Axel Roebel

Giovanni Bindi, Nils Demerlé, Rodrigo Diaz et al. · bytedance

Artificial Intelligence (AI) technologies such as deep learning are evolving very quickly bringing many changes to our everyday lives. To explore the future impact and potential of AI in the field of music and sound technologies a doctoral day was held between Queen Mary University of London (QMUL, UK) and Sciences et Technologies de la Musique et du Son (STMS, France). Prompt questions about current trends in AI and music were generated by academics from QMUL and STMS. Students from the two institutions then debated these questions. This report presents a summary of the student debates on the topics of: Data, Impact, and the Environment; Responsible Innovation and Creative Practice; Creativity and Bias; and From Tools to the Singularity. The students represent the future generation of AI and music researchers. The academics represent the incumbent establishment. The student debates reported here capture visions, dreams, concerns, uncertainties, and contentious issues for the future of AI and music as the establishment is rightfully challenged by the next generation.

Théodor Lemerle, Nicolas Obin, Axel Roebel

Recent advancements in text-to-speech (TTS) powered by language models have showcased remarkable capabilities in achieving naturalness and zero-shot voice cloning. Notably, the decoder-only transformer is the prominent architecture in this domain. However, transformers face challenges stemming from their quadratic complexity in sequence length, impeding training on lengthy sequences and resource-constrained hardware. Moreover they lack specific inductive bias with regards to the monotonic nature of TTS alignments. In response, we propose to replace transformers with emerging recurrent architectures and introduce specialized cross-attention mechanisms for reducing repeating and skipping issues. Consequently our architecture can be efficiently trained on long samples and achieve state-of-the-art zero-shot voice cloning against baselines of comparable size. Our implementation and demos are available at https://github.com/theodorblackbird/lina-speech.

Théodor Lemerle, Téo Guichoux, Axel Roebel et al.

Neural codec language models, built on transformer architecture, have revolutionized text-to-speech (TTS) synthesis, excelling in voice cloning by treating it as a prefix continuation task. However, their limited context length hinders their effectiveness to short speech samples. As a result, the voice cloning ability is restricted to a limited coverage and diversity of the speaker's prosody and style. Besides, adapting prosody, accent, or appropriate emotion from a short prefix remains a challenging task. Finally, the quadratic complexity of self-attention limits inference throughput. In this work, we introduce Lina-Speech, a TTS model with Gated Linear Attention (GLA) to replace standard self-attention as a principled backbone, improving inference throughput while matching state-of-the-art performance. Leveraging the stateful property of recurrent architecture, we introduce an Initial-State Tuning (IST) strategy that unlocks the possibility of multiple speech sample conditioning of arbitrary numbers and lengths and provides a comprehensive and efficient strategy for voice cloning and out-of-domain speaking style and emotion adaptation. We demonstrate the effectiveness of this approach for controlling fine-grained characteristics such as prosody and emotion. Code, checkpoints, and demo are freely available: https://github.com/theodorblackbird/lina-speech

Daniel Wolff, Rémi Mignot, Axel Roebel

With increasing amounts of music being digitally transferred from production to distribution, automatic means of determining media quality are needed. Protection mechanisms in digital audio processing tools have not eliminated the need of production entities located downstream the distribution chain to assess audio quality and detect defects inserted further upstream. Such analysis often relies on the received audio and scarce meta-data alone. Deliberate use of artefacts such as clicks in popular music as well as more recent defects stemming from corruption in modern audio encodings call for data-centric and context sensitive solutions for detection. We present a convolutional network architecture following end-to-end encoder decoder configuration to develop detectors for two exemplary audio defects. A click detector is trained and compared to a traditional signal processing method, with a discussion on context sensitivity. Additional post-processing is used for data augmentation and workflow simulation. The ability of our models to capture variance is explored in a detector for artefacts from decompression of corrupted MP3 compressed audio. For both tasks we describe the synthetic generation of artefacts for controlled detector training and evaluation. We evaluate our detectors on the large open-source Free Music Archive (FMA) and genre-specific datasets.

