WO2022263806A1

WO2022263806A1 - Text-to-speech system

Info

Publication number: WO2022263806A1
Application number: PCT/GB2022/051491
Authority: WO
Inventors: Ranniery Maia; Christophe Veaux; Kerem SOZUGECER; Taylan KAMIS
Original assignee: Deep Zen Limited
Priority date: 2021-06-14
Filing date: 2022-06-14
Publication date: 2022-12-22
Also published as: GB2607903A; GB202108468D0

Abstract

A text-to-speech method, comprising a training phase in which the system is trained with a plurality of emotional tagged data, tagged with a plurality of different emotions which are applied to a GST model to estimate emotion-dependent style embedding; generating a Gaussian mixture model (GMM) on said emotion-dependent style embedding, each Gaussian component representative of one emotion, and; at the time of synthesis, sampling said emotion-dependent style embeddings from the GMM and applying these as an input for controlled speech synthesis.

Description

Text-to-Speech System

This invention relates to a text-to-speech (TTS) system. In particular, it relates to a TTS system which can convey emotion.

TTS systems are well known, and can receive text as an input and output this as synthesised spoken speech. Early TTS systems outputted speech in a rather robotic, monotone manner. There is, however, an increasing desire for TTS systems which mimic the human voice more closely, including expressing emotion associated with the text. Thus, expressing in an "angry " voice where appropriate, in a "sad" voice and so on. Ultimately the aim is to generate a TTS system that a listener cannot distinguish from a human reading out the text.

A system which uses global style tokens (GST) for expressive speech synthesis has been proposed in:

[1] Y. Wang, D. Stanton, Y. Zhang, R.-S. Ryan, E. Battenberg, J. Shor, Y. Xiao, Y. Jia, F. Ren, and R. A. Saurous, "Style tokens: Unsupervised style modeling, control and transfer in end-to-end speech synthesis," in Proceedings of the 35th International Conference on Machine Learning (J. Dy and A. Krause, eds.), vol. 80 of Proceedings of Machine Learning Research, (Stockholm Sweden), pp. 5180- 5189, PMLR, 10-15 Jul 2018.

Since then, these have been used as a way of conveying different styles and emotions in neural TTS systems like that known as "Tacotron" and described in:

[2] J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y. Zhang, Y. Wang, R. Skerry- Ryan, R. A. Saurous, Y. Agiomyrgiannakis, and Y. Wu, "Natural TTS synthesis by conditioning wavenet on mel spectrogram predictions," in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), (Calgary, Canada), pp. 4779^1783, Apr. 2018. An additive multi-head attention module system building on this is described in:

[3] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, "Attention is all you need," in Advances in Neural Information Processing Systems (I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, eds.), vol. 30, Curran Associates, Inc., 2017. The GSTs are tokens that can be used to control various styles/emotions of speaking. They typically are a set of vectors that contain a target speaker's prosody and speaking style information. The goal of them was to represent and (to some extent) control speech style by using a large expressive speech database such as audio books. The assumption by the authors of these techniques was that underlying speech expressions can be factorised as a weighted sum of a set of style tokens or global style tokens by means of a style model which is trained with the TTS system in the machine learning environment.

Although these techniques are generally effective, they have several drawbacks. Firstly, GSTs are difficult to interpret in terms of specific styles, and second, it is difficult to synthesise speech having a specific style or mix of specific styles. Whilst some attempts have been made to control speech styles at the time of synthesis using GST, each of these have required a specific style to be used at the time of synthesis.

In particular, normal human speech generally uses a mix of emotions. Thus a human reader may express a degree of happiness but also simultaneously of fear and some excitement in a certain situation, or other mixed emotions. . This is very difficult or impossible to achieve with existing GST type methods which are based on single emotions. The present invention arose in an attempt to provide an improved method and system, expressive TTS systems and speech synthesis in which a wide range of emotions can be portrayed by the synthesised speech.

According to the present invention in a first aspect there is provided a text- to-speech method, comprising a training phrase in which the system is trained with a plurality of emotionally tagged data tagged with a plurality of different emotions which are applied to a GST model to estimate emotion-dependent style embedding; generating a Gaussian mixture model (GMM) on said emotion- dependent style embedding, one gaussian component for each of the plurality of emotions, and, at the time of synthesis, sampling said emotion-dependent style embeddings from each gaussian component of the GMM to obtain a combined mixed emotion scores and applying these as an input for controlled speech synthesis

Thus the speech synthesis output represents a combined emotion, which was not possible with previously proposed systems.

