Recent work suggests that large language models (LLMs) can improve performance of speech tasks compared to existing systems. To support their claims, results on LibriSpeech and Common Voice are often quoted. However, this work finds that a substantial amount of the LibriSpeech and Common Voice evaluation sets appear in public LLM pretraining corpora. This calls into question the reliability of findings drawn from these two datasets. To measure contamination impact, LLMs trained with/without contamination are compared. A contaminated LLM is more likely to generate test sentences it has seen during training. Then, speech recognisers based on LLMs are compared. They show only subtle error rate differences if the LLM is contaminated, but assign significantly higher probabilities to transcriptions seen during LLM training. Results show that LLM outputs can be biased by tiny amounts of data contamination, highlighting the importance of evaluating LLM-based speech systems with held-out data.

Self-attention relies on positional embeddings to encode input order. Relative Position (RelPos) embeddings are widely used in Automatic Speech Recognition (ASR). However, RelPos has quadratic time complexity to input length and is often incompatible with fast GPU implementations of attention. In contrast, Rotary Positional Embedding (RoPE) rotates each input vector based on its absolute position, taking linear time to sequence length, implicitly encoding relative distances through self-attention dot products. Thus, it is usually compatible with efficient attention. However, its use in ASR remains underexplored. This work evaluates RoPE across diverse ASR tasks with training data ranging from 100 to 50,000 hours, covering various speech types (read, spontaneous, clean, noisy) and different accents in both streaming and non-streaming settings. ASR error rates are similar or better than RelPos, while training time is reduced by up to $21 \%$. Code is available via the SpeechBrain toolkit.

Automatic speech recognition (ASR) with an encoder equipped with self-attention, whether streaming or non-streaming, takes quadratic time in the length of the speech utterance. This slows down training and decoding, increase the cost, and limits the deployment of the ASR in constrained devices. SummaryMixing is a promising linear-time complexity alternative to self-attention for non-streaming speech recognition that, for the first time, preserves or outperforms the accuracy of self-attention models. Unfortunately, the original definition of SummaryMixing is not suited to streaming speech recognition. Hence, this work extends SummaryMixing to a Conformer Transducer that works in both a streaming and an offline mode. It shows that this new linear-time complexity speech encoder outperforms self-attention in both scenarios while requiring less compute and memory during training and decoding.

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.

SpeechBrain is an open-source Conversational AI toolkit based on PyTorch, focused particularly on speech processing tasks such as speech recognition, speech enhancement, speaker recognition, text-to-speech, and much more. It promotes transparency and replicability by releasing both the pre-trained models and the complete recipes of code and algorithms required for training them. This paper presents SpeechBrain 1.0, a significant milestone in the evolution of the toolkit, which now has over 200 recipes for speech, audio, and language processing tasks, and more than 100 models available on Hugging Face. SpeechBrain 1.0 introduces new technologies to support diverse learning modalities, Large Language Model (LLM) integration, and advanced decoding strategies, along with novel models, tasks, and modalities. It also includes a new benchmark repository, offering researchers a unified platform for evaluating models across diverse tasks.

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.

The light gated recurrent units (Li-GRU) is well-known for achieving excellent results in automatic speech recognition (ASR) tasks while being lighter and faster to train than a standard gated recurrent units (GRU). However, the unbounded nature of its rectified linear unit on the candidate recurrent gate induces a gradient exploding phenomenon disrupting the training process and preventing it from being applied to medium to large ASR datasets. In this paper, we theoretically and empirically derive the necessary conditions for its stability as well as engineering mechanisms to speed up by a factor of five its training time, hence introducing a novel version of this architecture named SLi-GRU. Then, we evaluate its performance both on a toy task illustrating its newly acquired capabilities and a set of three different ASR datasets demonstrating lower word error rates compared to more complex recurrent neural networks.

International audience

International audience

International audience

Despite impressive results, deep learning-based technologies also raise severe privacy and environmental concerns induced by the training procedure often conducted in data centers. In response, alternatives to centralized training such as Federated Learning (FL) have emerged. FL is now starting to be deployed at a global scale by companies that must adhere to new legal demands and policies originating from governments and social groups advocating for privacy protection. However, the potential environmental impact related to FL remains unclear and unexplored. This article offers the first-ever systematic study of the carbon footprint of FL. We propose a rigorous model to quantify the carbon footprint, hence facilitating the investigation of the relationship between FL design and carbon emissions. We also compare the carbon footprint of FL to traditional centralized learning. Our findings show that, depending on the configuration, FL can emit up to two orders of magnitude more carbon than centralized training. However, in certain settings, it can be comparable to centralized learning due to the reduced energy consumption of embedded devices. Finally, we highlight and connect the results to the future challenges and trends in FL to reduce its environmental impact, including algorithms efficiency, hardware capabilities, and stronger industry transparency.