Mathilde Abrassart, Nicolas Obin, Axel Roebel

Precise control over speech characteristics, such as pitch, duration, and speech rate, remains a significant challenge in the field of voice conversion. The ability to manipulate parameters like pitch and syllable rate is an important element for effective identity conversion, but can also be used independently for voice transformation, achieving goals that were historically addressed by vocoder-based methods. In this work, we explore a convolutional neural network-based approach that aims to provide means for modifying fundamental frequency (F0), phoneme sequences, intensity, and speaker identity. Rather than relying on disentanglement techniques, our model is explicitly conditioned on these factors to generate mel spectrograms, which are then converted into waveforms using a universal neural vocoder. Accordingly, during inference, F0 contours, phoneme sequences, and speaker embeddings can be freely adjusted, allowing for intuitively controlled voice transformations. We evaluate our approach on speaker conversion and expressive speech tasks using both perceptual and objective metrics. The results suggest that the proposed method offers substantial flexibility, while maintaining high intelligibility and speaker similarity.

Diego Torres, Axel Roebel, Nicolas Obin

We present PitchFlower, a flow-based neural audio codec with explicit pitch controllability. Our approach enforces disentanglement through a simple perturbation: during training, F0 contours are flattened and randomly shifted, while the true F0 is provided as conditioning. A vector-quantization bottleneck prevents pitch recovery, and a flow-based decoder generates high quality audio. Experiments show that PitchFlower achieves more accurate pitch control than WORLD at much higher audio quality, and outperforms SiFiGAN in controllability while maintaining comparable quality. Beyond pitch, this framework provides a simple and extensible path toward disentangling other speech attributes.

Frederik Bous, Laurent Benaroya, Nicolas Obin et al.

This paper introduces voice reenactement as the task of voice conversion (VC) in which the expressivity of the source speaker is preserved during conversion while the identity of a target speaker is transferred. To do so, an original neural- VC architecture is proposed based on sequence-to-sequence voice conversion (S2S-VC) in which the speech prosody of the source speaker is preserved during conversion. First, the S2S-VC architecture is modified so as to synchronize the converted speech with the source speech by mean of phonetic duration encoding; second, the decoder is conditioned on the desired sequence of F0- values and an explicit F0-loss is formulated between the F0 of the source speaker and the one of the converted speech. Besides, an adversarial learning of conversions is integrated within the S2S-VC architecture so as to exploit both advantages of reconstruction of original speech and converted speech with manipulated attributes during training and then reducing the inconsistency between training and conversion. An experimental evaluation on the VCTK speech database shows that the speech prosody can be efficiently preserved during conversion, and that the proposed adversarial learning consistently improves the conversion and the naturalness of the reenacted speech.

Axel Roebel, Frederik Bous

This paper introduces the Multi-Band Excited WaveNet a neural vocoder for speaking and singing voices. It aims to advance the state of the art towards an universal neural vocoder, which is a model that can generate voice signals from arbitrary mel spectrograms extracted from voice signals. Following the success of the DDSP model and following the development of the recently proposed excitation vocoders we propose a vocoder structure consisting of multiple specialized DNN that are combined with dedicated signal processing components. All components are implemented as differentiable operators and therefore allow joined optimization of the model parameters. To prove the capacity of the model to reproduce high quality voice signals we evaluate the model on single and multi speaker/singer datasets. We conduct a subjective evaluation demonstrating that the models support a wide range of domain variations (unseen voices, languages, expressivity) achieving perceptive quality that compares with a state of the art universal neural vocoder, however using significantly smaller training datasets and significantly less parameters. We also demonstrate remaining limits of the universality of neural vocoders e.g. the creation of saturated singing voices.

Laurent Benaroya, Nicolas Obin, Axel Roebel

Voice conversion (VC) consists of digitally altering the voice of an individual to manipulate part of its content, primarily its identity, while maintaining the rest unchanged. Research in neural VC has accomplished considerable breakthroughs with the capacity to falsify a voice identity using a small amount of data with a highly realistic rendering. This paper goes beyond voice identity and presents a neural architecture that allows the manipulation of voice attributes (e.g., gender and age). Leveraging the latest advances on adversarial learning of structured speech representation, a novel structured neural network is proposed in which multiple auto-encoders are used to encode speech as a set of idealistically independent linguistic and extra-linguistic representations, which are learned adversariarly and can be manipulated during VC. Moreover, the proposed architecture is time-synchronized so that the original voice timing is preserved during conversion which allows lip-sync applications. Applied to voice gender conversion on the real-world VCTK dataset, our proposed architecture can learn successfully gender-independent representation and convert the voice gender with a very high efficiency and naturalness.

