Yannick Estève

Marcely Zanon Boito, John Ortega, Hugo Riguidel et al. · meta-ai

This paper describes the ON-TRAC Consortium translation systems developed for two challenge tracks featured in the Evaluation Campaign of IWSLT 2022: low-resource and dialect speech translation. For the Tunisian Arabic-English dataset (low-resource and dialect tracks), we build an end-to-end model as our joint primary submission, and compare it against cascaded models that leverage a large fine-tuned wav2vec 2.0 model for ASR. Our results show that in our settings pipeline approaches are still very competitive, and that with the use of transfer learning, they can outperform end-to-end models for speech translation (ST). For the Tamasheq-French dataset (low-resource track) our primary submission leverages intermediate representations from a wav2vec 2.0 model trained on 234 hours of Tamasheq audio, while our contrastive model uses a French phonetic transcription of the Tamasheq audio as input in a Conformer speech translation architecture jointly trained on automatic speech recognition, ST and machine translation losses. Our results highlight that self-supervised models trained on smaller sets of target data are more effective to low-resource end-to-end ST fine-tuning, compared to large off-the-shelf models. Results also illustrate that even approximate phonetic transcriptions can improve ST scores.

Haroun Elleuch, Salima Mdhaffar, Yannick Estève et al.

We present ADI-20, an extension of the previously published ADI-17 Arabic Dialect Identification (ADI) dataset. ADI-20 covers all Arabic-speaking countries' dialects. It comprises 3,556 hours from 19 Arabic dialects in addition to Modern Standard Arabic (MSA). We used this dataset to train and evaluate various state-of-the-art ADI systems. We explored fine-tuning pre-trained ECAPA-TDNN-based models, as well as Whisper encoder blocks coupled with an attention pooling layer and a classification dense layer. We investigated the effect of (i) training data size and (ii) the model's number of parameters on identification performance. Our results show a small decrease in F1 score while using only 30% of the original training data. We open-source our collected data and trained models to enable the reproduction of our work, as well as support further research in ADI.

Marcely Zanon Boito, Laurent Besacier, Natalia Tomashenko et al.

Self-supervised models for speech processing emerged recently as popular foundation blocks in speech processing pipelines. These models are pre-trained on unlabeled audio data and then used in speech processing downstream tasks such as automatic speech recognition (ASR) or speech translation (ST). Since these models are now used in research and industrial systems alike, it becomes necessary to understand the impact caused by some features such as gender distribution within pre-training data. Using French as our investigation language, we train and compare gender-specific wav2vec 2.0 models against models containing different degrees of gender balance in their pre-training data. The comparison is performed by applying these models to two speech-to-text downstream tasks: ASR and ST. Results show the type of downstream integration matters. We observe lower overall performance using gender-specific pre-training before fine-tuning an end-to-end ASR system. However, when self-supervised models are used as feature extractors, the overall ASR and ST results follow more complex patterns in which the balanced pre-trained model does not necessarily lead to the best results. Lastly, our crude 'fairness' metric, the relative performance difference measured between female and male test sets, does not display a strong variation from balanced to gender-specific pre-trained wav2vec 2.0 models.

Fethi Bougares, Salima Mdhaffar, Haroun Elleuch et al.

In this paper, we introduce TEDxTN, the first publicly available Tunisian Arabic to English speech translation dataset. This work is in line with the ongoing effort to mitigate the data scarcity obstacle for a number of Arabic dialects. We collected, segmented, transcribed and translated 108 TEDx talks following our internally developed annotations guidelines. The collected talks represent 25 hours of speech with code-switching that cover speakers with various accents from over 11 different regions of Tunisia. We make the annotation guidelines and corpus publicly available. This will enable the extension of TEDxTN to new talks as they become available. We also report results for strong baseline systems of Speech Recognition and Speech Translation using multiple pre-trained and fine-tuned end-to-end models. This corpus is the first open source and publicly available speech translation corpus of Code-Switching Tunisian dialect. We believe that this is a valuable resource that can motivate and facilitate further research on the natural language processing of Tunisian Dialect.

Tuan Nguyen, Salima Mdhaffar, Natalia Tomashenko et al.

This paper presents a study on the use of federated learning to train an ASR model based on a wav2vec 2.0 model pre-trained by self supervision. Carried out on the well-known TED-LIUM 3 dataset, our experiments show that such a model can obtain, with no use of a language model, a word error rate of 10.92% on the official TED-LIUM 3 test set, without sharing any data from the different users. We also analyse the ASR performance for speakers depending to their participation to the federated learning. Since federated learning was first introduced for privacy purposes, we also measure its ability to protect speaker identity. To do that, we exploit an approach to analyze information contained in exchanged models based on a neural network footprint on an indicator dataset. This analysis is made layer-wise and shows which layers in an exchanged wav2vec 2.0 based model bring the speaker identity information.