Preferably (but not exclusively) the plurality of emotions are selected from:

Anger

Happiness

Sadness

Excitement

Surprise

Fear

Disgust

The training step of the expressive TTS system preferably comprises providing a first set of training data of relatively large amount and diversity, and a second set of training data which can be of relatively short amount (compared to the first set) and which are tagged according to the predominant emotion. The invention further comprises a TTS system comprising apparatus configured to use the method.

Other aspects and features of the invention are disclosed in the dependent claims, description and drawings.

Embodiments of the invention will now be described, by way of example only, with reference to the accompanying drawings in which;

Figure 1 shows a conventional global style token (GST) emotion modelling system;

Figure 2 shows an overall diagram of a text to speech (TTS) system with emotion control;

Figure 3 shows a training method ofan emotion control module; and

Figure 4 shows a synthesis/influence method of an emotion control module.

As described, global style tokens (GST) for expressive speech synthesis have been proposed and used to convey different styles and emotions in TTS systems.

Essentially, a given speech signal is first compressed into a compact vector, known as a reference embedding, by an encoder. The reference embedding is then fed into an attention layer which determines scores of similarity between the embedding and entries from the set of style tokens. These then go through a softmax operation (as known in the art) which results in a so called "condition vector" (CV) which represents how similar the reference embedding is to each token. The combination of the condition vector and the GSTs results in a style embedding. The style embedding construction process is shown schematically in Figure 1 where a plurality of tokens 1 a, 1 b to 1 k are generated and input to the attention layer 2 to create a series of attention scores 2a, 2b to 2k. After a softmax operation the scores comprise a condition vector 3. The combination of the condition vector (CV) 3 and the GSTs 1 a to 1 k results in a style embedding, which is essentially a weighted summation of the GSTs where the weights are elements of the condition vector. This final style embedding is then used to condition the TTS or generate a specific style of emotion. Note that this will relate to a single emotion - Fear, say, or more generally a single style provided from the reference embedding.

The style layers may be implemented as an additive multi-head attention (MHA) model, as described in reference [3] above, ie a plurality of heads (or subvectors) in particular emotional styles.

In more detail, In the GST approach, a given speech signal is first compressed into a compact vector, a reference embedding, by an encoder. The reference embedding is then fed into an attention layer where the goal is not to align but to determine scores of similarity between the embedding and entries from a set of style tokens. After going through a softmax operation these scores compose a so-called condition vector (CV), and represent how similar is the reference embedding to each token of the GST bank. The combination of a CV and GSTs result into a style embedding, given by

where

and

are respectively the entries of the GST bank and components of the CV, with K being the number of tokens, and each GST being a D-dimensional vector, i.e.

The style embedding s is then used to condition the TTS onto a specific style or emotion. In [1] the style layer is implemented as an additive multi-head attention (MHA) module [3]. In this case the style embedding is a concatenation of individual head- dependent style embeddings

where

are respectively CV components and

GST bank entries for head h. In this case the dimension of each token is divided by the number of heads, i.e. . If the MHA has H heads,

the final style embedding becomes:

The MHA realization of GST is a specific case under the style factorization emotion control proposed by [1]

In embodiments of the present invention, a plurality of emotional training samples are used, in different emotional styles. These may, for example, be: anger, happiness, sadness, excitement, surprise, fear and disgust. A training piece is read by a reader using a particular emotion and this is labelled. Thus a user may read a piece in an "angry" voice and it is labelled as such, similarly with a "happy" voice, a "sad" voice, and so on. This generates a plurality of speech audio signals as training samples which are labelled with appropriate emotions. One emotion is associated with each labelled sample.

The training data for the TTS system may also comprise one or more typically longer samples of people reading text in a neutral and not over-expressive voice. A typical example may be four hours long, for example but it may be any length.

For training the entire system but emotionally tagged and neutral data will preferably be used. For style control training it may be possible to just use emotionally tagged data. The general technique is shown in Figure 2 and more specific training and synthesis parts are shown in more detail in Figures 3 and 4.

Referring to Figure 2, the training data, which may include text 5 and audio 6 samples, is stored in a database 7. This is then applied to a training file 8 where a style model is trained together with the TTS system, driven by the TTS loss. This means that the TTS system + emotion control are trained altogether in a way to reproduce a synthesized speech signal that can be as natural as possible.

Or, in other words, the goal of the joint training is to provide a speech signal that can be as close as possible to its natural version.

Thus, the training phase includes TTS training 9 and style model training 10 which provides style embeddings 11 to the TTS training. This results in trained models 12 including a TTS model 13 and a style model 14. As described below, at the training stage emotion condition vectors (CVs) are acquired and accumulated and a Gaussian Mixture Model (GMM) is fitted on these, where each GMM component is expected to represent each of the different emotions. These represent the trained models 12.

