The objective of this study is to generate high-quality speech from silent talking face videos, a task also known as video-to-speech synthesis. A significant challenge in video-to-speech synthesis lies in the substantial modality gap between silent video and multi-faceted speech. In this paper, we propose a novel video-to-speech system that effectively bridges this modality gap, significantly enhancing the quality of synthesized speech. This is achieved by learning of hierarchical representations from video to speech. Specifically, we gradually transform silent video into acoustic feature spaces through three sequential stages -- content, timbre, and prosody modeling. In each stage, we align visual factors -- lip movements, face identity, and facial expressions -- with corresponding acoustic counterparts to ensure the seamless transformation. Additionally, to generate realistic and coherent speech from the visual representations, we employ a flow matching model that estimates direct trajectories from a simple prior distribution to the target speech distribution. Extensive experiments demonstrate that our method achieves exceptional generation quality comparable to real utterances, outperforming existing methods by a significant margin.

The goal of voice conversion is to transform the speech of a source speaker to sound like that of a reference speaker while preserving the original content. A key challenge is to extract disentangled linguistic content from the source and voice style from the reference. While existing approaches leverage various methods to isolate the two, a generalization still requires further attention, especially for robustness in zero-shot scenarios. In this paper, we achieve successful disentanglement of content and speaker features by tuning self-supervised speech features with adapters. The adapters are trained to dynamically encode nuanced features from rich self-supervised features, and the decoder fuses them to produce speech that accurately resembles the reference with minimal loss of content. Moreover, we leverage a conditional flow matching decoder with cross-attention speaker conditioning to further boost the synthesis quality and efficiency. Subjective and objective evaluations in a zero-shot scenario demonstrate that the proposed method outperforms existing models in speech quality and similarity to the reference speech.

The goal of this paper is to generate realistic audio with a lightweight and fast diffusion-based vocoder, named FreGrad. Our framework consists of the following three key components: (1) We employ discrete wavelet transform that decomposes a complicated waveform into sub-band wavelets, which helps FreGrad to operate on a simple and concise feature space, (2) We design a frequency-aware dilated convolution that elevates frequency awareness, resulting in generating speech with accurate frequency information, and (3) We introduce a bag of tricks that boosts the generation quality of the proposed model. In our experiments, FreGrad achieves 3.7 times faster training time and 2.2 times faster inference speed compared to our baseline while reducing the model size by 0.6 times (only 1.78M parameters) without sacrificing the output quality. Audio samples are available at: https://mm.kaist.ac.kr/projects/FreGrad.

This paper addresses the problem of global tempo estimation in musical audio. Given that annotating tempo is time-consuming and requires certain musical expertise, few publicly available data sources exist to train machine learning models for this task. Towards alleviating this issue, we propose a fully self-supervised approach that does not rely on any human labeled data. Our method builds on the fact that generic (music) audio embeddings already encode a variety of properties, including information about tempo, making them easily adaptable for downstream tasks. While recent work in self-supervised tempo estimation aimed to learn a tempo specific representation that was subsequently used to train a supervised classifier, we reformulate the task into the binary classification problem of predicting whether a target track has the same or a different tempo compared to a reference. While the former still requires labeled training data for the final classification model, our approach uses arbitrary unlabeled music data in combination with time-stretching for model training as well as a small set of synthetically created reference samples for predicting the final tempo. Evaluation of our approach in comparison with the state-of-the-art reveals highly competitive performance when the constraint of finding the precise tempo octave is relaxed.

Audio embeddings enable large scale comparisons of the similarity of audio files for applications such as search and recommendation. Due to the subjectivity of audio similarity, it can be desirable to design systems that answer not only whether audio is similar, but similar in what way (e.g., wrt. tempo, mood or genre). Previous works have proposed disentangled embedding spaces where subspaces representing specific, yet possibly correlated, attributes can be weighted to emphasize those attributes in downstream tasks. However, no research has been conducted into the independence of these subspaces, nor their manipulation, in order to retrieve tracks that are similar but different in a specific way. Here, we explore the manipulation of tempo in embedding spaces as a case-study towards this goal. We propose tempo translation functions that allow for efficient manipulation of tempo within a pre-existing embedding space whilst maintaining other properties such as genre. As this translation is specific to tempo it enables retrieval of tracks that are similar but have specifically different tempi. We show that such a function can be used as an efficient data augmentation strategy for both training of downstream tempo predictors, and improved nearest neighbor retrieval of properties largely independent of tempo.

