Efficient training strategies for natural sounding speech synthesis and speaker adaptation based on FastPitch

This paper focuses on adapting the functionalities of the FastPitch model to the Romanian language; extending the set of speakers from one to eighteen; synthesising speech using an anonymous identity; and replicating the identities of new, unseen speakers. During this work, the effects of various configurations and training strategies were tested and discussed, along with their advantages and weaknesses. Finally, we settled on a new configuration, built on top of the FastPitch architecture, capable of producing natural speech synthesis, for both known (identities from the training dataset) and unknown (identities learnt through short reference samples) speakers. The anonymous speaker can be used for text-to-speech synthesis, if one wants to cancel out the identity information while keeping the semantic content whole and clear. At last, we discussed possible limitations of our work, which will form the basis for future investigations and advancements.

TBDM-Net: Bidirectional Dense Networks with Gender Information for Speech Emotion Recognition

This paper presents a novel deep neural network-based architecture tailored for Speech Emotion Recognition (SER). The architecture capitalises on dense interconnections among multiple layers of bidirectional dilated convolutions. A linear kernel dynamically fuses the outputs of these layers to yield the final emotion class prediction. This innovative architecture is denoted as TBDM-Net: Temporally-Aware Bi-directional Dense Multi-Scale Network. We conduct a comprehensive performance evaluation of TBDM-Net, including an ablation study, across six widely-acknowledged SER datasets for unimodal speech emotion recognition. Additionally, we explore the influence of gender-informed emotion prediction by appending either golden or predicted gender labels to the architecture's inputs or predictions. The implementation of TBDM-Net is accessible at: https://github.com/adrianastan/tbdm-net

WavLM model ensemble for audio deepfake detection

Audio deepfake detection has become a pivotal task over the last couple of years, as many recent speech synthesis and voice cloning systems generate highly realistic speech samples, thus enabling their use in malicious activities. In this paper we address the issue of audio deepfake detection as it was set in the ASVspoof5 challenge. First, we benchmark ten types of pretrained representations and show that the self-supervised representations stemming from the wav2vec2 and wavLM families perform best. Of the two, wavLM is better when restricting the pretraining data to LibriSpeech, as required by the challenge rules. To further improve performance, we finetune the wavLM model for the deepfake detection task. We extend the ASVspoof5 dataset with samples from other deepfake detection datasets and apply data augmentation. Our final challenge submission consists of a late fusion combination of four models and achieves an equal error rate of 6.56% and 17.08% on the two evaluation sets.

An analysis of large speech models-based representations for speech emotion recognition

Large speech models-derived features have recently shown increased performance over signal-based features across multiple downstream tasks, even when the networks are not finetuned towards the target task. In this paper we show the results of an analysis of several signal- and neural models-derived features for speech emotion recognition. We use pretrained models and explore their inherent potential abstractions of emotions. Simple classification methods are used so as to not interfere or add knowledge to the task. We show that, even without finetuning, some of these large neural speech models' representations can enclose information that enables performances close to, and even beyond state-of-the-art results across six standard speech emotion recognition datasets.

Towards generalisable and calibrated synthetic speech detection with self-supervised representations

Generalisation -- the ability of a model to perform well on unseen data -- is crucial for building reliable deep fake detectors. However, recent studies have shown that the current audio deep fake models fall short of this desideratum. In this paper we show that pretrained self-supervised representations followed by a simple logistic regression classifier achieve strong generalisation capabilities, reducing the equal error rate from 30% to 8% on the newly introduced In-the-Wild dataset. Importantly, this approach also produces considerably better calibrated models when compared to previous approaches. This means that we can trust our model's predictions more and use these for downstream tasks, such as uncertainty estimation. In particular, we show that the entropy of the estimated probabilities provides a reliable way of rejecting uncertain samples and further improving the accuracy.

An analysis on the effects of speaker embedding choice in non auto-regressive TTS

In this paper we introduce a first attempt on understanding how a non-autoregressive factorised multi-speaker speech synthesis architecture exploits the information present in different speaker embedding sets. We analyse if jointly learning the representations, and initialising them from pretrained models determine any quality improvements for target speaker identities. In a separate analysis, we investigate how the different sets of embeddings impact the network's core speech abstraction (i.e. zero conditioned) in terms of speaker identity and representation learning. We show that, regardless of the used set of embeddings and learning strategy, the network can handle various speaker identities equally well, with barely noticeable variations in speech output quality, and that speaker leakage within the core structure of the synthesis system is inevitable in the standard training procedures adopted thus far.

