Thomas Pellegrini

Ismail Khalfaoui-Hassani, Thomas Pellegrini, Timothée Masquelier

Dilated Convolution with Learnable Spacings (DCLS) is a recently proposed variation of the dilated convolution in which the spacings between the non-zero elements in the kernel, or equivalently their positions, are learnable. Non-integer positions are handled via interpolation. Thanks to this trick, positions have well-defined gradients. The original DCLS used bilinear interpolation, and thus only considered the four nearest pixels. Yet here we show that longer range interpolations, and in particular a Gaussian interpolation, allow improving performance on ImageNet1k classification on two state-of-the-art convolutional architectures (ConvNeXt and Conv\-Former), without increasing the number of parameters. The method code is based on PyTorch and is available at https://github.com/K-H-Ismail/Dilated-Convolution-with-Learnable-Spacings-PyTorch

Ismail Khalfaoui-Hassani, Timothée Masquelier, Thomas Pellegrini

Dilated convolution with learnable spacings (DCLS) is a recent convolution method in which the positions of the kernel elements are learned throughout training by backpropagation. Its interest has recently been demonstrated in computer vision (ImageNet classification and downstream tasks). Here we show that DCLS is also useful for audio tagging using the AudioSet classification benchmark. We took two state-of-the-art convolutional architectures using depthwise separable convolutions (DSC), ConvNeXt and ConvFormer, and a hybrid one using attention in addition, FastViT, and drop-in replaced all the DSC layers by DCLS ones. This significantly improved the mean average precision (mAP) with the three architectures without increasing the number of parameters and with only a low cost on the throughput. The method code is based on PyTorch and is available at https://github.com/K-H-Ismail/DCLS-Audio

Etienne Labbé, Thomas Pellegrini, Julien Pinquier

Automated Audio Captioning (AAC) aims to develop systems capable of describing an audio recording using a textual sentence. In contrast, Audio-Text Retrieval (ATR) systems seek to find the best matching audio recording(s) for a given textual query (Text-to-Audio) or vice versa (Audio-to-Text). These tasks require different types of systems: AAC employs a sequence-to-sequence model, while ATR utilizes a ranking model that compares audio and text representations within a shared projection subspace. However, this work investigates the relationship between AAC and ATR by exploring the ATR capabilities of an unmodified AAC system, without fine-tuning for the new task. Our AAC system consists of an audio encoder (ConvNeXt-Tiny) trained on AudioSet for audio tagging, and a transformer decoder responsible for generating sentences. For AAC, it achieves a high SPIDEr-FL score of 0.298 on Clotho and 0.472 on AudioCaps on average. For ATR, we propose using the standard Cross-Entropy loss values obtained for any audio/caption pair. Experimental results on the Clotho and AudioCaps datasets demonstrate decent recall values using this simple approach. For instance, we obtained a Text-to-Audio R@1 value of 0.382 for Au-dioCaps, which is above the current state-of-the-art method without external data. Interestingly, we observe that normalizing the loss values was necessary for Audio-to-Text retrieval.

Etienne Labbé, Thomas Pellegrini, Julien Pinquier

Automatic Audio Captioning (AAC) is the task that aims to describe an audio signal using natural language. AAC systems take as input an audio signal and output a free-form text sentence, called a caption. Evaluating such systems is not trivial, since there are many ways to express the same idea. For this reason, several complementary metrics, such as BLEU, CIDEr, SPICE and SPIDEr, are used to compare a single automatic caption to one or several captions of reference, produced by a human annotator. Nevertheless, an automatic system can produce several caption candidates, either using some randomness in the sentence generation process, or by considering the various competing hypothesized captions during decoding with beam-search, for instance. If we consider an end-user of an AAC system, presenting several captions instead of a single one seems relevant to provide some diversity, similarly to information retrieval systems. In this work, we explore the possibility to consider several predicted captions in the evaluation process instead of one. For this purpose, we propose SPIDEr-max, a metric that takes the maximum SPIDEr value among the scores of several caption candidates. To advocate for our metric, we report experiments on Clotho v2.1 and AudioCaps, with a transformed-based system. On AudioCaps for example, this system reached a SPIDEr-max value (with 5 candidates) close to the SPIDEr human score of reference.

