This paper presents a novel data augmentation technique for text-to-speech (TTS), that allows to generate new (text, audio) training examples without requiring any additional data. Our goal is to in-crease diversity of text conditionings available during training. This helps to reduce overfitting, especially in low-resource settings. Our method relies on substituting text and audio fragments in a way that preserves syntactical correctness. We take additional measures to ensure that synthesized speech does not contain artifacts caused by combining inconsistent audio samples. The perceptual evaluations show that our method improves speech quality over a number of datasets, speakers, and TTS architectures. We also demonstrate that it greatly improves robustness of attention-based TTS models.

Duration modelling has become an important research problem once more with the rise of non-attention neural textto-speech systems.The current approaches largely fall back to relying on previous statistical parametric speech synthesis technology for duration prediction, which poorly models the expressiveness and variability in speech.In this paper, we propose two alternate approaches to improve duration modelling.First, we propose a duration model conditioned on phrasing that improves the predicted durations and provides better modelling of pauses.We show that the duration model conditioned on phrasing improves the naturalness of speech over our baseline duration model.Second, we also propose a multi-speaker duration model called Cauliflow, that uses normalising flows to predict durations that better match the complex target duration distribution.Cauliflow performs on par with our other proposed duration model in terms of naturalness, whilst providing variable durations for the same prompt and variable levels of expressiveness.Lastly, we propose to condition Cauliflow on parameters that provide an intuitive control of the pacing and pausing in the synthesised speech in a novel way.

In this paper, we present CopyCat2 (CC2), a novel model capable of: a) synthesizing speech with different speaker identities, b) generating speech with expressive and contextually appropriate prosody, and c) transferring prosody at fine-grained level between any pair of seen speakers.We do this by activating distinct parts of the network for different tasks.We train our model using a novel approach to two-stage training.In Stage I, the model learns speaker-independent word-level prosody representations from speech which it uses for many-to-many finegrained prosody transfer.In Stage II, we learn to predict these prosody representations using the contextual information available in text, thereby, enabling multi-speaker TTS with contextually appropriate prosody.We compare CC2 to two strong baselines, one in TTS with contextually appropriate prosody, and one in fine-grained prosody transfer.CC2 reduces the gap in naturalness between our baseline and copy-synthesised speech by 22.79%.In fine-grained prosody transfer evaluations, it obtains a relative improvement of 33.15% in target speaker similarity.

Prosody is an integral part of communication, but remains an open problem in state-of-the-art speech synthesis. There are two major issues faced when modelling prosody: (1) prosody varies at a slower rate compared with other content in the acoustic signal (e.g. segmental information and background noise); (2) determining appropriate prosody without sufficient context is an ill-posed problem. In this paper, we propose solutions to both these issues. To mitigate the challenge of modelling a slow-varying signal, we learn to disentangle prosodic information using a word level representation. To alleviate the ill-posed nature of prosody modelling, we use syntactic and semantic information derived from text to learn a context-dependent prior over our prosodic space. Our context-aware model of prosody (CAMP) outperforms the state-of-the-art technique, closing the gap with natural speech by 26%. We also find that replacing attention with a jointly-trained duration model improves prosody significantly.

In this paper, we introduce Kathaka, a model trained with a novel two-stage training process for neural speech synthesis with contextually appropriate prosody. In Stage I, we learn a prosodic distribution at the sentence level from mel-spectrograms available during training. In Stage II, we propose a novel method to sample from this learnt prosodic distribution using the contextual information available in text. To do this, we use BERT on text, and graph-attention networks on parse trees extracted from text. We show a statistically significant relative improvement of 13.2% in naturalness over a strong baseline when compared to recordings. We also conduct an ablation study on variations of our sampling technique, and show a statistically significant improvement over the baseline in each case.

Many factors influence speech yielding different renditions of a given sentence. Generative models, such as variational autoencoders (VAEs), capture this variability and allow multiple renditions of the same sentence via sampling. The degree of prosodic variability depends heavily on the prior that is used when sampling. In this paper, we propose a novel method to compute an informative prior for the VAE latent space of a neural text-to-speech (TTS) system. By doing so, we aim to sample with more prosodic variability, while gaining controllability over the latent space's structure. By using as prior the posterior distribution of a secondary VAE, which we condition on a speaker vector, we can sample from the primary VAE taking explicitly the conditioning into account and resulting in samples from a specific region of the latent space for each condition (i.e. speaker). A formal preference test demonstrates significant preference of the proposed approach over standard Conditional VAE. We also provide visualisations of the latent space where well-separated condition-specific clusters appear, as well as ablation studies to better understand the behaviour of the system.

We present UTACO, a singing synthesis model based on an attention-based sequence-to-sequence mechanism and a vocoder based on dilated causal convolutions. These two classes of models have significantly affected the field of text-to-speech, but have never been thoroughly applied to the task of singing synthesis. UTACO demonstrates that attention can be successfully applied to the singing synthesis field and improves naturalness over the state of the art. The system requires considerably less explicit modelling of voice features such as F0 patterns, vibratos, and note and phoneme durations, than previous models in the literature. Despite this, it shows a strong improvement in naturalness with respect to previous neural singing synthesis models. The model does not require any durations or pitch patterns as inputs, and learns to insert vibrato autonomously according to the musical context. However, we observe that, by completely dispensing with any explicit duration modelling it becomes harder to obtain the fine control of timing needed to exactly match the tempo of a song.

