Most of the prevalent approaches in speech prosody modeling rely on learning global style representations or a continuous latent space which encode and transfer the attributes of reference speech. However, recent work on neural codecs which are based on Residual Vector Quantization (RVQ) already shows great potential offering distinct advantages. We investigate the prosody modeling capabilities of the discrete space of such an RVQ-VAE model, modifying it to operate on the phoneme-level. We condition both the encoder and decoder of the model on linguistic representations and apply a global speaker embedding in order to factor out both phonetic and speaker information. We conduct an extensive set of investigations based on subjective experiments and objective measures to show that the phoneme-level discrete latent representations obtained this way achieve a high degree of disentanglement, capturing fine-grained prosodic information that is robust and transferable. The latent space turns out to have interpretable structure with its principal components corresponding to pitch and energy.

Current state-of-the-art methods for automatic synthetic speech evaluation are based on MOS prediction neural models. Such MOS prediction models include MOSNet and LDNet that use spectral features as input, and SSL-MOS that relies on a pretrained selfsupervised learning model that directly uses the speech signal as input. In modern high-quality neural TTS systems, prosodic appropriateness with regard to the spoken content is a decisive factor for speech naturalness. For this reason, we propose to include prosodic and linguistic features as additional inputs in MOS prediction systems, and evaluate their impact on the prediction outcome. We consider phoneme-level F0 and duration features as prosodic inputs, as well as Tacotron encoder outputs, POS tags and BERT embeddings as higher-level linguistic inputs. All MOS prediction systems are trained on SOMOS, a neural TTS-only dataset with crowdsourced naturalness MOS evaluations. Results show that the proposed additional features are beneficial in the MOS prediction task, by improving the predicted MOS scores’ correlation with the ground truths, both at utterance-level and system-level predictions.

Existing singing voice synthesis models (SVS) are usually trained on singing data and depend on either error-prone time-alignment and duration features or explicit music score information. In this paper, we propose Karaoker, a multispeaker Tacotron-based model conditioned on voice characteristic features that is trained exclusively on spoken data without requiring time-alignments. Karaoker synthesizes singing voice and transfers style following a multi-dimensional template extracted from a source waveform of an unseen singer/speaker. The model is jointly conditioned with a single deep convolutional encoder on continuous data including pitch, intensity, harmonicity, formants, cepstral peak prominence and octaves. We extend the text-to-speech training objective with feature reconstruction, classification and speaker identification tasks that guide the model to an accurate result. In addition to multitasking, we also employ a Wasserstein GAN training scheme as well as new losses on the acoustic model's output to further refine the quality of the model.

Voice cloning is a difficult task which requires robust and informative features incorporated in a high quality TTS system in order to effectively copy an unseen speaker's voice.In our work, we utilize features learned in a self-supervised framework via the Bootstrap Your Own Latent (BYOL) method, which is shown to produce high quality speech representations when specific audio augmentations are applied to the vanilla algorithm.We further extend the augmentations in the training procedure to aid the resulting features to capture the speaker identity and to make them robust to noise and acoustic conditions.The learned features are used as pre-trained utterance-level embeddings and as inputs to a Non-Attentive Tacotron based architecture, aiming to achieve multispeaker speech synthesis without utilizing additional speaker features.This method enables us to train our model in an unlabeled multispeaker dataset as well as use unseen speaker embeddings to copy a speaker's voice.Subjective and objective evaluations are used to validate the proposed model, as well as the robustness to the acoustic conditions of the target utterance.

In this work, we present the SOMOS dataset, the first large-scale mean opinion scores (MOS) dataset consisting of solely neural text-to-speech (TTS) samples. It can be employed to train automatic MOS prediction systems focused on the assessment of modern synthesizers, and can stimulate advancements in acoustic model evaluation. It consists of 20K synthetic utterances of the LJ Speech voice, a public domain speech dataset which is a common benchmark for building neural acoustic models and vocoders. Utterances are generated from 200 TTS systems including vanilla neural acoustic models as well as models which allow prosodic variations. An LPCNet vocoder is used for all systems, so that the samples' variation depends only on the acoustic models. The synthesized utterances provide balanced and adequate domain and length coverage. We collect MOS naturalness evaluations on 3 English Amazon Mechanical Turk locales and share practices leading to reliable crowdsourced annotations for this task. We provide baseline results of state-of-the-art MOS prediction models on the SOMOS dataset and show the limitations that such models face when assigned to evaluate TTS utterances.