Lionel Pibre, Francisco Madrigal, Cyrille Equoy et al.

Meetings are a common activity in professional contexts, and it remains challenging to endow vocal assistants with advanced functionalities to facilitate meeting management. In this context, a task like active speaker detection can provide useful insights to model interaction between meeting participants. Motivated by our application context related to advanced meeting assistant, we want to combine audio and visual information to achieve the best possible performance. In this paper, we propose two different types of fusion for the detection of the active speaker, combining two visual modalities and an audio modality through neural networks. For comparison purpose, classical unsupervised approaches for audio feature extraction are also used. We expect visual data centered on the face of each participant to be very appropriate for detecting voice activity, based on the detection of lip and facial gestures. Thus, our baseline system uses visual data and we chose a 3D Convolutional Neural Network architecture, which is effective for simultaneously encoding appearance and movement. To improve this system, we supplemented the visual information by processing the audio stream with a CNN or an unsupervised speaker diarization system. We have further improved this system by adding visual modality information using motion through optical flow. We evaluated our proposal with a public and state-of-the-art benchmark: the AMI corpus. We analysed the contribution of each system to the merger carried out in order to determine if a given participant is currently speaking. We also discussed the results we obtained. Besides, we have shown that, for our application context, adding motion information greatly improves performance. Finally, we have shown that attention-based fusion improves performance while reducing the standard deviation.

Ismail Khalfaoui-Hassani, Thomas Pellegrini, Timothée Masquelier

Recent works indicate that convolutional neural networks (CNN) need large receptive fields (RF) to compete with visual transformers and their attention mechanism. In CNNs, RFs can simply be enlarged by increasing the convolution kernel sizes. Yet the number of trainable parameters, which scales quadratically with the kernel's size in the 2D case, rapidly becomes prohibitive, and the training is notoriously difficult. This paper presents a new method to increase the RF size without increasing the number of parameters. The dilated convolution (DC) has already been proposed for the same purpose. DC can be seen as a convolution with a kernel that contains only a few non-zero elements placed on a regular grid. Here we present a new version of the DC in which the spacings between the non-zero elements, or equivalently their positions, are no longer fixed but learnable via backpropagation thanks to an interpolation technique. We call this method "Dilated Convolution with Learnable Spacings" (DCLS) and generalize it to the n-dimensional convolution case. However, our main focus here will be on the 2D case. We first tried our approach on ResNet50: we drop-in replaced the standard convolutions with DCLS ones, which increased the accuracy of ImageNet1k classification at iso-parameters, but at the expense of the throughput. Next, we used the recent ConvNeXt state-of-the-art convolutional architecture and drop-in replaced the depthwise convolutions with DCLS ones. This not only increased the accuracy of ImageNet1k classification but also of typical downstream and robustness tasks, again at iso-parameters but this time with negligible cost on throughput, as ConvNeXt uses separable convolutions. Conversely, classic DC led to poor performance with both ResNet50 and ConvNeXt. The code of the method is available at: https://github.com/K-H-Ismail/Dilated-Convolution-with-Learnable-Spacings-PyTorch.

Thomas Pellegrini, Timothée Masquelier

Multi-label audio tagging consists of assigning sets of tags to audio recordings. At inference time, thresholds are applied on the confidence scores outputted by a probabilistic classifier, in order to decide which classes are detected active. In this work, we consider having at disposal a trained classifier and we seek to automatically optimize the decision thresholds according to a performance metric of interest, in our case F-measure (micro-F1). We propose a new method, called SGL-Thresh for Surrogate Gradient Learning of Thresholds, that makes use of gradient descent. Since F1 is not differentiable, we propose to approximate the thresholding operation gradients with the gradients of a sigmoid function. We report experiments on three datasets, using state-of-the-art pre-trained deep neural networks. In all cases, SGL-Thresh outperformed three other approaches: a default threshold value (defThresh), an heuristic search algorithm and a method estimating F1 gradients numerically. It reached 54.9\% F1 on AudioSet eval, compared to 50.7% with defThresh. SGL-Thresh is very fast and scalable to a large number of tags. To facilitate reproducibility, data and source code in Pytorch are available online: https://github.com/topel/SGL-Thresh

Romain Zimmer, Thomas Pellegrini, Srisht Fateh Singh et al.