Lucas Maison, Yannick Estève

Thanks to the rise of self-supervised learning, automatic speech recognition (ASR) systems now achieve near-human performance on a wide variety of datasets. However, they still lack generalization capability and are not robust to domain shifts like accent variations. In this work, we use speech audio representing four different French accents to create fine-tuning datasets that improve the robustness of pre-trained ASR models. By incorporating various accents in the training set, we obtain both in-domain and out-of-domain improvements. Our numerical experiments show that we can reduce error rates by up to 25% (relative) on African and Belgian accents compared to single-domain training while keeping a good performance on standard French.

Salima Mdhaffar, Jarod Duret, Titouan Parcollet et al.

Recent works showed that end-to-end neural approaches tend to become very popular for spoken language understanding (SLU). Through the term end-to-end, one considers the use of a single model optimized to extract semantic information directly from the speech signal. A major issue for such models is the lack of paired audio and textual data with semantic annotation. In this paper, we propose an approach to build an end-to-end neural model to extract semantic information in a scenario in which zero paired audio data is available. Our approach is based on the use of an external model trained to generate a sequence of vectorial representations from text. These representations mimic the hidden representations that could be generated inside an end-to-end automatic speech recognition (ASR) model by processing a speech signal. An SLU neural module is then trained using these representations as input and the annotated text as output. Last, the SLU module replaces the top layers of the ASR model to achieve the construction of the end-to-end model. Our experiments on named entity recognition, carried out on the QUAERO corpus, show that this approach is very promising, getting better results than a comparable cascade approach or than the use of synthetic voices.

Gaëlle Laperrière, Valentin Pelloin, Mickaël Rouvier et al.

In this paper we examine the use of semantically-aligned speech representations for end-to-end spoken language understanding (SLU). We employ the recently-introduced SAMU-XLSR model, which is designed to generate a single embedding that captures the semantics at the utterance level, semantically aligned across different languages. This model combines the acoustic frame-level speech representation learning model (XLS-R) with the Language Agnostic BERT Sentence Embedding (LaBSE) model. We show that the use of the SAMU-XLSR model instead of the initial XLS-R model improves significantly the performance in the framework of end-to-end SLU. Finally, we present the benefits of using this model towards language portability in SLU.

Léo Jacqmin, Lucas Druart, Yannick Estève et al.

Though Dialogue State Tracking (DST) is a core component of spoken dialogue systems, recent work on this task mostly deals with chat corpora, disregarding the discrepancies between spoken and written language.In this paper, we propose OLISIA, a cascade system which integrates an Automatic Speech Recognition (ASR) model and a DST model. We introduce several adaptations in the ASR and DST modules to improve integration and robustness to spoken conversations.With these adaptations, our system ranked first in DSTC11 Track 3, a benchmark to evaluate spoken DST. We conduct an in-depth analysis of the results and find that normalizing the ASR outputs and adapting the DST inputs through data augmentation, along with increasing the pre-trained models size all play an important role in reducing the performance discrepancy between written and spoken conversations.

Lucas Maison, Yannick Estève

Automatic speech recognition (ASR) systems become increasingly efficient thanks to new advances in neural network training like self-supervised learning. However, they are known to be unfair toward certain groups, for instance, people speaking with an accent. In this work, we use the French Common Voice dataset to quantify the biases of a pre-trained wav2vec~2.0 model toward several demographic groups. By fine-tuning the pre-trained model on a variety of fixed-size, carefully crafted training sets, we demonstrate the importance of speaker diversity. We also run an in-depth analysis of the Common Voice corpus and identify important shortcomings that should be taken into account by users of this dataset.

Jarod Duret, Titouan Parcollet, Yannick Estève

We propose a method for speech-to-speech emotionpreserving translation that operates at the level of discrete speech units. Our approach relies on the use of multilingual emotion embedding that can capture affective information in a language-independent manner. We show that this embedding can be used to predict the pitch and duration of speech units in a target language, allowing us to resynthesize the source speech signal with the same emotional content. We evaluate our approach to English and French speech signals and show that it outperforms a baseline method that does not use emotional information, including when the emotion embedding is extracted from a different language. Even if this preliminary study does not address directly the machine translation issue, our results demonstrate the effectiveness of our approach for cross-lingual emotion preservation in the context of speech resynthesis.

Ryan Whetten, Titouan Parcollet, Marco Dinarelli et al.

Self-supervised learning (SSL) has transformed speech processing, yet its reliance on massive pre-training datasets remains a bottleneck. While robustness is often attributed to scale and diversity, the role of the data distribution is less understood. We systematically examine how curated subsets of pre-training data influence Automatic Speech Recognition (ASR) performance. Surprisingly, optimizing for acoustic, speaker, or linguistic diversity yields no clear improvements over random sampling. Instead, we find that prioritizing the longest utterances achieves superior ASR results while using only half the original dataset, reducing pre-training time by 24% on a large corpora. These findings suggest that for pre-training speech SSL models, data length is a more critical factor than either data diversity or overall data quantity for performance and efficiency, offering a new perspective for data selection strategies in SSL speech processing.

Gaëlle Laperrière, Ha Nguyen, Sahar Ghannay et al.

