arXiv subjects

Romain Serizel

Publications and source records attributed to Romain Serizel.

At least 19 recordsLinked to original sources

arXiv2026

Investigating the Performance and Energy Costs of Replicating Band-Split RNN for Music Source Separation

Band-split recurrent neural network (BSRNN) is a popular music source separation model that yields close to state-of-the-art results using reasonable computational resources and public datasets. It is therefore interesting from a reproducible research perspective, but achieving its performance is not straightforward since its full code is not available. In this paper, we conduct a replication of BSRNN via implementing the full pipeline. We extend the original paper's analysis by experimentally studying various design choices about data preprocessing, the optimization protocol, and architectural parameters. We report and discuss this project's energy cost, and we underline how its footprint could have been substantial lower upon availability of the full pipeline, which advocates for more reproducible research practices. To comply with this objective, we publicly release our code and pre-trained models.

cs.SD↗

arXiv2026

It Takes Few to TANGO: A Quantized Distributed Model for Binaural Speech Enhancement

Neural network-based multichannel speech enhancement systems achieve strong enhancement performance, but their computational and memory requirements limit deployment on resource-constrained devices. This paper investigates low-precision inference for TANGO, a hybrid distributed binaural speech enhancement system combining neural mask estimation with spatial filtering. We evaluate post-training quantization and quantization-aware training for the neural components, and analyze how quantization errors in the mask estimators propagate through the downstream spatial filtering stage. Our analysis shows that, although quantization degrades intermediate mask estimates, the spatial filtering stage compensates for most quantization-induced errors. Leveraging this robustness, we simplify TANGO into MN-TANGO, reducing both model size and computational complexity while maintaining comparable final performance. By combining INT8 weight-and-activation quantization with ERB compression and grouped recurrent layers, the most compact MN-TANGO reaches 4.65 MMAC/s and 0.177 MB.

cs.SD↗

arXiv2026

Time-Frequency Weighted Losses for Phoneme Reconstruction in DNN-Based Speech Enhancement

Conventional training losses for speech enhancement based on the signal-to-distortion ratio (SDR) treat all time-frequency (TF) regions uniformly, overlooking the fine-grained spectral cues that are relevant to specific phoneme intelligibility. We propose a TF weighting framework that modulates the SDR objective based on local speech presence, speech-to-interference ratio (SIR), and spectral flux. By integrating these factors into a differentiable objective, the framework emphasizes TF bins with high speech-noise competition while also accounting for transient cues such as consonant bursts. Experimental results show that our approach improves objective frequency-weighted enhancement metrics, as well as phoneme recognition accuracy, particularly for consonants. Spectral analysis shows better reconstruction of mid-frequency structures at less adverse SIR.

cs.SD↗

arXiv2026

Audio-visual Contrastive Alignment for Diffusion-based Visual-conditioned Speech Enhancement

Audio-visual speech enhancement (AVSE) exploits visual cues such as lip movements to recover speech in noisy environments. Recent work introduced diffusion-based unsupervised AVSE, where a speech diffusion model conditioned on visual features via cross-attention is trained and used as a data-driven prior for posterior sampling-based speech enhancement. Despite promising performance over its audio-only counterpart, the impact of explicitly enforcing cross-modal alignment in the fusion remains unclear. In this work, we propose to augment the diffusion training objective with a contrastive audio-visual loss to encourage stronger use of visual information while keeping the posterior sampling framework unchanged. Experiments across matched and mismatched test data show consistent improvements in interference suppression, signal reconstruction, and perceptual quality, with the largest gains at low SNRs. Code is available at https://github.com/ cexauce/AV-CA-DiffUSE

eess.SP↗

arXiv2026

Metric Analysis for Spatial Semantic Segmentation of Sound Scenes

Spatial semantic segmentation of sound scenes (S5) consists of jointly performing audio source separation and sound event classification from a multichannel audio mixture. Evaluating S5 systems with separation and classification metrics individually makes system comparison difficult, whereas existing joint metrics, such as the class-aware signal-to-distortion ratio (CA-SDR), can conflate separation and labeling errors. In particular, CA-SDR relies on predicted class labels for source matching, which may obscure label swaps or misclassifications when the underlying source estimates remain perceptually correct. In this work, we introduce the class and source-aware signal-to-distortion ratio (CASA-SDR), a new metric that performs permutation-invariant source matching before computing classification errors, thereby shifting from a classification-focused approach to a separation-focused approach. We first analyze CA-SDR in controlled scenarios with oracle separation and synthetic classification errors, as well as under controlled cross-contamination between sources, and compare its behavior to that of the classical SDR and CASA-SDR. We also study the impact of classification errors on the metrics by introducing error-based and source-based aggregation strategies. Finally, we compare CA-SDR and CASA-SDR on systems submitted to Task 4 of the DCASE 2025 challenge, highlighting the cases where CA-SDR over-penalizes label swaps or poorly separated sources, while CASA-SDR provides a more interpretable separation-centric assessment of S5 performance.