Recently, it has been shown that spiking neural networks (SNNs) can be trained efficiently, in a supervised manner, using backpropagation through time. Indeed, the most commonly used spiking neuron model, the leaky integrate-and-fire neuron, obeys a differential equation which can be approximated using discrete time steps, leading to a recurrent relation for the potential. The firing threshold causes optimization issues, but they can be overcome using a surrogate gradient. Here, we extend previous approaches in two ways. Firstly, we show that the approach can be used to train convolutional layers. Convolutions can be done in space, time (which simulates conduction delays), or both. Secondly, we include fast horizontal connections à la Denève: when a neuron N fires, we subtract to the potentials of all the neurons with the same receptive the dot product between their weight vectors and the one of neuron N. As Denève et al. showed, this is useful to represent a dynamic multidimensional analog signal in a population of spiking neurons. Here we demonstrate that, in addition, such connections also allow implementing a multidimensional send-on-delta coding scheme. We validate our approach on one speech classification benchmarks: the Google speech command dataset. We managed to reach nearly state-of-the-art accuracy (94%) while maintaining low firing rates (about 5Hz). Our code is based on PyTorch and is available in open source at http://github.com/romainzimmer/s2net

Sebastião Quintas, Isabelle Ferrané, Thomas Pellegrini

The use of synthetic speech as data augmentation is gaining increasing popularity in fields such as automatic speech recognition and speech classification tasks. Despite novel text-to-speech systems with voice cloning capabilities, that allow the usage of a larger amount of voices based on short audio segments, it is known that these systems tend to hallucinate and oftentimes produce bad data that will most likely have a negative impact on the downstream task. In the present work, we conduct a set of experiments around zero-shot learning with synthetic speech data for the specific task of speech commands classification. Our results on the Google Speech Commands dataset show that a simple ASR-based filtering method can have a big impact in the quality of the generated data, translating to a better performance. Furthermore, despite the good quality of the generated speech data, we also show that synthetic and real speech can still be easily distinguishable when using self-supervised (WavLM) features, an aspect further explored with a CycleGAN to bridge the gap between the two types of speech material.

Lucas Block Medin, Thomas Pellegrini, Lucile Gelin

Child speech recognition is still an underdeveloped area of research due to the lack of data (especially on non-English languages) and the specific difficulties of this task. Having explored various architectures for child speech recognition in previous work, in this article we tackle recent self-supervised models. We first compare wav2vec 2.0, HuBERT and WavLM models adapted to phoneme recognition in French child speech, and continue our experiments with the best of them, WavLM base+. We then further adapt it by unfreezing its transformer blocks during fine-tuning on child speech, which greatly improves its performance and makes it significantly outperform our base model, a Transformer+CTC. Finally, we study in detail the behaviour of these two models under the real conditions of our application, and show that WavLM base+ is more robust to various reading tasks and noise levels. Index Terms: speech recognition, child speech, self-supervised learning

Ludovic Tuncay, Etienne Labbé, Thomas Pellegrini

AudioSet is one of the most used and largest datasets in audio tagging, containing about 2 million audio samples that are manually labeled with 527 event categories organized into an ontology. However, the annotations contain inconsistencies, particularly where categories that should be labeled as positive according to the ontology are frequently mislabeled as negative. To address this issue, we apply Hierarchical Label Propagation (HLP), which propagates labels up the ontology hierarchy, resulting in a mean increase in positive labels per audio clip from 1.98 to 2.39 and affecting 109 out of the 527 classes. Our results demonstrate that HLP provides performance benefits across various model architectures, including convolutional neural networks (PANN's CNN6 and ConvNeXT) and transformers (PaSST), with smaller models showing more improvements. Finally, on FSD50K, another widely used dataset, models trained on AudioSet with HLP consistently outperformed those trained without HLP. Our source code will be made available on GitHub.

Ludovic Tuncay, Etienne Labbé, Emmanouil Benetos et al.