Clément Le Moine, Nicolas Obin, Axel Roebel

Speech Emotion Recognition (SER) task has known significant improvements over the last years with the advent of Deep Neural Networks (DNNs). However, even the most successful methods are still rather failing when adaptation to specific speakers and scenarios is needed, inevitably leading to poorer performances when compared to humans. In this paper, we present novel work based on the idea of teaching the emotion recognition network about speaker identity. Our system is a combination of two ACRNN classifiers respectively dedicated to speaker and emotion recognition. The first informs the latter through a Self Speaker Attention (SSA) mechanism that is shown to considerably help to focus on emotional information of the speech signal. Experiments on social attitudes database Att-HACK and IEMOCAP corpus demonstrate the effectiveness of the proposed method and achieve the state-of-the-art performance in terms of unweighted average recall.

Clément Le Moine Veillon, Nicolas Obin, Axel Roebel

This paper presents a end-to-end framework for the F0 transformation in the context of expressive voice conversion. A single neural network is proposed, in which a first module is used to learn F0 representation over different temporal scales and a second adversarial module is used to learn the transformation from one emotion to another. The first module is composed of a convolution layer with wavelet kernels so that the various temporal scales of F0 variations can be efficiently encoded. The single decomposition/transformation network allows to learn in a end-to-end manner the F0 decomposition that are optimal with respect to the transformation, directly from the raw F0 signal.

Frederik Bous, Luc Ardaillon, Axel Roebel

This article investigates into recently emerging approaches that use deep neural networks for the estimation of glottal closure instants (GCI). We build upon our previous approach that used synthetic speech exclusively to create perfectly annotated training data and that had been shown to compare favourably with other training approaches using electroglottograph (EGG) signals. Here we introduce a semi-supervised training strategy that allows refining the estimator by means of an analysis-synthesis setup using real speech signals, for which GCI ground truth does not exist. Evaluation of the analyser is performed by means of comparing the GCI extracted from the glottal flow signal generated by the analyser with the GCI extracted from EGG on the CMU arctic dataset, where EGG signals were recorded in addition to speech. We observe that (1.) the artificial increase of the diversity of pulse shapes that has been used in our previous construction of the synthetic database is beneficial, (2.) training the GCI network in the analysis-synthesis setup allows achieving a very significant improvement of the GCI analyser, (3.) additional regularisation strategies allow improving the final analysis network when trained in the analysis-synthesis setup.

Rafael Ferro, Nicolas Obin, Axel Roebel

This paper tackles GAN optimization and stability issues in the context of voice conversion. First, to simplify the conversion task, we propose to use spectral envelopes as inputs. Second we propose two adversarial weight training paradigms, the generalized weighted GAN and the generator impact GAN, both aim at reducing the impact of the generator on the discriminator, so both can learn more gradually and efficiently during training. Applying an energy constraint to the cycleGAN paradigm considerably improved conversion quality. A subjective experiment conducted on a voice conversion task on the voice conversion challenge 2018 dataset shows first that despite a significantly reduced network complexity, the proposed method achieves state-of-the-art results, and second that the proposed weighted GAN methods outperform a previously proposed one.

Hugo Caracalla, Axel Roebel

This article introduces a new parametric synthesis method for sound textures based on existing works in visual and sound texture synthesis. Starting from a base sound signal, an optimization process is performed until the cross-correlations between the feature-maps of several untrained 2D Convolutional Neural Networks (CNN) resemble those of an original sound texture. We use compressed RI spectrograms as input to the CNN: this time-frequency representation is the stacking of the real and imaginary part of the Short Time Fourier Transform (STFT) and thus implicitly contains both the magnitude and phase information, allowing for convincing syntheses of various audio events. The optimization is however performed directly on the time signal to avoid any STFT consistency issue. The results of an online perceptual evaluation are also detailed, and show that this method achieves results that are more realistic-sounding than existing parametric methods on a wide array of textures.