Over the past few years, self-supervised learned speech representations have emerged as fruitful replacements for conventional surface representations when solving Spoken Language Understanding (SLU) tasks. Simultaneously, multilingual models trained on massive textual data were introduced to encode language agnostic semantics. Recently, the SAMU-XLSR approach introduced a way to make profit from such textual models to enrich multilingual speech representations with language agnostic semantics. By aiming for better semantic extraction on a challenging Spoken Language Understanding task and in consideration with computation costs, this study investigates a specific in-domain semantic enrichment of the SAMU-XLSR model by specializing it on a small amount of transcribed data from the downstream task. In addition, we show the benefits of the use of same-domain French and Italian benchmarks for low-resource language portability and explore cross-domain capacities of the enriched SAMU-XLSR.

Salima Mdhaffar, Haroun Elleuch, Fethi Bougares et al.

Speech encoders pretrained through self-supervised learning (SSL) have demonstrated remarkable performance in various downstream tasks, including Spoken Language Understanding (SLU) and Automatic Speech Recognition (ASR). For instance, fine-tuning SSL models for such tasks has shown significant potential, leading to improvements in the SOTA performance across challenging datasets. In contrast to existing research, this paper contributes by comparing the effectiveness of SSL approaches in the context of (i) the low-resource spoken Tunisian Arabic dialect and (ii) its combination with a low-resource SLU and ASR scenario, where only a few semantic annotations are available for fine-tuning. We conduct experiments using many SSL speech encoders on the TARIC-SLU dataset. We use speech encoders that were pre-trained on either monolingual or multilingual speech data. Some of them have also been refined without in-domain nor Tunisian data through multimodal supervised teacher-student paradigm. This study yields numerous significant findings that we are discussing in this paper.

Ryan Whetten, Titouan Parcollet, Adel Moumen et al.

Self-Supervised Learning (SSL) has proven to be effective in various domains, including speech processing. However, SSL is computationally and memory expensive. This is in part due the quadratic complexity of multi-head self-attention (MHSA). Alternatives for MHSA have been proposed and used in the speech domain, but have yet to be investigated properly in an SSL setting. In this work, we study the effects of replacing MHSA with recent state-of-the-art alternatives that have linear complexity, namely, HyperMixing, Fastformer, SummaryMixing, and Mamba. We evaluate these methods by looking at the speed, the amount of VRAM consumed, and the performance on the SSL MP3S benchmark. Results show that these linear alternatives maintain competitive performance compared to MHSA while, on average, decreasing VRAM consumption by around 20% to 60% and increasing speed from 7% to 65% for input sequences ranging from 20 to 80 seconds.

Jarod Duret, Mickael Rouvier, Yannick Estève

In this work, we detail our submission to the 2024 edition of the MSP-Podcast Speech Emotion Recognition (SER) Challenge. This challenge is divided into two distinct tasks: Categorical Emotion Recognition and Emotional Attribute Prediction. We concentrated our efforts on Task 1, which involves the categorical classification of eight emotional states using data from the MSP-Podcast dataset. Our approach employs an ensemble of models, each trained independently and then fused at the score level using a Support Vector Machine (SVM) classifier. The models were trained using various strategies, including Self-Supervised Learning (SSL) fine-tuning across different modalities: speech alone, text alone, and a combined speech and text approach. This joint training methodology aims to enhance the system's ability to accurately classify emotional states. This joint training methodology aims to enhance the system's ability to accurately classify emotional states. Thus, the system obtained F1-macro of 0.35\% on development set.

Haroun Elleuch, Youssef Saidi, Salima Mdhaffar et al.

This paper describes Elyadata \& LIA's joint submission to the NADI multi-dialectal Arabic Speech Processing 2025. We participated in the Spoken Arabic Dialect Identification (ADI) and multi-dialectal Arabic ASR subtasks. Our submission ranked first for the ADI subtask and second for the multi-dialectal Arabic ASR subtask among all participants. Our ADI system is a fine-tuned Whisper-large-v3 encoder with data augmentation. This system obtained the highest ADI accuracy score of \textbf{79.83\%} on the official test set. For multi-dialectal Arabic ASR, we fine-tuned SeamlessM4T-v2 Large (Egyptian variant) separately for each of the eight considered dialects. Overall, we obtained an average WER and CER of \textbf{38.54\%} and \textbf{14.53\%}, respectively, on the test set. Our results demonstrate the effectiveness of large pre-trained speech models with targeted fine-tuning for Arabic speech processing.

Manon Macary, Marie Tahon, Yannick Estève et al.

The goal of our research is to automatically retrieve the satisfaction and the frustration in real-life call-center conversations. This study focuses an industrial application in which the customer satisfaction is continuously tracked down to improve customer services. To compensate the lack of large annotated emotional databases, we explore the use of pre-trained speech representations as a form of transfer learning towards AlloSat corpus. Moreover, several studies have pointed out that emotion can be detected not only in speech but also in facial trait, in biological response or in textual information. In the context of telephone conversations, we can break down the audio information into acoustic and linguistic by using the speech signal and its transcription. Our experiments confirms the large gain in performance obtained with the use of pre-trained features. Surprisingly, we found that the linguistic content is clearly the major contributor for the prediction of satisfaction and best generalizes to unseen data. Our experiments conclude to the definitive advantage of using CamemBERT representations, however the benefit of the fusion of acoustic and linguistic modalities is not as obvious. With models learnt on individual annotations, we found that fusion approaches are more robust to the subjectivity of the annotation task. This study also tackles the problem of performances variability and intends to estimate this variability from different views: weights initialization, confidence intervals and annotation subjectivity. A deep analysis on the linguistic content investigates interpretable factors able to explain the high contribution of the linguistic modality for this task.

