Darius Petermann

Pooneh Mousavi, Gallil Maimon, Adel Moumen et al.

Discrete audio tokens are compact representations that aim to preserve perceptual quality, phonetic content, and speaker characteristics while enabling efficient storage and inference, as well as competitive performance across diverse downstream tasks. They provide a practical alternative to continuous features, enabling the integration of speech and audio into modern large language models (LLMs). As interest in token-based audio processing grows, various tokenization methods have emerged, and several surveys have reviewed the latest progress in the field. However, existing studies often focus on specific domains or tasks and lack a unified comparison across various benchmarks. This paper presents a systematic review and benchmark of discrete audio tokenizers, covering three domains: speech, music, and general audio. We propose a taxonomy of tokenization approaches based on encoder-decoder, quantization techniques, training paradigm, streamability, and application domains. We evaluate tokenizers on multiple benchmarks for reconstruction, downstream performance, and acoustic language modeling, and analyze trade-offs through controlled ablation studies. Our findings highlight key limitations, practical considerations, and open challenges, providing insight and guidance for future research in this rapidly evolving area. For more information, including our main results and tokenizer database, please refer to our website: https://poonehmousavi.github.io/dates-website/.

Darius Petermann, Mahdi M. Kalayeh

Training audio-to-image generative models requires an abundance of diverse audio-visual pairs that are semantically aligned. Such data is almost always curated from in-the-wild videos, given the cross-modal semantic correspondence that is inherent to them. In this work, we hypothesize that insisting on the absolute need for ground truth audio-visual correspondence, is not only unnecessary, but also leads to severe restrictions in scale, quality, and diversity of the data, ultimately impairing its use in the modern generative models. That is, we propose a scalable image sonification framework where instances from a variety of high-quality yet disjoint uni-modal origins can be artificially paired through a retrieval process that is empowered by reasoning capabilities of modern vision-language models. To demonstrate the efficacy of this approach, we use our sonified images to train an audio-to-image generative model that performs competitively against state-of-the-art. Finally, through a series of ablation studies, we exhibit several intriguing auditory capabilities like semantic mixing and interpolation, loudness calibration and acoustic space modeling through reverberation that our model has implicitly developed to guide the image generation process.

Darius Petermann, Minje Kim

With the recent advancements of data driven approaches using deep neural networks, music source separation has been formulated as an instrument-specific supervised problem. While existing deep learning models implicitly absorb the spatial information conveyed by the multi-channel input signals, we argue that a more explicit and active use of spatial information could not only improve the separation process but also provide an entry-point for many user-interaction based tools. To this end, we introduce a control method based on the stereophonic location of the sources of interest, expressed as the panning angle. We present various conditioning mechanisms, including the use of raw angle and its derived feature representations, and show that spatial information helps. Our proposed approaches improve the separation performance compared to location agnostic architectures by 1.8 dB SI-SDR in our Slakh-based simulated experiments. Furthermore, the proposed methods allow for the disentanglement of same-class instruments, for example, in mixtures containing two guitar tracks. Finally, we also demonstrate that our approach is robust to incorrect source panning information, which can be incurred by our proposed user interaction.

Darius Petermann, Gordon Wichern, Zhong-Qiu Wang et al.

The cocktail party problem aims at isolating any source of interest within a complex acoustic scene, and has long inspired audio source separation research. Recent efforts have mainly focused on separating speech from noise, speech from speech, musical instruments from each other, or sound events from each other. However, separating an audio mixture (e.g., movie soundtrack) into the three broad categories of speech, music, and sound effects (understood to include ambient noise and natural sound events) has been left largely unexplored, despite a wide range of potential applications. This paper formalizes this task as the cocktail fork problem, and presents the Divide and Remaster (DnR) dataset to foster research on this topic. DnR is built from three well-established audio datasets (LibriSpeech, FMA, FSD50k), taking care to reproduce conditions similar to professionally produced content in terms of source overlap and relative loudness, and made available at CD quality. We benchmark standard source separation algorithms on DnR, and further introduce a new multi-resolution model to better address the variety of acoustic characteristics of the three source types. Our best model produces SI-SDR improvements over the mixture of 11.0 dB for music, 11.2 dB for speech, and 10.8 dB for sound effects.

Darius Petermann, Seungkwon Beack, Minje Kim

An autoencoder-based codec employs quantization to turn its bottleneck layer activation into bitstrings, a process that hinders information flow between the encoder and decoder parts. To circumvent this issue, we employ additional skip connections between the corresponding pair of encoder-decoder layers. The assumption is that, in a mirrored autoencoder topology, a decoder layer reconstructs the intermediate feature representation of its corresponding encoder layer. Hence, any additional information directly propagated from the corresponding encoder layer helps the reconstruction. We implement this kind of skip connections in the form of additional autoencoders, each of which is a small codec that compresses the massive data transfer between the paired encoder-decoder layers. We empirically verify that the proposed hyper-autoencoded architecture improves perceptual audio quality compared to an ordinary autoencoder baseline.

Darius Petermann, Pritish Chandna, Helena Cuesta et al.

Choral singing is a widely practiced form of ensemble singing wherein a group of people sing simultaneously in polyphonic harmony. The most commonly practiced setting for choir ensembles consists of four parts; Soprano, Alto, Tenor and Bass (SATB), each with its own range of fundamental frequencies (F$0$s). The task of source separation for this choral setting entails separating the SATB mixture into the constituent parts. Source separation for musical mixtures is well studied and many deep learning based methodologies have been proposed for the same. However, most of the research has been focused on a typical case which consists in separating vocal, percussion and bass sources from a mixture, each of which has a distinct spectral structure. In contrast, the simultaneous and harmonic nature of ensemble singing leads to high structural similarity and overlap between the spectral components of the sources in a choral mixture, making source separation for choirs a harder task than the typical case. This, along with the lack of an appropriate consolidated dataset has led to a dearth of research in the field so far. In this paper we first assess how well some of the recently developed methodologies for musical source separation perform for the case of SATB choirs. We then propose a novel domain-specific adaptation for conditioning the recently proposed U-Net architecture for musical source separation using the fundamental frequency contour of each of the singing groups and demonstrate that our proposed approach surpasses results from domain-agnostic architectures.