Less Forgetting for Better Generalization: Exploring Continual-learning Fine-tuning Methods for Speech Self-supervised Representations

Despite being trained on massive and diverse datasets, speech self-supervised encoders are generally used for downstream purposes as mere frozen feature extractors or model initializers before fine-tuning. The former severely limits the exploitation of large encoders, while the latter hurts the robustness acquired during pretraining, especially in low-resource scenarios. This work explores middle-ground solutions, conjecturing that reducing the forgetting of the self-supervised task during the downstream fine-tuning leads to better generalization. To prove this, focusing on speech recognition, we benchmark different continual-learning approaches during fine-tuning and show that they improve both in-domain and out-of-domain generalization abilities. Relative performance gains reach 15.7% and 22.5% with XLSR used as the encoder on two English and Danish speech recognition tasks. Further probing experiments show that these gains are indeed linked to less forgetting.

Winner-takes-all learners are geometry-aware conditional density estimators

Winner-takes-all training is a simple learning paradigm, which handles ambiguous tasks by predicting a set of plausible hypotheses. Recently, a connection was established between Winner-takes-all training and centroidal Voronoi tessellations, showing that, once trained, hypotheses should quantize optimally the shape of the conditional distribution to predict. However, the best use of these hypotheses for uncertainty quantification is still an open question.In this work, we show how to leverage the appealing geometric properties of the Winner-takes-all learners for conditional density estimation, without modifying its original training scheme. We theoretically establish the advantages of our novel estimator both in terms of quantization and density estimation, and we demonstrate its competitiveness on synthetic and real-world datasets, including audio data.

A lightweight dual-stage framework for personalized speech enhancement based on DeepFilterNet2

Isolating the desired speaker's voice amidst multiplespeakers in a noisy acoustic context is a challenging task. Per-sonalized speech enhancement (PSE) endeavours to achievethis by leveraging prior knowledge of the speaker's voice.Recent research efforts have yielded promising PSE mod-els, albeit often accompanied by computationally intensivearchitectures, unsuitable for resource-constrained embeddeddevices. In this paper, we introduce a novel method to per-sonalize a lightweight dual-stage Speech Enhancement (SE)model and implement it within DeepFilterNet2, a SE modelrenowned for its state-of-the-art performance. We seek anoptimal integration of speaker information within the model,exploring different positions for the integration of the speakerembeddings within the dual-stage enhancement architec-ture. We also investigate a tailored training strategy whenadapting DeepFilterNet2 to a PSE task. We show that ourpersonalization method greatly improves the performancesof DeepFilterNet2 while preserving minimal computationaloverhead.

Online speaker diarization of meetings guided by speech separation

Overlapped speech is notoriously problematic for speaker diarization systems. Consequently, the use of speech separation has recently been proposed to improve their performance. Although promising, speech separation models struggle with realistic data because they are trained on simulated mixtures with a fixed number of speakers. In this work, we introduce a new speech separation-guided diarization scheme suitable for the online speaker diarization of long meeting recordings with a variable number of speakers, as present in the AMI corpus. We envisage ConvTasNet and DPRNN as alternatives for the separation networks, with two or three output sources. To obtain the speaker diarization result, voice activity detection is applied on each estimated source. The final model is fine-tuned end-to-end, after first adapting the separation to real data using AMI. The system operates on short segments, and inference is performed by stitching the local predictions using speaker embeddings and incremental clustering. The results show that our system improves the state-of-the-art on the AMI headset mix, using no oracle information and under full evaluation (no collar and including overlapped speech). Finally, we show the strength of our system particularly on overlapped speech sections.

On the choice of the optimal temporal support for audio classification with Pre-trained embeddings

Current state-of-the-art audio analysis systems rely on pre-trained embedding models, often used off-the-shelf as (frozen) feature extractors. Choosing the best one for a set of tasks is the subject of many recent publications. However, one aspect often overlooked in these works is the influence of the duration of audio input considered to extract an embedding, which we refer to as Temporal Support (TS). In this work, we study the influence of the TS for well-established or emerging pre-trained embeddings, chosen to represent different types of architectures and learning paradigms. We conduct this evaluation using both musical instrument and environmental sound datasets, namely OpenMIC, TAU Urban Acoustic Scenes 2020 Mobile, and ESC-50. We especially highlight that Audio Spectrogram Transformer-based systems (PaSST and BEATs) remain effective with smaller TS, which therefore allows for a drastic reduction in memory and computational cost. Moreover, we show that by choosing the optimal TS we reach competitive results across all tasks. In particular, we improve the state-of-the-art results on OpenMIC, using BEATs and PaSST without any fine-tuning.