Phuong-Hang Le, Valentin Pelloin, Arnault Chatelain et al.

We release Pantagruel models, a new family of self-supervised encoder models for French text and speech. Instead of predicting modality-tailored targets such as textual tokens or speech units, Pantagruel learns contextualized target representations in the feature space, allowing modality-specific encoders to capture linguistic and acoustic regularities more effectively. Separate models are pre-trained on large-scale French corpora, including Wikipedia, OSCAR and CroissantLLM for text, together with MultilingualLibriSpeech, LeBenchmark, and INA-100k for speech. INA-100k is a newly introduced 100,000-hour corpus of French audio derived from the archives of the Institut National de l'Audiovisuel (INA), the national repository of French radio and television broadcasts, providing highly diverse audio data. We evaluate Pantagruel across a broad range of downstream tasks spanning both modalities, including those from the standard French benchmarks such as FLUE or LeBenchmark. Across these tasks, Pantagruel models show competitive or superior performance compared to strong French baselines such as CamemBERT, FlauBERT, and LeBenchmark2.0, while maintaining a shared architecture that can seamlessly handle either speech or text inputs. These results confirm the effectiveness of feature-space self-supervised objectives for French representation learning and highlight Pantagruel as a robust foundation for multimodal speech-text understanding.

Ryan Whetten, Titouan Parcollet, Marco Dinarelli et al.

Self-Supervised Learning (SSL) has proven to be useful in various speech tasks. However, these methods are generally very demanding in terms of data, memory, and computational resources. BERT-based Speech pre-Training with Random-projection Quantizer (BEST-RQ), is an SSL method that has shown great performance on Automatic Speech Recognition (ASR) while being simpler than other SSL methods, such as wav2vec 2.0. Despite BEST-RQ's great performance, details are lacking in the original paper, such as the amount of GPU/TPU hours used in pre-training, and there is no official easy-to-use open-source implementation. Furthermore, BEST-RQ has not been evaluated on other downstream tasks aside from ASR and speech translation. In this work, we describe a re-implementation of a Random-projection quantizer and perform a preliminary study with a comparison to wav2vec 2.0 on four downstream tasks. We discuss the details and differences of our implementation. We show that a random projection quantizer can achieve similar downstream performance as wav2vec 2.0 while decreasing training time by over a factor of two.

Jarod Duret, Benjamin O'Brien, Yannick Estève et al.

Textless speech-to-speech translation systems are rapidly advancing, thanks to the integration of self-supervised learning techniques. However, existing state-of-the-art systems fall short when it comes to capturing and transferring expressivity accurately across different languages. Expressivity plays a vital role in conveying emotions, nuances, and cultural subtleties, thereby enhancing communication across diverse languages. To address this issue this study presents a novel method that operates at the discrete speech unit level and leverages multilingual emotion embeddings to capture language-agnostic information. Specifically, we demonstrate how these embeddings can be used to effectively predict the pitch and duration of speech units in the target language. Through objective and subjective experiments conducted on a French-to-English translation task, our findings highlight the superior expressivity transfer achieved by our approach compared to current state-of-the-art systems.

Jarod Duret, Yannick Estève, Titouan Parcollet

Recent advancements in textless speech-to-speech translation systems have been driven by the adoption of self-supervised learning techniques. Although most state-of-the-art systems adopt a similar architecture to transform source language speech into sequences of discrete representations in the target language, the criteria for selecting these target speech units remains an open question. This work explores the selection process through a study of downstream tasks such as automatic speech recognition, speech synthesis, speaker recognition, and emotion recognition. Interestingly, our findings reveal a discrepancy in the optimization of discrete speech units: units that perform well in resynthesis performance do not necessarily correlate with those that enhance translation efficacy. This discrepancy underscores the nuanced complexity of target feature selection and its impact on the overall performance of speech-to-speech translation systems.

Salima Mdhaffar, Haroun Elleuch, Chaimae Chellaf et al.

This paper introduces SENSE (Shared Embedding for N-lingual Speech and tExt), an open-source solution inspired by the SAMU-XLSR framework and conceptually similar to Meta AI's SONAR models. These approaches rely on a teacher-student framework to align a self-supervised speech encoder with the language-agnostic continuous representations of a text encoder at the utterance level. We describe how the original SAMU-XLSR method has been updated by selecting a stronger teacher text model and a better initial speech encoder. The source code for training and using SENSE models has been integrated into the SpeechBrain toolkit, and the first SENSE model we trained has been publicly released. We report experimental results on multilingual and multimodal semantic tasks, where our SENSE model achieves highly competitive performance. Finally, this study offers new insights into how semantics are captured in such semantically aligned speech encoders.

Tom Labiausse, Romain Fabre, Yannick Estève et al.

