'' designed for text-to-speech use. It is derived from the original audio and text materials of the LibriSpeech corpus, which has been used for training and evaluating automatic speech recognition systems. The new corpus inherits desired properties of the LibriSpeech corpus while addressing a number of issues which make LibriSpeech less than ideal for text-to-speech work. The released corpus consists of 585 hours of speech data at 24kHz sampling rate from 2,456 speakers and the corresponding texts. Experimental results show that neural end-to-end TTS models trained from the LibriTTS corpus achieved above 4.0 in mean opinion scores in naturalness in five out of six evaluation speakers. The corpus is freely available for download from http://www.openslr.org/60/. View details
    Preview abstract We present an extension to the Tacotron speech synthesis architecture that learns a latent embedding space of prosody, derived from a reference acoustic representation containing the desired prosody. We show that conditioning Tacotron on this learned embedding space results in synthesized audio that matches the reference signal’s prosody with fine time detail. We define several quantitative and subjective metrics for evaluating prosody transfer, and report results and audio samples from a single-speaker and 44-speaker Tacotron model on a prosody transfer task. View details
    Uncovering Latent Style Factors for Expressive Speech Synthesis
    Yuxuan Wang
    Ying Xiao
    NIPS Workshop on Machine Learning for Audio Signal Processing (ML4Audio) (2017) (to appear)
    Preview abstract Prosodic modeling is a core problem in speech synthesis. The key challenge is producing desirable prosody from textual input containing only phonetic information. In this preliminary study, we introduce the concept of "style tokens" in Tacotron, a recently proposed end-to-end neural speech synthesis model. Using style tokens, we aim to extract independent prosodic styles from training data. We show that without annotation data or an explicit supervision signal, our approach can automatically learn a variety of prosodic variations in a purely data-driven way. Importantly, each style token corresponds to a fixed style factor regardless of the given text sequence. As a result, we can control the prosodic style of synthetic speech in a somewhat predictable and globally consistent way. View details
    Preview abstract A text-to-speech synthesis system typically consists of multiple stages, such as a text analysis frontend, an acoustic model and an audio synthesis module. Building these components often requires extensive domain expertise and may contain brittle design choices. In this paper, we present Tacotron, an end-to-end generative text-to-speech model that synthesizes speech directly from characters. Given (text, audio) pairs, the model can be trained completely from scratch with random initialization. We present several key techniques to make the sequence-to-sequence framework perform well for this challenging task. Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English, outperforming a production parametric system in terms of naturalness. In addition, since Tacotron generates speech at the frame level, it's substantially faster than sample-level autoregressive methods. View details
    No Results Found