Building on the Joint-Embedding Predictive Architecture (JEPA) paradigm, a recent self-supervised learning framework that predicts latent representations of masked regions in high-level feature spaces, we propose Audio-JEPA (Audio Joint-Embedding Predictive Architecture), tailored specifically for audio data. Audio-JEPA uses a simple Vision Transformer backbone to predict latent representations of masked spectrogram patches rather than reconstructing raw audio. We pre-train on unlabeled AudioSet clips (10s, 32kHz) with random patch masking on mel-spectrograms. We evaluate on the X-ARES suite covering speech, music, and environmental sound tasks. Although our implementation is a straightforward translation of the original model to audio, the results still show comparable performance to wav2vec 2.0 and data2vec while using less than one-fifth of their training data and with no hyper-parameter tuning. All code and pretrained checkpoints will be released on GitHub.

Lucile Gelin, Morgane Daniel, Julien Pinquier et al.

Automatic recognition systems for child speech are lagging behind those dedicated to adult speech in the race of performance. This phenomenon is due to the high acoustic and linguistic variability present in child speech caused by their body development, as well as the lack of available child speech data. Young readers speech additionally displays peculiarities, such as slow reading rate and presence of reading mistakes, that hardens the task. This work attempts to tackle the main challenges in phone acoustic modelling for young child speech with limited data, and improve understanding of strengths and weaknesses of a wide selection of model architectures in this domain. We find that transfer learning techniques are highly efficient on end-to-end architectures for adult-to-child adaptation with a small amount of child speech data. Through transfer learning, a Transformer model complemented with a Connectionist Temporal Classification (CTC) objective function, reaches a phone error rate of 28.1%, outperforming a state-of-the-art DNN-HMM model by 6.6% relative, as well as other end-to-end architectures by more than 8.5% relative. An analysis of the models' performance on two specific reading tasks (isolated words and sentences) is provided, showing the influence of the utterance length on attention-based and CTC-based models. The Transformer+CTC model displays an ability to better detect reading mistakes made by children, that can be attributed to the CTC objective function effectively constraining the attention mechanisms to be monotonic.

Léo Cances, Etienne Labbé, Thomas Pellegrini

In this article, we adapted five recent SSL methods to the task of audio classification. The first two methods, namely Deep Co-Training (DCT) and Mean Teacher (MT), involve two collaborative neural networks. The three other algorithms, called MixMatch (MM), ReMixMatch (RMM), and FixMatch (FM), are single-model methods that rely primarily on data augmentation strategies. Using the Wide-ResNet-28-2 architecture in all our experiments, 10% of labeled data and the remaining 90% as unlabeled data for training, we first compare the error rates of the five methods on three standard benchmark audio datasets: Environmental Sound Classification (ESC-10), UrbanSound8K (UBS8K), and Google Speech Commands (GSC). In all but one cases, MM, RMM, and FM outperformed MT and DCT significantly, MM and RMM being the best methods in most experiments. On UBS8K and GSC, MM achieved 18.02% and 3.25% error rate (ER), respectively, outperforming models trained with 100% of the available labeled data, which reached 23.29% and 4.94%, respectively. RMM achieved the best results on ESC-10 (12.00% ER), followed by FM which reached 13.33%. Second, we explored adding the mixup augmentation, used in MM and RMM, to DCT, MT, and FM. In almost all cases, mixup brought consistent gains. For instance, on GSC, FM reached 4.44% and 3.31% ER without and with mixup. Our PyTorch code will be made available upon paper acceptance at https:// github. com/ Labbe ti/ SSLH.

Thomas Pellegrini, Romain Zimmer, Timothée Masquelier

Deep Neural Networks (DNNs) are the current state-of-the-art models in many speech related tasks. There is a growing interest, though, for more biologically realistic, hardware friendly and energy efficient models, named Spiking Neural Networks (SNNs). Recently, it has been shown that SNNs can be trained efficiently, in a supervised manner, using backpropagation with a surrogate gradient trick. In this work, we report speech command (SC) recognition experiments using supervised SNNs. We explored the Leaky-Integrate-Fire (LIF) neuron model for this task, and show that a model comprised of stacked dilated convolution spiking layers can reach an error rate very close to standard DNNs on the Google SC v1 dataset: 5.5%, while keeping a very sparse spiking activity, below 5%, thank to a new regularization term. We also show that modeling the leakage of the neuron membrane potential is useful, since the LIF model outperformed its non-leaky model counterpart significantly.