Simultaneous speech translation requires translating source speech into a target language in real-time while handling non-monotonic word dependencies. Traditional approaches rely on supervised training with word-level aligned data, which is difficult to collect at scale and thus depends on synthetic alignments using language-specific heuristics that are suboptimal. We propose Hibiki-Zero, which eliminates the need for word-level alignments entirely. This fundamentally simplifies the training pipeline and enables seamless scaling to diverse languages with varying grammatical structures, removing the bottleneck of designing language-specific alignment heuristics. We first train on sentence-level aligned data to learn speech translation at high latency, then apply a novel reinforcement learning strategy using GRPO to optimize latency while preserving translation quality. Hibiki-Zero achieves state-of-the-art performance in translation accuracy, latency, voice transfer, and naturalness across five X-to-English tasks. Moreover, we demonstrate that our model can be adapted to support a new input language with less than 1000h of speech. We provide examples, model weights, inference code and we release a benchmark containing 45h of multilingual data for speech translation evaluation.

Thibault Gaudier, Marie Tahon, Anthony Larcher et al.

Speech resynthesis is a generic task for which we want to synthesize audio with another audio as input, which finds applications for media monitors and journalists.Among different tasks addressed by speech resynthesis, voice conversion preserves the linguistic information while modifying the identity of the speaker, and speech edition preserves the identity of the speaker but some words are modified.In both cases, we need to disentangle speaker and phonetic contents in intermediate representations.Phonetic PosteriorGrams (PPG) are a frame-level probabilistic representation of phonemes, and are usually considered speaker-independent.This paper presents a PPG-based speech resynthesis system.A perceptive evaluation assesses that it produces correct audio quality.Then, we demonstrate that an automatic speaker verification model is not able to recover the source speaker after re-synthesis with PPG, even when the model is trained on synthetic data.

Lucas Druart, Valentin Vielzeuf, Yannick Estève

In Task-Oriented Dialogue (TOD) systems, correctly updating the system's understanding of the user's requests (\textit{a.k.a} dialogue state tracking) is key to a smooth interaction. Traditionally, TOD systems perform this update in three steps: transcription of the user's utterance, semantic extraction of the key concepts, and contextualization with the previously identified concepts. Such cascade approaches suffer from cascading errors and separate optimization. End-to-End approaches have been proven helpful up to the turn-level semantic extraction step. This paper goes one step further and provides (1) a novel approach for completely neural spoken DST, (2) an in depth comparison with a state of the art cascade approach and (3) avenues towards better context propagation. Our study highlights that jointly-optimized approaches are also competitive for contextually dependent tasks, such as Dialogue State Tracking (DST), especially in audio native settings. Context propagation in DST systems could benefit from training procedures accounting for the previous' context inherent uncertainty.

Chloé Sekkat, Fanny Leroy, Salima Mdhaffar et al.

Recent works demonstrate that voice assistants do not perform equally well for everyone, but research on demographic robustness of speech technologies is still scarce. This is mainly due to the rarity of large datasets with controlled demographic tags. This paper introduces the Sonos Voice Control Bias Assessment Dataset, an open dataset composed of voice assistant requests for North American English in the music domain (1,038 speakers, 166 hours, 170k audio samples, with 9,040 unique labelled transcripts) with a controlled demographic diversity (gender, age, dialectal region and ethnicity). We also release a statistical demographic bias assessment methodology, at the univariate and multivariate levels, tailored to this specific use case and leveraging spoken language understanding metrics rather than transcription accuracy, which we believe is a better proxy for user experience. To demonstrate the capabilities of this dataset and statistical method to detect demographic bias, we consider a pair of state-of-the-art Automatic Speech Recognition and Spoken Language Understanding models. Results show statistically significant differences in performance across age, dialectal region and ethnicity. Multivariate tests are crucial to shed light on mixed effects between dialectal region, gender and age.

Ryan Whetten, Lucas Maison, Titouan Parcollet et al.

In Self-Supervised Learning (SSL), pre-training and evaluation are resource intensive. In the speech domain, current indicators of the quality of SSL models during pre-training, such as the loss, do not correlate well with downstream performance. Consequently, it is often difficult to gauge the final downstream performance in a cost efficient manner during pre-training. In this work, we propose unsupervised efficient methods that give insights into the quality of the pre-training of SSL speech models, namely, measuring the cluster quality and rank of the embeddings of the SSL model. Results show that measures of cluster quality and rank correlate better with downstream performance than the pre-training loss with only one hour of unlabeled audio, reducing the need for GPU hours and labeled data in SSL model evaluation.

Jarod Duret, Salima Mdhaffar, Gaëlle Laperrière et al.

In this study, we investigate the benefits of domain-specific self-supervised pre-training for both offline and streaming ASR in Air Traffic Control (ATC) environments. We train BEST-RQ models on 4.5k hours of unlabeled ATC data, then fine-tune on a smaller supervised ATC set. To enable real-time processing, we propose using chunked attention and dynamic convolutions, ensuring low-latency inference. We compare these in-domain SSL models against state-of-the-art, general-purpose speech encoders such as w2v-BERT 2.0 and HuBERT. Results show that domain-adapted pre-training substantially improves performance on standard ATC benchmarks, significantly reducing word error rates when compared to models trained on broad speech corpora. Furthermore, the proposed streaming approach further improves word error rate under tighter latency constraints, making it particularly suitable for safety-critical aviation applications. These findings highlight that specializing SSL representations for ATC data is a practical path toward more accurate and efficient ASR systems in real-world operational settings.