Audio embeddings are crucial tools in understanding large catalogs of music. Typically embeddings are evaluated on the basis of the performance they provide in a wide range of downstream tasks, however few studies have investigated the local properties of the embedding spaces themselves which are important in nearest neighbor algorithms, commonly used in music search and recommendation. In this work we show that when learning audio representations on music datasets via contrastive learning, musical properties that are typically homogeneous within a track (e.g., key and tempo) are reflected in the locality of neighborhoods in the resulting embedding space. By applying appropriate data augmentation strategies, localisation of such properties can not only be reduced but the localisation of other attributes is increased. For example, locality of features such as pitch and tempo that are less relevant to non-expert listeners, may be mitigated while improving the locality of more salient features such as genre and mood, achieving state-of-the-art performance in nearest neighbor retrieval accuracy. Similarly, we show that the optimal selection of data augmentation strategies for contrastive learning of music audio embeddings is dependent on the downstream task, highlighting this as an important embedding design decision.

The goal of this work is to reconstruct high quality speech from lip motions alone, a task also known as lip-to-speech. A key challenge of lip-to-speech systems is the one-to-many mapping caused by (1) the existence of homophenes and (2) multiple speech variations, resulting in a mispronounced and over-smoothed speech. In this paper, we propose a novel lip-to-speech system that significantly improves the generation quality by alleviating the one-to-many mapping problem from multiple perspectives. Specifically, we incorporate (1) self-supervised speech representations to disambiguate homophenes, and (2) acoustic variance information to model diverse speech styles. Additionally, to better solve the aforementioned problem, we employ a flow based post-net which captures and refines the details of the generated speech. We perform extensive experiments on two datasets, and demonstrate that our method achieves the generation quality close to that of real human utterance, outperforming existing methods in terms of speech naturalness and intelligibility by a large margin. Synthesised samples are available at our demo page: https://mm.kaist.ac.kr/projects/LTBS.

-- This paper proposes a framework which is able to generate a sequence of three-dimensional human dance poses for a given music. The proposed framework consists of three components: a music feature encoder, a pose generator, and a music genre classifier . We focus on integrating these components for generating a realistic 3D human dancing move from music, which can be applied to artificial agents and humanoid robots. The trained dance pose generator, which is a generative autoregressive model, is able to synthesize a dance sequence longer than 5,000 pose frames. Experimental results of generated dance sequences from various songs show how the proposed method generates humanlike dancing move to a given music. In addition, a generated 3D dance sequence is applied to a humanoid robot, showing that the proposed framework can make a robot to dance just by listening to music. Dance is one of the most important form of performing arts that having been emerged in all known cultures. As one of the specific subcategory of under theatrical dance, choreography associated with music is also one of the most popular forms that have usually been designed and physically performed by professional choreographers.

Deep neural networks have frequently been used to directly learn representations useful for a given task from raw input data. In terms of overall performance metrics, machine learning solutions employing deep representations frequently have been reported to greatly outperform those using hand-crafted feature representations. At the same time, they may pick up on aspects that are predominant in the data, yet not actually meaningful or interpretable. In this paper, we therefore propose a systematic way to diagnose the trustworthiness of deep music representations, considering musical semantics. The underlying assumption is that in case a deep representation is to be trusted, distance consistency between known related points should be maintained both in the input audio space and corresponding latent deep space. We generate known related points through semantically meaningful transformations, both considering imperceptible and graver transformations. Then, we examine within- and between-space distance consistencies, both considering audio space and latent embedded space, the latter either being a result of a conventional feature extractor or a deep encoder. We illustrate how our method, as a complement to task-specific performance, provides interpretable insight into what a network may have captured from training data signals.

Existing high-performance deep learning models require very intensive computing. For this reason, it is difficult to embed a deep learning model into a system with limited resources. In this paper, we propose the novel idea of the network compression as a method to solve this limitation. The principle of this idea is to make iterative pruning more effective and sophisticated by simulating the reduced network. A simple experiment was conducted to evaluate the method; the results showed that the proposed method achieved higher performance than existing methods at the same pruning level.