Leo Cances, Patrice Guyot, Thomas Pellegrini

Sound event detection (SED) aims at identifying audio events (audio tagging task) in recordings and then locating them temporally (localization task). This last task ends with the segmentation of the frame-level class predictions, that determines the onsets and offsets of the audio events. Yet, this step is often overlooked in scientific publications. In this paper, we focus on the post-processing algorithms used to identify the audio event boundaries. Different post-processing steps are investigated, through smoothing, thresholding, and optimization. In particular, we evaluate different approaches for temporal segmentation, namely statistic-based and parametric methods. Experiments are carried out on the DCASE 2018 challenge task 4 data. We compared post-processing algorithms on the temporal prediction curves of two models: one based on the challenge's baseline and a Multiple Instance Learning (MIL) model. Results show the crucial impact of the post-processing methods on the final detection score. Statistic-based methods yield a 22.9% event-based F-score on the evaluation set with our MIL model. Moreover, the best results were obtained using class-dependent parametric methods with 32.0% F-score.

Thomas Pellegrini, Léo Cances

The design of new methods and models when only weakly-labeled data are available is of paramount importance in order to reduce the costs of manual annotation and the considerable human effort associated with it. In this work, we address Sound Event Detection in the case where a weakly annotated dataset is available for training. The weak annotations provide tags of audio events but do not provide temporal boundaries. The objective is twofold: 1) audio tagging, i.e. multi-label classification at recording level, 2) sound event detection, i.e. localization of the event boundaries within the recordings. This work focuses mainly on the second objective. We explore an approach inspired by Multiple Instance Learning, in which we train a convolutional recurrent neural network to give predictions at frame-level, using a custom loss function based on the weak labels and the statistics of the frame-based predictions. Since some sound classes cannot be distinguished with this approach, we improve the method by penalizing similarity between the predictions of the positive classes during training. On the test set used in the DCASE 2018 challenge, consisting of 288 recordings and 10 sound classes, the addition of a penalty resulted in a localization F-score of 34.75%, and brought 10% relative improvement compared to not using the penalty. Our best model achieved a 26.20% F-score on the DCASE-2018 official Eval subset close to the 10-system ensemble approach that ranked second in the challenge with a 29.9% F-score.

Thomas Pellegrini, Jérôme Farinas, Estelle Delpech et al.

In this paper, we describe the outcomes of the challenge organized and run by Airbus and partners in 2018. The challenge consisted of two tasks applied to Air Traffic Control (ATC) speech in English: 1) automatic speech-to-text transcription, 2) call sign detection (CSD). The registered participants were provided with 40 hours of speech along with manual transcriptions. Twenty-two teams submitted predictions on a five hour evaluation set. ATC speech processing is challenging for several reasons: high speech rate, foreign-accented speech with a great diversity of accents, noisy communication channels. The best ranked team achieved a 7.62% Word Error Rate and a 82.41% CSD F1-score. Transcribing pilots' speech was found to be twice as harder as controllers' speech. Remaining issues towards solving ATC ASR are also discussed.

Thomas Pellegrini

Detecting bird sounds in audio recordings automatically, if accurate enough, is expected to be of great help to the research community working in bio- and ecoacoustics, interested in monitoring biodiversity based on audio field recordings. To estimate how accurate the state-of-the-art machine learning approaches are, the Bird Audio Detection challenge involving large audio datasets was recently organized. In this paper, experiments using several types of convolutional neural networks (i.e. standard CNNs, residual nets and densely connected nets) are reported in the framework of this challenge. DenseNets were the preferred solution since they were the best performing and most compact models, leading to a 88.22% area under the receiver operator curve score on the test set of the challenge. Performance gains were obtained thank to data augmentation through time and frequency shifting, model parameter averaging during training and ensemble methods using the geometric mean. On the contrary, the attempts to enlarge the training dataset with samples of the test set with automatic predictions used as pseudo-groundtruth labels consistently degraded performance.