Lucas Druart, Valentin Vielzeuf, Yannick Estève

In spoken Task-Oriented Dialogue (TOD) systems, the choice of the semantic representation describing the users' requests is key to a smooth interaction. Indeed, the system uses this representation to reason over a database and its domain knowledge to choose its next action. The dialogue course thus depends on the information provided by this semantic representation. While textual datasets provide fine-grained semantic representations, spoken dialogue datasets fall behind. This paper provides insights into automatic enhancement of spoken dialogue datasets' semantic representations. Our contributions are three fold: (1) assess the relevance of Large Language Model fine-tuning, (2) evaluate the knowledge captured by the produced annotations and (3) highlight semi-automatic annotation implications.

Gaëlle Laperrière, Sahar Ghannay, Bassam Jabaian et al.

Self-Supervised Learning is vastly used to efficiently represent speech for Spoken Language Understanding, gradually replacing conventional approaches. Meanwhile, textual SSL models are proposed to encode language-agnostic semantics. SAMU-XLSR framework employed this semantic information to enrich multilingual speech representations. A recent study investigated SAMU-XLSR in-domain semantic enrichment by specializing it on downstream transcriptions, leading to state-of-the-art results on a challenging SLU task. This study's interest lies in the loss of multilingual performances and lack of specific-semantics training induced by such specialization in close languages without any SLU implication. We also consider SAMU-XLSR's loss of initial cross-lingual abilities due to a separate SLU fine-tuning. Therefore, this paper proposes a dual task learning approach to improve SAMU-XLSR semantic enrichment while considering distant languages for multilingual and language portability experiments.

Marcely Zanon Boito, Fethi Bougares, Florentin Barbier et al.

In this paper we present two datasets for Tamasheq, a developing language mainly spoken in Mali and Niger. These two datasets were made available for the IWSLT 2022 low-resource speech translation track, and they consist of collections of radio recordings from daily broadcast news in Niger (Studio Kalangou) and Mali (Studio Tamani). We share (i) a massive amount of unlabeled audio data (671 hours) in five languages: French from Niger, Fulfulde, Hausa, Tamasheq and Zarma, and (ii) a smaller 17 hours parallel corpus of audio recordings in Tamasheq, with utterance-level translations in the French language. All this data is shared under the Creative Commons BY-NC-ND 3.0 license. We hope these resources will inspire the speech community to develop and benchmark models using the Tamasheq language.

Salima Mdhaffar, Jean-François Bonastre, Marc Tommasi et al.

The widespread of powerful personal devices capable of collecting voice of their users has opened the opportunity to build speaker adapted speech recognition system (ASR) or to participate to collaborative learning of ASR. In both cases, personalized acoustic models (AM), i.e. fine-tuned AM with specific speaker data, can be built. A question that naturally arises is whether the dissemination of personalized acoustic models can leak personal information. In this paper, we show that it is possible to retrieve the gender of the speaker, but also his identity, by just exploiting the weight matrix changes of a neural acoustic model locally adapted to this speaker. Incidentally we observe phenomena that may be useful towards explainability of deep neural networks in the context of speech processing. Gender can be identified almost surely using only the first layers and speaker verification performs well when using middle-up layers. Our experimental study on the TED-LIUM 3 dataset with HMM/TDNN models shows an accuracy of 95% for gender detection, and an Equal Error Rate of 9.07% for a speaker verification task by only exploiting the weights from personalized models that could be exchanged instead of user data.

Natalia Tomashenko, Salima Mdhaffar, Marc Tommasi et al.

This paper investigates methods to effectively retrieve speaker information from the personalized speaker adapted neural network acoustic models (AMs) in automatic speech recognition (ASR). This problem is especially important in the context of federated learning of ASR acoustic models where a global model is learnt on the server based on the updates received from multiple clients. We propose an approach to analyze information in neural network AMs based on a neural network footprint on the so-called Indicator dataset. Using this method, we develop two attack models that aim to infer speaker identity from the updated personalized models without access to the actual users' speech data. Experiments on the TED-LIUM 3 corpus demonstrate that the proposed approaches are very effective and can provide equal error rate (EER) of 1-2%.

Sahar Ghannay, Antoine Caubrière, Salima Mdhaffar et al.

Spoken language understanding (SLU) topic has seen a lot of progress these last three years, with the emergence of end-to-end neural approaches. Spoken language understanding refers to natural language processing tasks related to semantic extraction from speech signal, like named entity recognition from speech or slot filling task in a context of human-machine dialogue. Classically, SLU tasks were processed through a cascade approach that consists in applying, firstly, an automatic speech recognition process, followed by a natural language processing module applied to the automatic transcriptions. These three last years, end-to-end neural approaches, based on deep neural networks, have been proposed in order to directly extract the semantics from speech signal, by using a single neural model. More recent works on self-supervised training with unlabeled data open new perspectives in term of performance for automatic speech recognition and natural language processing. In this paper, we present a brief overview of the recent advances on the French MEDIA benchmark dataset for SLU, with or without the use of additional data. We also present our last results that significantly outperform the current state-of-the-art with a Concept Error Rate (CER) of 11.2%, instead of 13.6% for the last state-of-the-art system presented this year.