Training Automatic Speech Recognition (ASR) models under federated learning (FL) settings has attracted a lot of attention recently. However, the FL scenarios often presented in the literature are artificial and fail to capture the complexity of real FL systems. In this paper, we construct a challenging and realistic ASR federated experimental setup consisting of clients with heterogeneous data distributions using the French and Italian sets of the CommonVoice dataset, a large heterogeneous dataset containing thousands of different speakers, acoustic environments and noises. We present the first empirical study on an attention-based sequence-to-sequence End-to-End (E2E) ASR model with three aggregation weighting strategies – standard FedAvg, loss-based aggregation and a novel word error rate (WER)-based aggregation, compared in two realistic FL scenarios: cross-silo with 10 clients and cross-device with 2K and 4K clients. This 4K cross-device ASR experiment is the largest ever performed. Our first-of-its-kind analysis on E2E ASR from heterogeneous and realistic federated acoustic models provides the foundations for future research and development of realistic FL ASR applications.

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.A 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.

International audience

Self-supervised learning (SSL) has proven vital in speech and audio-related applications. The paradigm trains a general model on unlabeled data that can later be used to solve specific downstream tasks. This type of model is costly to train as it requires manipulating long input sequences that can only be handled by powerful centralised servers. Surprisingly, despite many attempts to increase training efficiency through model compression, the effects of truncating input sequence lengths to reduce computation have not been studied. In this paper, we provide the first empirical study of SSL pre-training for different specified sequence lengths and link this to various downstream tasks. We find that training on short sequences can dramatically reduce resource costs while retaining a satisfactory performance for all tasks. This simple one-line change would promote the migration of SSL training from data centres to user-end edge devices for more realistic and personalised applications.

Knowledge distillation has been widely used to compress existing deep learning models while preserving the performance on a wide range of applications. In the specific context of Automatic Speech Recognition (ASR), distillation from ensembles of acoustic models has recently shown promising results in increasing recognition performance. In this paper, we propose an extension of multi-teacher distillation methods to joint CTC-attention end-to-end ASR systems. We also introduce three novel distillation strategies. The core intuition behind them is to integrate the error rate metric to the teacher selection rather than solely focusing on the observed losses. In this way, we directly distill and optimize the student toward the relevant metric for speech recognition. We evaluate these strategies under a selection of training procedures on different datasets (TIMIT, Librispeech, Common Voice) and various languages (English, French, Italian). In particular, state-of-the-art error rates are reported on the Common Voice French, Italian and TIMIT datasets.

Self-Supervised Learning (SSL) using huge unlabeled data has been successfully explored for image and natural language processing. Recent works also investigated SSL from speech. They were notably successful to improve performance on downstream tasks such as automatic speech recognition (ASR). While these works suggest it is possible to reduce dependence on labeled data for building efficient speech systems, their evaluation was mostly made on ASR and using multiple and heterogeneous experimental settings (most of them for English). This questions the objective comparison of SSL approaches and the evaluation of their impact on building speech systems. In this paper, we propose LeBenchmark: a reproducible framework for assessing SSL from speech. It not only includes ASR (high and low resource) tasks but also spoken language understanding, speech translation and emotion recognition. We also focus on speech technologies in a language different than English: French. SSL models of different sizes are trained from carefully sourced and documented datasets. Experiments show that SSL is beneficial for most but not all tasks which confirms the need for exhaustive and reliable benchmarks to evaluate its real impact. LeBenchmark is shared with the scientific community for reproducible research in SSL from speech.

In speech technologies, speaker's voice representation is used in many applications such as speech recognition, voice conversion, speech synthesis and, obviously, user authentication. Modern vocal representations of the speaker are based on neural embeddings. In addition to the targeted information, these representations usually contain sensitive information about the speaker, like the age, sex, physical state, education level or ethnicity. In order to allow the user to choose which information to protect, we introduce in this paper the concept of attribute-driven privacy preservation in speaker voice representation. It allows a person to hide one or more personal aspects to a potential malicious interceptor and to the application provider. As a first solution to this concept, we propose to use an adversarial autoencoding method that disentangles in the voice representation a given speaker attribute thus allowing its concealment. We focus here on the sex attribute for an Automatic Speaker Verification (ASV) task. Experiments carried out using the VoxCeleb datasets have shown that the proposed method enables the concealment of this attribute while preserving ASV ability.