Residual Information in Deep Speaker Embedding Architectures

Speaker embeddings represent a means to extract representative vectorial representations from a speech signal such that the representation pertains to the speaker identity alone. The embeddings are commonly used to classify and discriminate between different speakers. However, there is no objective measure to evaluate the ability of a speaker embedding to disentangle the speaker identity from the other speech characteristics. This means that the embeddings are far from ideal, highly dependent on the training corpus and still include a degree of residual information pertaining to factors such as linguistic content, recording conditions or speaking style of the utterance. This paper introduces an analysis over six sets of speaker embeddings extracted with some of the most recent and high-performing DNN architectures, and in particular, the degree to which they are able to truly disentangle the speaker identity from the speech signal. To correctly evaluate the architectures, a large multi-speaker parallel speech dataset is used. The dataset includes 46 speakers uttering the same set of prompts, recorded in either a professional studio or their home environments. The analysis looks into the intra- and inter-speaker similarity measures computed over the different embedding sets, as well as if simple classification and regression methods are able to extract several residual information factors from the speaker embeddings. The results show that the discriminative power of the analyzed embeddings is very high, yet across all the analyzed architectures, residual information is still present in the representations in the form of a high correlation to the recording conditions, linguistic contents and utterance duration.

The ZevoMOS entry to VoiceMOS Challenge 2022

This paper introduces the ZevoMOS entry to the main track of the VoiceMOS Challenge 2022. The ZevoMOS submission is based on a two-step finetuning of pretrained self-supervised learning (SSL) speech models. The first step uses a task of classifying natural versus synthetic speech, while the second step's task is to predict the MOS scores associated with each training sample. The results of the finetuning process are then combined with the confidence scores extracted from an automatic speech recognition model, as well as the raw embeddings of the training samples obtained from a wav2vec SSL speech model. The team id assigned to the ZevoMOS system within the VoiceMOS Challenge is T01. The submission was placed on the 14th place with respect to the system-level SRCC, and on the 9th place with respect to the utterance-level MSE. The paper also introduces additional evaluations of the intermediate results.

FlexLip: A Controllable Text-to-Lip System

The task of converting text input into video content is becoming an important topic for synthetic media generation. Several methods have been proposed with some of them reaching close-to-natural performances in constrained tasks. In this paper, we tackle a subissue of the text-to-video generation problem, by converting the text into lip landmarks. However, we do this using a modular, controllable system architecture and evaluate each of its individual components. Our system, entitled FlexLip, is split into two separate modules: text-to-speech and speech-to-lip, both having underlying controllable deep neural network architectures. This modularity enables the easy replacement of each of its components, while also ensuring the fast adaptation to new speaker identities by disentangling or projecting the input features. We show that by using as little as 20 min of data for the audio generation component, and as little as 5 min for the speech-to-lip component, the objective measures of the generated lip landmarks are comparable with those obtained when using a larger set of training samples. We also introduce a series of objective evaluation measures over the complete flow of our system by taking into consideration several aspects of the data and system configuration. These aspects pertain to the quality and amount of training data, the use of pretrained models, and the data contained therein, as well as the identity of the target speaker; with regard to the latter, we show that we can perform zero-shot lip adaptation to an unseen identity by simply updating the shape of the lips in our model.

An objective evaluation of the effects of recording conditions and speaker characteristics in multi-speaker deep neural speech synthesis

Multi-speaker spoken datasets enable the creation of text-to-speech synthesis (TTS) systems which can output several voice identities. The multi-speaker (MSPK) scenario also enables the use of fewer training samples per speaker. However, in the resulting acoustic model, not all speakers exhibit the same synthetic quality, and some of the voice identities cannot be used at all. In this paper we evaluate the influence of the recording conditions, speaker gender, and speaker particularities over the quality of the synthesised output of a deep neural TTS architecture, namely Tacotron2. The evaluation is possible due to the use of a large Romanian parallel spoken corpus containing over 81 hours of data. Within this setup, we also evaluate the influence of different types of text representations: orthographic, phonetic, and phonetic extended with syllable boundaries and lexical stress markings. We evaluate the results of the MSPK system using the objective measures of equal error rate (EER) and word error rate (WER), and also look into the distances between natural and synthesised t-SNE projections of the embeddings computed by an accurate speaker verification network. The results show that there is indeed a large correlation between the recording conditions and the speaker's synthetic voice quality. The speaker gender does not influence the output, and that extending the input text representation with syllable boundaries and lexical stress information does not equally enhance the generated audio across all speaker identities. The visualisation of the t-SNE projections of the natural and synthesised speaker embeddings show that the acoustic model shifts some of the speakers' neural representation, but not all of them. As a result, these speakers have lower performances of the output speech.