Ha Nguyen, Yannick Estève, Laurent Besacier

Boosted by the simultaneous translation shared task at IWSLT 2020, promising end-to-end online speech translation approaches were recently proposed. They consist in incrementally encoding a speech input (in a source language) and decoding the corresponding text (in a target language) with the best possible trade-off between latency and translation quality. This paper investigates two key aspects of end-to-end simultaneous speech translation: (a) how to encode efficiently the continuous speech flow, and (b) how to segment the speech flow in order to alternate optimally between reading (R: encoding input) and writing (W: decoding output) operations. We extend our previously proposed end-to-end online decoding strategy and show that while replacing BLSTM by ULSTM encoding degrades performance in offline mode, it actually improves both efficiency and performance in online mode. We also measure the impact of different methods to segment the speech signal (using fixed interval boundaries, oracle word boundaries or randomly set boundaries) and show that our best end-to-end online decoding strategy is surprisingly the one that alternates R/W operations on fixed size blocks on our English-German speech translation setup.

Ha Nguyen, Yannick Estève, Laurent Besacier

This paper proposes a decoding strategy for end-to-end simultaneous speech translation. We leverage end-to-end models trained in offline mode and conduct an empirical study for two language pairs (English-to-German and English-to-Portuguese). We also investigate different output token granularities including characters and Byte Pair Encoding (BPE) units. The results show that the proposed decoding approach allows to control BLEU/Average Lagging trade-off along different latency regimes. Our best decoding settings achieve comparable results with a strong cascade model evaluated on the simultaneous translation track of IWSLT 2020 shared task.

Valentin Pelloin, Nathalie Camelin, Antoine Laurent et al.

In this paper, we propose a novel end-to-end sequence-to-sequence spoken language understanding model using an attention mechanism. It reliably selects contextual acoustic features in order to hypothesize semantic contents. An initial architecture capable of extracting all pronounced words and concepts from acoustic spans is designed and tested. With a shallow fusion language model, this system reaches a 13.6 concept error rate (CER) and an 18.5 concept value error rate (CVER) on the French MEDIA corpus, achieving an absolute 2.8 points reduction compared to the state-of-the-art. Then, an original model is proposed for hypothesizing concepts and their values. This transduction reaches a 15.4 CER and a 21.6 CVER without any new type of context.

Manon Macary, Marie Tahon, Yannick Estève et al.

Pre-training for feature extraction is an increasingly studied approach to get better continuous representations of audio and text content. In the present work, we use wav2vec and camemBERT as self-supervised learned models to represent our data in order to perform continuous emotion recognition from speech (SER) on AlloSat, a large French emotional database describing the satisfaction dimension, and on the state of the art corpus SEWA focusing on valence, arousal and liking dimensions. To the authors' knowledge, this paper presents the first study showing that the joint use of wav2vec and BERT-like pre-trained features is very relevant to deal with continuous SER task, usually characterized by a small amount of labeled training data. Evaluated by the well-known concordance correlation coefficient (CCC), our experiments show that we can reach a CCC value of 0.825 instead of 0.592 when using MFCC in conjunction with word2vec word embedding on the AlloSat dataset.

Paul Tardy, Louis de Seynes, François Hernandez et al.

Supervised approaches for Neural Abstractive Summarization require large annotated corpora that are costly to build. We present a French meeting summarization task where reports are predicted based on the automatic transcription of the meeting audio recordings. In order to build a corpus for this task, it is necessary to obtain the (automatic or manual) transcription of each meeting, and then to segment and align it with the corresponding manual report to produce training examples suitable for training. On the other hand, we have access to a very large amount of unaligned data, in particular reports without corresponding transcription. Reports are professionally written and well formatted making pre-processing straightforward. In this context, we study how to take advantage of this massive amount of unaligned data using two approaches (i) self-supervised pre-training using a target-side denoising encoder-decoder model; (ii) back-summarization i.e. reversing the summarization process by learning to predict the transcription given the report, in order to align single reports with generated transcription, and use this synthetic dataset for further training. We report large improvements compared to the previous baseline (trained on aligned data only) for both approaches on two evaluation sets. Moreover, combining the two gives even better results, outperforming the baseline by a large margin of +6 ROUGE-1 and ROUGE-L and +5 ROUGE-2 on two evaluation sets

Paul Tardy, David Janiszek, Yannick Estève et al.