Deep learning contributes to reaching higher levels of artificial intelligence. Due to its pervasive adoption, however, growing concerns on the environmental impact of this technology have been raised. In particular, the energy consumed at training and inference time by modern neural networks is far from being negligible and will increase even further due to the deployment of ever larger models. This work investigates for the first time the carbon cost of end-to-end automatic speech recognition (ASR). First, it quantifies the amount of CO2 emitted while training state-of-the-art (SOTA) ASR systems on a university-scale cluster. Then, it shows that a tiny performance improvement comes at an extremely high carbon cost. For instance, the conducted experiments reveal that a SOTA Transformer emits 50% of its total training released CO2 solely to achieve a final decrease of 0.3 of the word error rate. With this study, we hope to raise awareness on this crucial topic and we provide guidelines, insights, and estimates enabling researchers to better assess the environmental impact of training speech technologies.

Through solving pretext tasks, self-supervised learning (SSL) leverages unlabeled data to extract useful latent representations replacing traditional input features in the downstream task. A common pretext task consists in pretraining a SSL model on pseudo-labels derived from the original signal. This technique is particularly relevant for speech data where various meaningful signal processing features may serve as pseudo-labels. However, the process of selecting pseudo-labels, for speech or other types of data, remains mostly unexplored and currently relies on observing the results on the final downstream task. Nevertheless, this methodology is not sustainable at scale due to substantial computational (hence carbon) costs. Thus, this paper introduces a practical and theoretical framework to select relevant pseudo-labels with respect to a given downstream task. More precisely, we propose a functional estimator of the pseudo-label utility grounded in the conditional independence theory, which does not require any training. The experiments conducted on speaker recognition and automatic speech recognition validate our estimator, showing a significant correlation between the performance observed on the downstream task and the utility estimates obtained with our approach, facilitating the prospection of relevant pseudo-labels for self-supervised speech representation learning.

Convolutional Neural Networks (CNN) have been used in Automatic Speech Recognition (ASR) to learn representations directly from the raw signal instead of hand-crafted acoustic features, providing a richer and lossless input signal. Recent researches propose to inject prior acoustic knowledge to the first convolutional layer by integrating the shape of the impulse responses in order to increase both the interpretability of the learnt acoustic model, and its performances. We propose to combine the complex Gabor filter with complex-valued deep neural networks to replace usual CNN weights kernels, to fully take advantage of its optimal time-frequency resolution and of the complex domain. The conducted experiments on the TIMIT phoneme recognition task shows that the proposed approach reaches top-of-the-line performances while remaining interpretable.

Modern end-to-end (E2E) Automatic Speech Recognition (ASR) systems rely on Deep Neural Networks (DNN) that are mostly trained on handcrafted and pre-computed acoustic features such as Mel-filter-banks or Mel-frequency cepstral coefficients. Nonetheless, and despite worse performances, E2E ASR models processing raw waveforms are an active research field due to the lossless nature of the input signal. In this paper, we propose the E2E-SincNet, a novel fully E2E ASR model that goes from the raw waveform to the text transcripts by merging two recent and powerful paradigms: SincNet and the joint CTC-attention training scheme. The conducted experiments on two different speech recognition tasks show that our approach outperforms previously investigated E2E systems relying either on the raw waveform or pre-computed acoustic features, with a reported top-of-the-line Word Error Rate (WER) of 4.7% on the Wall Street Journal (WSJ) dataset.

International audience

Machine learning (ML) and deep learning with deep neural networks (DNN), have drastically improved the performances of modern systems on numerous spoken language understanding (SLU) related tasks. Since most of current researches focus on new neural architectures to enhance the performances in realistic conditions, few recent works investigated the use of different algebras with neural networks (NN), to better represent the nature of the data being processed. To this extent, quaternion-valued neural networks (QNN) have shown better performances, and an important reduction of the number of neural parameters compared to traditional real-valued neural networks, when dealing with multidimensional signal. Nonetheless, the use of QNNs is strictly limited to quaternion input or output features. This article introduces a new unsupervised method based on a hybrid autoencoder (AE) called real-to-quaternion autoencoder (R2H), to extract a quaternion-valued input signal from any real-valued data, to be processed by QNNs. The experiments performed to identify the most related theme of a given telephone conversation from a customer care service (CCS), demonstrate that the R2H approach outperforms all the previously established models, either real- or quaternion-valued ones, in term of accuracy and with up to four times fewer neural parameters.