At synthesis time 15 the text to be output as TTS 16 is applied to a synthesis system together with the emotional scores (GAMs - see below) 17. These are then mixed in the synthesis model 15 using TTS influence 18 and style embedding creation 19, which provides style embeddings to the TTS inference 18, in order to generate a synthetic speech (ie TTS output) with emotions. As described, this may have a mix of emotions and may have different degrees of each of emotion.

Turning now to the training phase. As described, a plurality of training samples are obtained, each labelled with a different emotion. From these, a GMM is fitted to the samples, resulting in one Gaussian component representing each emotion. Referring to Figure 3, the labelled emotional data samples 20, comprising emotional data 1, emotional data 2. emotional data J are applied to a style model 21. This results in a series of attention weights 20 for each emotion. Thus, there will be a first attention weight for emotion 1 (say sadness) another attention weight for emotion 2 (say happiness) and so on. These are then used to initialise a gaussian component for each component 23. After this, the GMMs are trained by maximum-likelihood using an expectation-maximisation algorithm. The procedure below describes an embodiment of how the GMM is estimated, by way of example only: 1. Initialize means by taking the average of the corresponding emotion dependent samples:

where

is the number of samples in emotional dataset j;

where

is a J-size identity matrix;

3. perform the E step of the expectation-maximization algorithm, by estimating the posterior probabilities,

4. perform the M step of the expectation-maximization algorithm by estimating the new component means, covariances and coefficients by making

5. repeat steps 3 and 4 several times until there is not much change in the log- likelihood cost below:

This results in a plurality of GMM components 24, one for each emotion. That is, a GMM component 1 , GMM component 2... GMM component J. In a preferred embodiment, there are seven emotions but different numbers and types of emotions may be used.

As the GMM components are initialised from the sample, each one is given one emotional label (sadness, happiness etc) so that each GMM component set 24 represents one emotion. The set of components is then applied during synthesis time.

At this time the actual text to be synthesised is analysed and the emotional content of the text is determined. This will typically comprise a plurality of emotions and with different degrees of "intensity" for each emotions. The amount of each emotion (from the relevant GMM component) in the text to be synthesised is determined. Typically, this results in a score between a zero value and maximum value, for example between 0 and 1 for each emotion for the text, where 0 represents a situation whether the text has none of that particular emotion and 1 where it is entirely of that emotion.

Thus, a passage which is entirely happy without any other emotion may have a score of 1 for the happiness emotion and a score of 0 for every other emotion. A text by a user who is angry but has no other emotion will have a score of 1 for the anger model, and 0 for the others. In practice, there will be a degree of each emotion, and thus a typical text may have a score of say 0.5 for happiness, 0.75 for anger, 0.24 for disgust, and so on, and thus represents and includes a range and plurality of emotions. By using all or some of the GMM components, one for each emotion, and combining these, the actual synthesis text can be provided with this range of emotions.

Referring more particularly to Figure 4, the emotional control model 24 (from the training stage) is shown comprising the GMM components, one for each emotion. These are then sampled 25 to provide an attention weight for each emotion (representing the degree of information that should be taken from the GST bank to be used in a particular text). After that, these attention weights are multiplied a corresponding emotional score. These emotional scores, that come from the frontend or the user, may be, for example, 0.2 for happiness, 0.8 for excitement, and so on. These are then combined at stage 28, a softmax process 29 is applied, and a CV 13 is generated which is used for the actual TTS output.

To train the GMM, in effect, all of the samples of one emotion (eg all the "happy samples") are used to calculate the mean vector of the "happy component". This is then used as the initial mean of the corresponding generated Gaussian component for the GMM component relating to happiness. The same is done for all the different emotions. Once this initial mean is set in the initialising step, then the GMM can be trained and its means are iteratively updated during the GMM training. They have already been labelled with a particular emotion and therefore this emotion is inherently linked to a particular GMM component 24. In Figure 4, the set of emotional scores (GMM components) are provided either by a user or by a front end of a system. The attention score vector is sampled 25 from each component of the trained GMM and these are then combined with the provided emotional scores to generate the synthetic CV 30 used for synthesis.

Once this synthetic CV 30 is obtained, then style embedding can be constructed using a process as shown in Figure 2 above and speech can be generated using a general process (Tacotron/ GTS) as shown in Figure 1 .

As is known, in neural network and machine learning technologies an embedding represents a vector that represents specific information. In the present invention, an embedding can represent a speaker, resulting in a set of vectors or embeddings in which each of them represents a specific speaker. Thus, the embeddings may represent styles.