Collaborating Foundation models for Domain Generalized Semantic Segmentation

Domain Generalized Semantic Segmentation (DGSS) deals with training a model on a labeled source domain with the aim of generalizing to unseen domains during inference. Existing DGSS methods typically effectuate robust features by means of Domain Randomization (DR). Such an approach is often limited as it can only account for style diversification and not content. In this work, we take an orthogonal approach to DGSS and propose to use an assembly of CoLlaborative FOUndation models for Domain Generalized Semantic Segmentation (CLOUDS). In detail, CLOUDS is a framework that integrates FMs of various kinds: (i) CLIP backbone for its robust feature representation, (ii) generative models to diversify the content, thereby covering various modes of the possible target distribution, and (iii) Segment Anything Model (SAM) for iteratively refining the predictions of the segmentation model. Extensive experiments show that our CLOUDS excels in adapting from synthetic to real DGSS benchmarks and under varying weather conditions, notably outperforming prior methods by 5.6% and 6.7% on averaged miou, respectively. The code is available at : https://github.com/yasserben/CLOUDS

Resilient Multiple Choice Learning: A learned scoring scheme with application to audio scene analysis

We introduce Resilient Multiple Choice Learning (rMCL), an extension of the MCL approach for conditional distribution estimation in regression settings where multiple targets may be sampled for each training input. Multiple Choice Learning is a simple framework to tackle multimodal density estimation, using the Winner-Takes-All (WTA) loss for a set of hypotheses. In regression settings, the existing MCL variants focus on merging the hypotheses, thereby eventually sacrificing the diversity of the predictions. In contrast, our method relies on a novel learned scoring scheme underpinned by a mathematical framework based on Voronoi tessellations of the output space, from which we can derive a probabilistic interpretation. After empirically validating rMCL with experiments on synthetic data, we further assess its merits on the sound source localization problem, demonstrating its practical usefulness and the relevance of its interpretation.

Speech Self-Supervised Representations Benchmarking: a Case for Larger Probing Heads

Self-supervised learning (SSL) leverages large datasets of unlabeled speech to reach impressive performance with reduced amounts of annotated data. The high number of proposed approaches fostered the emergence of comprehensive benchmarks that evaluate their performance on a set of downstream tasks exploring various aspects of the speech signal. However, while the number of considered tasks has been growing, most proposals rely upon a single downstream architecture that maps the frozen SSL representations to the task labels. This study examines how benchmarking results are affected by changes in the probing head architecture. Interestingly, we found that altering the downstream architecture structure leads to significant fluctuations in the performance ranking of the evaluated models. Against common practices in speech SSL benchmarking, we evaluate larger-capacity probing heads, showing their impact on performance, inference costs, generalization and multi-level feature exploitation.

SAMbA: Speech enhancement with Asynchronous ad-hoc Microphone Arrays

Speech enhancement in ad-hoc microphone arrays is often hindered by the asynchronization of the devices composing the microphone array. Asynchronization comes from sampling time offset and sampling rate offset which inevitably occur when the microphones are embedded in different hardware components. In this paper, we propose a deep neural network (DNN)-based speech enhancement solution that is suited for applications in ad-hoc microphone arrays because it is distributed and copes with asynchronization. We show that asynchronization has a limited impact on the spatial filtering and mostly affects the performance of the DNNs. Instead of resynchronising the signals, which requires costly processing steps, we use an attention mechanism which makes the DNNs, thus our whole pipeline, robust to asynchronization. We also show that the attention mechanism leads to the asynchronization parameters in an unsupervised manner.

Speech Self-Supervised Representation Benchmarking: Are We Doing it Right?

Self-supervised learning (SSL) has recently allowed leveraging large datasets of unlabeled speech signals to reach impressive performance on speech tasks using only small amounts of annotated data. The high number of proposed approaches fostered the need and rise of extended benchmarks that evaluate their performance on a set of downstream tasks exploring various aspects of the speech signal. However, and while the number of considered tasks has been growing, most rely upon a single decoding architecture that maps the frozen SSL representations to the downstream labels. This work investigates the robustness of such benchmarking results to changes in the decoder architecture. Interestingly, it appears that varying the architecture of the downstream decoder leads to significant variations in the leaderboards of most tasks. Concerningly, our study reveals that benchmarking using limited decoders may cause a counterproductive increase in the sizes of the developed SSL models.