A text-to-speech (TTS) model typically factorizes speech attributes such as content, speaker and prosody into disentangled representations.Recent works aim to additionally model the acoustic conditions explicitly, in order to disentangle the primary speech factors, i.e. linguistic content, prosody and timbre from any residual factors, such as recording conditions and background noise.This paper proposes unsupervised, interpretable and fine-grained noise and prosody modeling.We incorporate adversarial training, representation bottleneck and utterance-to-frame modeling in order to learn frame-level noise representations.To the same end, we perform fine-grained prosody modeling via a Fully Hierarchical Variational AutoEncoder (FVAE) which additionally results in more expressive speech synthesis.Experimental results support our claims and ablation studies verify the importance of each proposed component.Audio samples are available in our demo page 1 .

This paper presents a method for controlling the prosody at the phoneme level in an autoregressive attention-based text-to-speech system. Instead of learning latent prosodic features with a variational framework as is commonly done, we directly extract phoneme-level F0 and duration features from the speech data in the training set. Each prosodic feature is discretized using unsupervised clustering in order to produce a sequence of prosodic labels for each utterance. This sequence is used in parallel to the phoneme sequence in order to condition the decoder with the utilization of a prosodic encoder and a corresponding attention module. Experimental results show that the proposed method retains the high quality of generated speech, while allowing phoneme-level control of F0 and duration. By replacing the F0 cluster centroids with musical notes, the model can also provide control over the note and octave within the range of the speaker.

The idea of using phonological features instead of phonemes as input to sequence-to-sequence TTS has been recently proposed for zero-shot multilingual speech synthesis. This approach is useful for code-switching, as it facilitates the seamless uttering of foreign text embedded in a stream of native text. In our work, we train a language-agnostic multispeaker model conditioned on a set of phonologically derived features common across different languages, with the goal of achieving cross-lingual speaker adaptation. We first experiment with the effect of language phonological similarity on cross-lingual TTS of several source-target language combinations. Subsequently, we fine-tune the model with very limited data of a new speaker's voice in either a seen or an unseen language, and achieve synthetic speech of equal quality, while preserving the target speaker's identity. With as few as 32 and 8 utterances of target speaker data, we obtain high speaker similarity scores and naturalness comparable to the corresponding literature. In the extreme case of only 2 available adaptation utterances, we find that our model behaves as a few-shot learner, as the performance is similar in both the seen and unseen adaptation language scenarios.

This paper presents an end-to-end text-to-speech system with low latency on a CPU, suitable for real-time applications. The system is composed of an autoregressive attention-based sequence-to-sequence acoustic model and the LPCNet vocoder for waveform generation. An acoustic model architecture that adopts modules from both the Tacotron 1 and 2 models is proposed, while stability is ensured by using a recently proposed purely location-based attention mechanism, suitable for arbitrary sentence length generation. During inference, the decoder is unrolled and acoustic feature generation is performed in a streaming manner, allowing for a nearly constant latency which is independent from the sentence length. Experimental results show that the acoustic model can produce feature sequences with minimal latency about 31 times faster than real-time on a computer CPU and 6.5 times on a mobile CPU, enabling it to meet the conditions required for real-time applications on both devices. The full end-to-end system can generate almost natural quality speech, which is verified by listening tests.

We examine the use of linear and non-linear dimensionality reduction algorithms for extracting low-rank feature representations for speech emotion recognition. Two feature sets are used, one based on low-level descriptors and their aggregations (IS10) and one modeling recurrence dynamics of speech (RQA), as well as their fusion. We report speech emotion recognition (SER) results for learned representations on two databases using different classification methods. Classification with low-dimensional representations yields performance improvement in a variety of settings. This indicates that dimensionality reduction is an effective way to combat the curse of dimensionality for SER. Visualization of features in two dimensions provides insight into discriminatory abilities of reduced feature sets.