Thus, in embodiments of the invention, emotion control can be divided into training and inference stages.. At training time emotional CVs are accumulated and a GMM fitted on them. At synthesis time CVs from each Gaussian component are sampled and mixed, based on the scores provided by the TTS front end.

Traininq

Starting from a fully trained tacotron and style model [1]], attention scores are collected from emotionally tagged data before being applied to the softmax layer, to obtain emotion-dependent attention scores,

, where j and n are respectively emotion and sample indices, and the scores prior to softmax are

where means the k-th attention score before softmax of the n-th sample of emotional dataset j. After that, a GMM is fitted on

, where J is the number of intended styles or emotions. To enforce emotion controllability each component mean m® is initializing by making

where is the number of samples in emotional dataset j. In order to enable interpretable emotion control at synthesis it is assumed that each component represents one emotion.

Synthesis

At synthesis time, first a set of emotional scores are provided by the user or

TTS frontend:

Then an attention score vector is sampled independently from each component of the trained GMM,

with S® being the covariance matrix of component j and

meaning a normal distribution. After that the frontend emotion scores,

are combined, with the sampled attention scores,

where become the mixed emotion scores. The adjusted CV

components are calculated as

and the final style embedding is given by

Claims

1. A text-to-speech method, comprising a training phrase in which the system is trained with a plurality of emotionally tagged data tagged with a plurality of different emotions which are applied to a GST model to estimate emotion- dependent style embedding; generating a Gaussian mixture model (GMM) on said emotion-dependent style embedding, one GMM component for each of the plurality of emotions, and, at the time of synthesis, sampling said emotion-dependent style embeddings from the GMM components, to obtain combined mixed emotion scores and applying these as an input for controlled speech synthesis

2. A method of claimed in claim 1 wherein one GMM component is generated for each one of the plurality of different emotions.

3. A method as claimed in claim 1 and claim 2 wherein the plurality of emotions are selected from anger, happiness, sadness, excitement, surprise, fear and disgust.

4. A method as claimed in any preceding claim wherein the training step comprises providing a first set of training data of relatively large amount and a second set of training data of relatively short amount compared to the first set of training data , each of which are tagged according to a predominant emotion.

5. A method as claimed in any preceding claim wherein, in the training phase, emotionally tagged data is applied to an attention weights estimation module and attention weights are generated for each emotion.

6. A method as claimed in claim 5 wherein the attention weight on each emotion is used to generate a GMM component for that emotion.

7. A method as claimed in claim 6 wherein the number of GMM components corresponds to the number of intended emotions.

8. A method as claimed in any preceding claim wherein during the training phase each GMM component is initialised with a particular emotion.

9. A method as claimed in any preceding claim wherein, at the time of synthesis, the text to be synthesised is analysed, and the emotional content of the text is determined.

10. A method as claimed in claim 9 wherein at the time of synthesis, a score is allocated to each GMM component representative of the amount of each emotion in the text, ranging from a minimum value, where it has none of that particular emotion, to a maximum value (where the text is wholly of that emotion).

11. A method as claimed in claim 10 wherein a value is given to each GMM component between a minimum and maximum value, inclusive, establishing the amount of that particular emotion in the text, to provide an emotion score for each GMM component and therefore each emotion, and using these to obtain a condition vector for synthesis comprising desired amounts of the respective emotions.

12. A method as claimed in claim 11 wherein the GMM components are sampled to provide an attention weight for each emotion.

13. A method as claimed in claim 12 wherein the attention weights are used to generate an emotional score which is then combined, and further, comprising applying a softmax process and generating a condition vector.

14. A method as claimed in any preceding claim, wherein, in the training phase, attention scores are collected from emotionally tagged data before being applied to a softmax layer, to obtain emotion-dependent attention scores,

where means the k-th attention score before softmax of the n-th sample of emotional dataset j. and wherein, a GMM is fitted on where J is the number of

intended styles or emotions. To enforce emotion controllability each component mean m® is initializing by making

where is the number of samples in emotional dataset j.

15. A method as claimed in any preceding claim wherein, during the synthesis stage, after a set of emotional scores are provided

an attention score vector is sampled independently from each component of the trained GMM,

with being the covariance matrix of component j and

meaning a normal distribution; the frontend emotion scores,

are combined, with the sampled attention scores,

where

become the mixed emotion scores; a set of adjusted CV components are calculated as

and a final style embedding is given by

16. A TTS system comprising apparatus configured to use the method of any of the preceding claims.