cs.SD↗

arXiv2026

Description and Discussion on DCASE 2026 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes

This paper presents an overview of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2026 Challenge Task 4, Spatial Semantic Segmentation of Sound Scenes (S5). The S5 task focuses on the joint detection and separation of sound events in complex spatial audio mixtures, contributing to the foundation of immersive communication. First introduced in DCASE 2025, the S5 task continues in DCASE 2026 Task 4 with key changes to better reflect real-world conditions, including allowing mixtures to contain multiple sources of the same class and to contain no target sources. In this paper, we describe task setting, along with the corresponding updates to the evaluation metrics and dataset. The experimental results of the submitted systems are also reported and analyzed. The official access point for data and code is https://github.com/nttcslab/dcase2026_task4_baseline.

eess.AS↗

arXiv2026

Diffusion-based Frameworks for Unsupervised Speech Enhancement

This paper addresses unsupervised diffusion-based single-channel speech enhancement (SE). Prior work in this direction combines a score-based diffusion model trained on clean speech with a Gaussian noise model whose covariance is structured by non-negative matrix factorization (NMF). This combination is used within an iterative expectation-maximization (EM) scheme, in which a diffusion-based posterior-sampling E-step estimates the clean speech. We first revisit this framework and propose to explicitly model both speech and acoustic noise as latent variables, jointly sampling them in the E-step instead of sampling speech alone as in previous approaches. We then introduce a new semi-supervised SE framework that replaces the NMF noise prior with a diffusion-based noise model, learned jointly with the speech prior in a single conditional score model. Within this framework, we derive two variants: one that implicitly accounts for noise and one that explicitly treats noise as a latent variable. Experiments on WSJ0-QUT and VoiceBank-DEMAND show that explicit noise modeling systematically improves SE performance for both NMF-based and diffusion-based noise priors. Under matched conditions, the diffusion-based noise model attains the best overall quality and intelligibility among unsupervised methods, while under mismatched conditions the proposed NMF-based explicit-noise framework is more robust and suffers less degradation than several supervised baselines. Code, demo, and supplementary materials are publicly available.

cs.SD↗

arXiv2026

The Costs of Reproducibility in Music Separation Research: a Replication of Band-Split RNN

Music source separation is the task of isolating the instrumental tracks from a music song. Despite its spectacular recent progress, the trend towards more complex architectures and training protocols exacerbates reproducibility issues. The band-split recurrent neural networks (BSRNN) model is promising in this regard, since it yields close to state-of-the-art results on public datasets, and requires reasonable resources for training. Unfortunately, it is not straightforward to reproduce since its full code is not available. In this paper, we attempt to replicate BSRNN as closely as possible to the original paper through extensive experiments, which allows us to conduct a critical reflection on this reproducibility issue. Our contributions are three-fold. First, this study yields several insights on the model design and training pipeline, which sheds light on potential future improvements. In particular, since we were unsuccessful in reproducing the original results, we explore additional variants that ultimately yield an optimized BSRNN model, whose performance largely improves that of the original. Second, we discuss reproducibility issues from both methodological and practical perspectives. We notably underline how substantial time and energy costs could have been saved upon availability of the full pipeline. Third, our code and pre-trained models are released publicly to foster reproducible research. We hope that this study will contribute to spread awareness on the importance of reproducible research in the music separation community, and help promoting more transparent and sustainable practices.

cs.SD↗

arXiv2026

Domain-Invariant Representation Learning of Bird Sounds

Passive acoustic monitoring (PAM) is crucial for bioacoustic research, enabling non-invasive species tracking and biodiversity monitoring. Citizen science platforms provide large annotated datasets from focal recordings, where the target species is intentionally recorded. However, PAM requires monitoring in passive soundscapes, creating a domain shift between focal and passive recordings, challenging deep learning models trained on focal recordings. To address domain generalization, we leverage supervised contrastive learning by enforcing domain invariance across same-class examples from different domains. Additionally, we propose ProtoCLR, an alternative to SupCon loss which reduces the computational complexity by comparing examples to class prototypes instead of pairwise comparisons. We conduct few-shot classification based on BIRB, a large-scale bird sound benchmark to assess pre-trained bioacoustic models. Our findings suggest that ProtoCLR is a better alternative to SupCon.