Prosody Transfer (PT) is a technique that aims to use the prosody from a source audio as a reference while synthesising speech. Fine-grained PT aims at capturing prosodic aspects like rhythm, emphasis, melody, duration, and loudness, from a source audio at a very granular level and transferring them when synthesising speech in a different target speaker's voice. Current approaches for fine-grained PT suffer from source speaker leakage, where the synthesised speech has the voice identity of the source speaker as opposed to the target speaker. In order to mitigate this issue, they compromise on the quality of PT. In this paper, we propose CopyCat, a novel, many-to-many PT system that is robust to source speaker leakage, without using parallel data. We achieve this through a novel reference encoder architecture capable of capturing temporal prosodic representations which are robust to source speaker leakage. We compare CopyCat against a state-of-the-art fine-grained PT model through various subjective evaluations, where we show a relative improvement of $47\%$ in the quality of prosody transfer and $14\%$ in preserving the target speaker identity, while still maintaining the same naturalness.

We present an approach to synthesize whisper by applying a handcrafted signal processing recipe and Voice Conversion (VC) techniques to convert normally phonated speech to whispered speech. We investigate using Gaussian Mixture Models (GMM) and Deep Neural Networks (DNN) to model the mapping between acoustic features of normal speech and those of whispered speech. We evaluate naturalness and speaker similarity of the converted whisper on an internal corpus and on the publicly available wTIMIT corpus. We show that applying VC techniques is significantly better than using rule-based signal processing methods and it achieves results that are indistinguishable from copy-synthesis of natural whisper recordings. We investigate the ability of the DNN model to generalize on unseen speakers, when trained with data from multiple speakers. We show that excluding the target speaker from the training set has little or no impact on the perceived naturalness and speaker similarity of the converted whisper. The proposed DNN method is used in the newly released Whisper Mode of Amazon Alexa.

This paper explores the potential universality of neural vocoders.We train a WaveRNN-based vocoder on 74 speakers coming from 17 languages.This vocoder is shown to be capable of generating speech of consistently good quality (98% relative mean MUSHRA when compared to natural speech) regardless of whether the input spectrogram comes from a speaker or style seen during training or from an out-of-domain scenario when the recording conditions are studio-quality.When the recordings show significant changes in quality, or when moving towards non-speech vocalizations or singing, the vocoder still significantly outperforms speaker-dependent vocoders, but operates at a lower average relative MUSHRA of 75%.These results are shown to be consistent across languages, regardless of them being seen during training (e.g.English or Japanese) or unseen (e.g.

Recurrent Neural Networks (RNN) have recently proved to be effective in acoustic modeling for TTS. Various techniques such as the Maximum Likelihood Parameter Generation (MLPG) algorithm have been naturally inherited from the HMM-based speech synthesis framework. This paper investigates in which situations parameter generation and variance restoration approaches help for RNN-based TTS. We explore how their performance is affected by various factors such as the choice of the loss function, the application of regularization methods and the amount of training data. We propose an efficient way to calculate MLPG using a convolutional kernel. Our results show that the use of the L1 loss with proper regularization outperforms any system built with the conventional L2 loss and does not require to apply MLPG (which is necessary otherwise). We did not observe perceptual improvements when embedding MLPG into the acoustic model. Finally, we show that variance restoration approaches are important for cepstral features but only yield minor perceptual gains for the prediction of F0.

Statistical TTS systems that directly predict the speech waveform have recently reported improvements in synthesis quality. This investigation evaluates Amazon's statistical speech waveform synthesis (SSWS) system. An in-depth evaluation of SSWS is conducted across a number of domains to better understand the consistency in quality. The results of this evaluation are validated by repeating the procedure on a separate group of testers. Finally, an analysis of the nature of speech errors of SSWS compared to hybrid unit selection synthesis is conducted to identify the strengths and weaknesses of SSWS. Having a deeper insight into SSWS allows us to better define the focus of future work to improve this new technology.

Pitch detection is a fundamental problem in speech processing as F0 is used in a large number of applications. Recent papers have proposed deep learning for robust pitch tracking. In this letter, we consider voicing detection as a classification problem and F0 contour estimation as a regression problem. For both tasks, acoustic features from multiple domains and traditional machine learning methods are used. The discrimination power of existing and proposed features is assessed through mutual information. Multiple supervised and unsupervised approaches are compared. A significant relative reduction of voicing errors over the best baseline is obtained-20% with the best clustering method (K-means) and 45% with a multi-layer perceptron. For F0 contour estimation, the benefits of regression techniques are limited though. We investigate whether those objective gains translate in a parametric synthesis task. Clear perceptual preferences are observed for the proposed approach over two widely used baselines (robust algorithm for pitch tracking (RAPT) and distributed inline-filter operation (DIO)).

This paper proposes a method to improve the quality delivered by statistical parametric speech synthesizers. For this, we use a codebook of pitch-synchronous residual frames, so as to construct a more realistic source signal. First a limited codebook of typical excitations is built from some training database. During the synthesis part, HMMs are used to generate filter and source coefficients. The latter coefficients contain both the pitch and a compact representation of target residual frames. The source signal is obtained by concatenating excitation frames picked up from the codebook, based on a selection criterion and taking target residual coefficients as input. Subjective results show a relevant improvement compared to the basic technique.

INTERSPEECH2009: 10th Annual Conference of the International Speech Communication Association, September 6-10, 2009, Brighton, UK.

The subject of this paper is the conversion of a given speaker's voice (the source speaker) into another identified voice (the target one). We assume we have at our disposal a large amount of speech samples from source and target voice with at least a part of them being parallel. The proposed system is built on a mapping function between source and target spectral envelopes followed by a frame selection algorithm to produce final spectral envelopes. Converted speech is produced by a basic LP analysis of the source and LP synthesis using the converted spectral envelopes. We compared three types of conversion: without mapping, with mapping and using the excitation of the source speaker and finally with mapping using the excitation of the target. Results show that the combination of mapping and frame selection provide the best results, and underline the interest to work on methods to convert the LP excitation.