Summarizing texts is not a straightforward task. Before even considering text summarization, one should determine what kind of summary is expected. How much should the information be compressed? Is it relevant to reformulate or should the summary stick to the original phrasing? State-of-the-art on automatic text summarization mostly revolves around news articles. We suggest that considering a wider variety of tasks would lead to an improvement in the field, in terms of generalization and robustness. We explore meeting summarization: generating reports from automatic transcriptions. Our work consists in segmenting and aligning transcriptions with respect to reports, to get a suitable dataset for neural summarization. Using a bootstrapping approach, we provide pre-alignments that are corrected by human annotators, making a validation set against which we evaluate automatic models. This consistently reduces annotators' efforts by providing iteratively better pre-alignment and maximizes the corpus size by using annotations from our automatic alignment models. Evaluation is conducted on \publicmeetings, a novel corpus of aligned public meetings. We report automatic alignment and summarization performances on this corpus and show that automatic alignment is relevant for data annotation since it leads to large improvement of almost +4 on all ROUGE scores on the summarization task.

Maha Elbayad, Ha Nguyen, Fethi Bougares et al.

This paper describes the ON-TRAC Consortium translation systems developed for two challenge tracks featured in the Evaluation Campaign of IWSLT 2020, offline speech translation and simultaneous speech translation. ON-TRAC Consortium is composed of researchers from three French academic laboratories: LIA (Avignon Université), LIG (Université Grenoble Alpes), and LIUM (Le Mans Université). Attention-based encoder-decoder models, trained end-to-end, were used for our submissions to the offline speech translation track. Our contributions focused on data augmentation and ensembling of multiple models. In the simultaneous speech translation track, we build on Transformer-based wait-k models for the text-to-text subtask. For speech-to-text simultaneous translation, we attach a wait-k MT system to a hybrid ASR system. We propose an algorithm to control the latency of the ASR+MT cascade and achieve a good latency-quality trade-off on both subtasks.

Antoine Caubrière, Natalia Tomashenko, Antoine Laurent et al.

We present an end-to-end approach to extract semantic concepts directly from the speech audio signal. To overcome the lack of data available for this spoken language understanding approach, we investigate the use of a transfer learning strategy based on the principles of curriculum learning. This approach allows us to exploit out-of-domain data that can help to prepare a fully neural architecture. Experiments are carried out on the French MEDIA and PORTMEDIA corpora and show that this end-to-end SLU approach reaches the best results ever published on this task. We compare our approach to a classical pipeline approach that uses ASR, POS tagging, lemmatizer, chunker... and other NLP tools that aim to enrich ASR outputs that feed an SLU text to concepts system. Last, we explore the promising capacity of our end-to-end SLU approach to address the problem of domain portability.

Sahar Ghannay, Antoine Caubrière, Yannick Estève et al.

Named entity recognition (NER) is among SLU tasks that usually extract semantic information from textual documents. Until now, NER from speech is made through a pipeline process that consists in processing first an automatic speech recognition (ASR) on the audio and then processing a NER on the ASR outputs. Such approach has some disadvantages (error propagation, metric to tune ASR systems sub-optimal in regards to the final task, reduced space search at the ASR output level...) and it is known that more integrated approaches outperform sequential ones, when they can be applied. In this paper, we present a first study of end-to-end approach that directly extracts named entities from speech, though a unique neural architecture. On a such way, a joint optimization is able for both ASR and NER. Experiments are carried on French data easily accessible, composed of data distributed in several evaluation campaign. Experimental results show that this end-to-end approach provides better results (F-measure=0.69 on test data) than a classical pipeline approach to detect named entity categories (F-measure=0.65).

François Hernandez, Vincent Nguyen, Sahar Ghannay et al.

In this paper, we present TED-LIUM release 3 corpus dedicated to speech recognition in English, that multiplies by more than two the available data to train acoustic models in comparison with TED-LIUM 2. We present the recent development on Automatic Speech Recognition (ASR) systems in comparison with the two previous releases of the TED-LIUM Corpus from 2012 and 2014. We demonstrate that, passing from 207 to 452 hours of transcribed speech training data is really more useful for end-to-end ASR systems than for HMM-based state-of-the-art ones, even if the HMM-based ASR system still outperforms end-to-end ASR system when the size of audio training data is 452 hours, with respectively a Word Error Rate (WER) of 6.6% and 13.7%. Last, we propose two repartitions of the TED-LIUM release 3 corpus: the legacy one that is the same as the one existing in release 2, and a new one, calibrated and designed to make experiments on speaker adaptation. Like the two first releases, TED-LIUM 3 corpus will be freely available for the research community.

Edwin Simonnet, Sahar Ghannay, Nathalie Camelin et al.

This paper addresses the problem of automatic speech recognition (ASR) error detection and their use for improving spoken language understanding (SLU) systems. In this study, the SLU task consists in automatically extracting, from ASR transcriptions , semantic concepts and concept/values pairs in a e.g touristic information system. An approach is proposed for enriching the set of semantic labels with error specific labels and by using a recently proposed neural approach based on word embeddings to compute well calibrated ASR confidence measures. Experimental results are reported showing that it is possible to decrease significantly the Concept/Value Error Rate with a state of the art system, outperforming previously published results performance on the same experimental data. It also shown that combining an SLU approach based on conditional random fields with a neural encoder/decoder attention based architecture , it is possible to effectively identifying confidence islands and uncertain semantic output segments useful for deciding appropriate error handling actions by the dialogue manager strategy .