Hugo Caracalla, Axel Roebel

The following article introduces a new parametric synthesis algorithm for sound textures inspired by existing methods used for visual textures. Using a 2D Convolutional Neural Network (CNN), a sound signal is modified until the temporal cross-correlations of the feature maps of its log-spectrogram resemble those of a target texture. We show that the resulting synthesized sound signal is both different from the original and of high quality, while being able to reproduce singular events appearing in the original. This process is performed in the time domain, discarding the harmful phase recovery step which usually concludes synthesis performed in the time-frequency domain. It is also straightforward and flexible, as it does not require any fine tuning between several losses when synthesizing diverse sound textures. A way of extending the synthesis in order to produce a sound of any length is also presented, after which synthesized spectrograms and sound signals are showcased. We also discuss on the choice of CNN, on border effects in our synthesized signals and on possible ways of modifying the algorithm in order to improve its current long computation time.

Celine Jacques, Axel Roebel

A recurrent issue in deep learning is the scarcity of data, in particular precisely annotated data. Few publicly available databases are correctly annotated and generating correct labels is very time consuming. The present article investigates into data augmentation strategies for Neural Networks training, particularly for tasks related to drum transcription. These tasks need very precise annotations. This article investigates state-of-the-art sound transformation algorithms for remixing noise and sinusoidal parts, remixing attacks, transposing with and without time compensation and compares them to basic regularization methods such as using dropout and additive Gaussian noise. And it shows how a drum transcription algorithm based on CNN benefits from the proposed data augmentation strategy.

Alice Cohen-Hadria, Axel Roebel, Geoffroy Peeters

State-of-the-art singing voice separation is based on deep learning making use of CNN structures with skip connections (like U-net model, Wave-U-Net model, or MSDENSELSTM). A key to the success of these models is the availability of a large amount of training data. In the following study, we are interested in singing voice separation for mono signals and will investigate into comparing the U-Net and the Wave-U-Net that are structurally similar, but work on different input representations. First, we report a few results on variations of the U-Net model. Second, we will discuss the potential of state of the art speech and music transformation algorithms for augmentation of existing data sets and demonstrate that the effect of these augmentations depends on the signal representations used by the model. The results demonstrate a considerable improvement due to the augmentation for both models. But pitch transposition is the most effective augmentation strategy for the U-Net model, while transposition, time stretching, and formant shifting have a much more balanced effect on the Wave-U-Net model. Finally, we compare the two models on the same dataset.

Frederik Bous, Axel Roebel

We conduct an investigation on various hyper-parameters regarding neural networks used to generate spectral envelopes for singing synthesis. Two perceptive tests, where the first compares two models directly and the other ranks models with a mean opinion score, are performed. With these tests we show that when learning to predict spectral envelopes, 2d-convolutions are superior over previously proposed 1d-convolutions and that predicting multiple frames in an iterated fashion during training is superior over injecting noise to the input data. An experimental investigation whether learning to predict a probability distribution vs.\ single samples was performed but turned out to be inconclusive. A network architecture is proposed that incorporates the improvements which we found to be useful and we show in our experiments that this network produces better results than other stat-of-the-art methods.

Mathieu Lagrange, Grégoire Lafay, Mathias Rossignol et al.

This paper introduces a model of environmental acoustic scenes which adopts a morphological approach by ab-stracting temporal structures of acoustic scenes. To demonstrate its potential, this model is employed to evaluate the performance of a large set of acoustic events detection systems. This model allows us to explicitly control key morphological aspects of the acoustic scene and isolate their impact on the performance of the system under evaluation. Thus, more information can be gained on the behavior of evaluated systems, providing guidance for further improvements. The proposed model is validated using submitted systems from the IEEE DCASE Challenge; results indicate that the proposed scheme is able to successfully build datasets useful for evaluating some aspects the performance of event detection systems, more particularly their robustness to new listening conditions and the increasing level of background sounds.