Convolutional neural networks (CNN) have recently achieved state-of-the-art results in various applications. In the case of image recognition, an ideal model has to learn independently of the training data, both local dependencies between the three components (R,G,B) of a pixel, and the global relations describing edges or shapes, making it efficient with small or heterogeneous datasets. Quaternion-valued convolutional neural networks (QCNN) solved this problematic by introducing multidimensional algebra to CNN. This paper proposes to explore the fundamental reason of the success of QCNN over CNN, by investigating the impact of the Hamilton product on a color image reconstruction task performed from a gray-scale only training. By learning independently both internal and external relations and with less parameters than real valued convolutional encoder-decoder (CAE), quaternion convolutional encoder-decoders (QCAE) perfectly reconstructed unseen color images while CAE produced worst and gray-scale versions.

Recurrent neural networks (RNN) are at the core of modern automatic speech recognition (ASR) systems. In particular, long short-term memory (LSTM) recurrent neural networks have achieved state-of-the-art results in many speech recognition tasks, due to their efficient representation of long and short term dependencies in sequences of inter-dependent features. Nonetheless, internal dependencies within the element composing multidimensional features are weakly considered by traditional real-valued representations. We propose a novel quaternion long short-term memory (QL-STM) recurrent neural network that takes into account both the external relations between the features composing a sequence, and these internal latent structural dependencies with the quaternion algebra. QLSTMs are compared to LSTMs during a memory copy-task and a realistic application of speech recognition on the Wall Street Journal (WSJ) dataset. QLSTM reaches better performances during the two experiments with up to 2.8 times less learning parameters, leading to a more expressive representation of the information.

The availability of open-source software is playing a remarkable role in the popularization of speech recognition and deep learning. Kaldi, for instance, is nowadays an established framework used to develop state-of-the-art speech recognizers. PyTorch is used to build neural networks with the Python language and has recently spawn tremendous interest within the machine learning community thanks to its simplicity and flexibility. The PyTorch-Kaldi project aims to bridge the gap between these popular toolkits, trying to inherit the efficiency of Kaldi and the flexibility of PyTorch. PyTorch-Kaldi is not only a simple interface between these software, but it embeds several useful features for developing modern speech recognizers. For instance, the code is specifically designed to naturally plug-in user-defined acoustic models. As an alternative, users can exploit several pre-implemented neural networks that can be customized using intuitive configuration files. PyTorch-Kaldi supports multiple feature and label streams as well as combinations of neural networks, enabling the use of complex neural architectures. The toolkit is publicly-released along with a rich documentation and is designed to properly work locally or on HPC clusters. Experiments, that are conducted on several datasets and tasks, show that PyTorch-Kaldi can effectively be used to develop modern state-of-the-art speech recognizers.

International audience

Deep neural networks (DNNs) and more precisely recurrent neural networks (RNNs) are at the core of modern automatic speech recognition systems, due to their efficiency to process input sequences. Recently, it has been shown that different input representations, based on multidimensional algebras, such as complex and quaternion numbers, are able to bring to neural networks a more natural, compressive and powerful representation of the input signal by outperforming common real-valued NNs. Indeed, quaternion-valued neural networks (QNNs) better learn both internal dependencies, such as the relation between the Mel-filter-bank value of a specific time frame and its time derivatives, and global dependencies, describing the relations that exist between time frames. Nonetheless, QNNs are limited to quaternion-valued input signals, and it is difficult to benefit from this powerful representation with real-valued input data. This paper proposes to tackle this weakness by introducing a real-to-quaternion encoder that allows QNNs to process any one dimensional input features, such as traditional Mel-filter-banks for automatic speech recognition.

Recently, the connectionist temporal classification (CTC) model coupled with recurrent (RNN) or convolutional neural networks (CNN), made it easier to train speech recognition systems in an end-to-end fashion. However in real-valued models, time frame components such as mel-filter-bank energies and the cepstral coefficients obtained from them, together with their first and second order derivatives, are processed as individual elements, while a natural alternative is to process such components as composed entities. We propose to group such elements in the form of quaternions and to process these quaternions using the established quaternion algebra. Quaternion numbers and quaternion neural networks have shown their efficiency to process multidimensional inputs as entities, to encode internal dependencies, and to solve many tasks with less learning parameters than real-valued models. This paper proposes to integrate multiple feature views in quaternion-valued convolutional neural network (QCNN), to be used for sequence-to-sequence mapping with the CTC model. Promising results are reported using simple QCNNs in phoneme recognition experiments with the TIMIT corpus. More precisely, QCNNs obtain a lower phoneme error rate (PER) with less learning parameters than a competing model based on real-valued CNNs.