cs.SD↗

arXiv2025

Towards Low-Latency Tracking of Multiple Speakers With Short-Context Speaker Embeddings

Speaker embeddings are promising identity-related features that can enhance the identity assignment performance of a tracking system by leveraging its spatial predictions, i.e, by performing identity reassignment. Common speaker embedding extractors usually struggle with short temporal contexts and overlapping speech, which imposes long-term identity reassignment to exploit longer temporal contexts. However, this increases the probability of tracking system errors, which in turn impacts negatively on identity reassignment. To address this, we propose a Knowledge Distillation (KD) based training approach for short context speaker embedding extraction from two speaker mixtures. We leverage the spatial information of the speaker of interest using beamforming to reduce overlap. We study the feasibility of performing identity reassignment over blocks of fixed size, i.e., blockwise identity reassignment, to go towards a low-latency speaker embedding based tracking system. Results demonstrate that our distilled models are effective at short-context embedding extraction and more robust to overlap. Although, blockwise reassignment results indicate that further work is needed to handle simultaneous speech more effectively.

eess.AS↗

arXiv2025

Diffused Responsibility: Analyzing the Energy Consumption of Generative Text-to-Audio Diffusion Models

Text-to-audio models have recently emerged as a powerful technology for generating sound from textual descriptions. However, their high computational demands raise concerns about energy consumption and environmental impact. In this paper, we conduct an analysis of the energy usage of 7 state-of-the-art text-to-audio diffusion-based generative models, evaluating to what extent variations in generation parameters affect energy consumption at inference time. We also aim to identify an optimal balance between audio quality and energy consumption by considering Pareto-optimal solutions across all selected models. Our findings provide insights into the trade-offs between performance and environmental impact, contributing to the development of more efficient generative audio models.

eess.AS↗

arXiv2025

Posterior Transition Modeling for Unsupervised Diffusion-Based Speech Enhancement

We explore unsupervised speech enhancement using diffusion models as expressive generative priors for clean speech. Existing approaches guide the reverse diffusion process using noisy speech through an approximate, noise-perturbed likelihood score, combined with the unconditional score via a trade-off hyperparameter. In this work, we propose two alternative algorithms that directly model the conditional reverse transition distribution of diffusion states. The first method integrates the diffusion prior with the observation model in a principled way, removing the need for hyperparameter tuning. The second defines a diffusion process over the noisy speech itself, yielding a fully tractable and exact likelihood score. Experiments on the WSJ0-QUT and VoiceBank-DEMAND datasets demonstrate improved enhancement metrics and greater robustness to domain shifts compared to both supervised and unsupervised baselines.

cs.SD↗

arXiv2025

Evaluating Multichannel Speech Enhancement Algorithms at the Phoneme Scale Across Genders

Multichannel speech enhancement algorithms are essential for improving the intelligibility of speech signals in noisy environments. These algorithms are usually evaluated at the utterance level, but this approach overlooks the disparities in acoustic characteristics that are observed in different phoneme categories and between male and female speakers. In this paper, we investigate the impact of gender and phonetic content on speech enhancement algorithms. We motivate this approach by outlining phoneme- and gender-specific spectral features. Our experiments reveal that while utterance-level differences between genders are minimal, significant variations emerge at the phoneme level. Results show that the tested algorithms better reduce interference with fewer artifacts on female speech, particularly in plosives, fricatives, and vowels. Additionally, they demonstrate greater performance for female speech in terms of perceptual and speech recognition metrics.

cs.SD↗

arXiv2025

Frequency-Weighted Training Losses for Phoneme-Level DNN-based Speech Enhancement

Recent advances in deep learning have significantly improved multichannel speech enhancement algorithms, yet conventional training loss functions such as the scale-invariant signal-to-distortion ratio (SDR) may fail to preserve fine-grained spectral cues essential for phoneme intelligibility. In this work, we propose perceptually-informed variants of the SDR loss, formulated in the time-frequency domain and modulated by frequency-dependent weighting schemes. These weights are designed to emphasize time-frequency regions where speech is prominent or where the interfering noise is particularly strong. We investigate both fixed and adaptive strategies, including ANSI band-importance weights, spectral magnitude-based weighting, and dynamic weighting based on the relative amount of speech and noise. We train the FaSNet multichannel speech enhancement model using these various losses. Experimental results show that while standard metrics such as the SDR are only marginally improved, their perceptual frequency-weighted counterparts exhibit a more substantial improvement. Besides, spectral and phoneme-level analysis indicates better consonant reconstruction, which points to a better preservation of certain acoustic cues.

cs.SD↗

arXiv2025

Speaker Embeddings to Improve Tracking of Intermittent and Moving Speakers

Speaker tracking methods often rely on spatial observations to assign coherent track identities over time. This raises limits in scenarios with intermittent and moving speakers, i.e., speakers that may change position when they are inactive, thus leading to discontinuous spatial trajectories. This paper proposes to investigate the use of speaker embeddings, in a simple solution to this issue. We propose to perform identity reassignment post-tracking, using speaker embeddings. We leverage trajectory-related information provided by an initial tracking step and multichannel audio signal. Beamforming is used to enhance the signal towards the speakers' positions in order to compute speaker embeddings. These are then used to assign new track identities based on an enrollment pool. We evaluate the performance of the proposed speaker embedding-based identity reassignment method on a dataset where speakers change position during inactivity periods. Results show that it consistently improves the identity assignment performance of neural and standard tracking systems. In particular, we study the impact of beamforming and input duration for embedding extraction.

eess.AS↗

arXiv2025

Description and Discussion on DCASE 2025 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes

Spatial Semantic Segmentation of Sound Scenes (S5) aims to enhance technologies for sound event detection and separation from multi-channel input signals that mix multiple sound events with spatial information. This is a fundamental basis of immersive communication. The ultimate goal is to separate sound event signals with 6 Degrees of Freedom (6DoF) information into dry sound object signals and metadata about the object type (sound event class) and representing spatial information, including direction. However, because several existing challenge tasks already provide some of the subset functions, this task for this year focuses on detecting and separating sound events from multi-channel spatial input signals. This paper outlines the S5 task setting of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2025 Challenge Task 4 and the DCASE2025 Task 4 Dataset, newly recorded and curated for this task. We also report experimental results for an S5 system trained and evaluated on this dataset. The full version of this paper will be published after the challenge results are made public.

cs.SD↗

arXiv2025

Tracking of Intermittent and Moving Speakers : Dataset and Metrics

This paper presents the problem of tracking intermittent and moving sources, i.e, sources that may change position when they are inactive. This issue is seldom explored, and most current tracking methods rely on spatial observations for track identity management. They are either based on a previous localization step, or designed to perform joint localization and tracking by predicting ordered position estimates. This raises concerns about whether such methods can maintain reliable track identity assignment performance when dealing with discontinuous spatial tracks, which may be caused by a change of direction during silence. We introduce LibriJump, a novel dataset of acoustic scenes in the First Order Ambisonics format focusing on speaker tracking. The dataset contains speakers with changing positions during inactivity periods, thus simulating discontinuous tracks. To measure the identity assignment performance, we propose to use tracking association metrics adapted from the computer vision community. We provide experiments showing the complementarity of association metrics with previously used tracking metrics, given continuous and discontinuous spatial tracks.

eess.AS↗

arXiv2025

Diffusion-based Unsupervised Audio-visual Speech Enhancement

This paper proposes a new unsupervised audio-visual speech enhancement (AVSE) approach that combines a diffusion-based audio-visual speech generative model with a non-negative matrix factorization (NMF) noise model. First, the diffusion model is pre-trained on clean speech conditioned on corresponding video data to simulate the speech generative distribution. This pre-trained model is then paired with the NMF-based noise model to estimate clean speech iteratively. Specifically, a diffusion-based posterior sampling approach is implemented within the reverse diffusion process, where after each iteration, a speech estimate is obtained and used to update the noise parameters. Experimental results confirm that the proposed AVSE approach not only outperforms its audio-only counterpart but also generalizes better than a recent supervised-generative AVSE method. Additionally, the new inference algorithm offers a better balance between inference speed and performance compared to the previous diffusion-based method. Code and demo available at: https://jeaneudesayilo.github.io/fast_UdiffSE

cs.SD↗