Quaternion convolutional neural networks (QCNN) are powerful architectures to learn and model external dependencies that exist between neighbor features of an input vector, and internal latent dependencies within the feature. This paper proposes to evaluate the effectiveness of the QCNN on a realistic theme identification task of spoken telephone conversations between agents and customers from the call center of the Paris transportation system (RATP). We show that QCNNs are more suitable than real-valued CNN to process multidimensional data and to code internal dependencies. Indeed, real-valued CNNs deal with both internal and external relations at the same level since components of an entity are processed independently. Experimental evidence is provided that the proposed QCNN architecture always outperforms real-valued equivalent CNN models in the theme identification task of the DECODA corpus. It is also shown that QCNN accuracy results are the best achieved so far on this task, while reducing by a factor of 4 the number of model parameters.

Deep Neural Networks (DNN) received a great interest from researchers due to their capability to construct robust abstract representations of heterogeneous documents in a latent subspace. Nonetheless, mere real-valued deep neural networks require an appropriate adaptation, such as the convolution process, to capture latent relations between input features. Moreover, real-valued deep neural networks reveal little in way of document internal dependencies, by only considering words or topics contained in the document as an isolate basic element. Quaternion-valued multi-layer perceptrons (QMLP), and autoencoders (QAE) have been introduced to capture such latent dependencies, alongside to represent multidimensional data. Nonetheless, a three-layered neural network does not benefit from the high abstraction capability of DNNs. The paper proposes first to extend the hyper-complex algebra to deep neural networks (QDNN) and, then, introduces pre-trained deep quaternion neural networks (QDNN-AE) with dedicated quaternion encoder-decoders (QAE). The experiments conduced on a theme identification task of spoken dialogues from the DECODA data set show, inter alia, that the QDNN-AE reaches a promising gain of 2.2% compared to the standard real-valued DNN-AE.

International audience

Automatically translating textual documents from one language to another inevitably results in translation errors. In addition to language specificities, this automatic translation appears more difficult in the context of spoken dialogues since, for example, the language register is far from “clean speech”. Speech analytics suffer from these translation errors. To tackle this difficulty, a solution consists in mapping translations into a space of hidden topics. In the classical topic-based representation obtained from a Latent Dirichlet Allocation (LDA), distribution of words into each topic is estimated automatically. Nonetheless, the targeted classes are ignored in the particular context of a classification task. In the DSTC5 main task, this targeted class information is crucial, the main objective being to track dialog states for sub-dialog segments. For this challenge, we propose to apply an original topic-based representation for each sub-dialogue based not only on the sub-dialogue content itself (words), but also on the dialogue state related to the sub-dialogue. This original representation is based on the Author-Topic (AT) model, previously successfully applied on a different classification task. Promising results confirmed the interest of such a method, the AT model reaching performance slightly better in terms of F-measure than baseline ones given by the task's organizers.

Machine Learning (ML) techniques have allowed a great performance improvement of different challenging Spoken Language Understanding (SLU) tasks. Among these methods, Neural Networks (NN), or Multilayer Perceptron (MLP), recently received a great interest from researchers due to their representation capability of complex internal structures in a low dimensional subspace. However, MLPs employ document representations based on basic word level or topic-based features. Therefore, these basic representations reveal little in way of document statistical structure by only considering words or topics contained in the document as a “bag-of-words”, ignoring relations between them. We propose to remedy this weakness by extending the complex features based on Quaternion algebra presented in [1] to neural networks called QMLP. This original QMLP approach is based on hyper-complex algebra to take into consideration features dependencies in documents. New document features, based on the document structure itself, used as input of the QMLP, are also investigated in this paper, in comparison to those initially proposed in [1]. Experiments made on a SLU task from a real framework of human spoken dialogues showed that our QMLP approach associated with the proposed document features outperforms other approaches, with an accuracy gain of 2% with respect to the MLP based on real numbers and more than 3% with respect to the first Quaternion-based features proposed in [1]. We finally demonstrated that less iterations are needed by our QMLP architecture to be efficient and to reach promising accuracies.