Neil Zeghidour

Chris Donahue, Antoine Caillon, Adam Roberts et al. · stanford

We present SingSong, a system that generates instrumental music to accompany input vocals, potentially offering musicians and non-musicians alike an intuitive new way to create music featuring their own voice. To accomplish this, we build on recent developments in musical source separation and audio generation. Specifically, we apply a state-of-the-art source separation algorithm to a large corpus of music audio to produce aligned pairs of vocals and instrumental sources. Then, we adapt AudioLM (Borsos et al., 2022) -- a state-of-the-art approach for unconditional audio generation -- to be suitable for conditional "audio-to-audio" generation tasks, and train it on the source-separated (vocal, instrumental) pairs. In a pairwise comparison with the same vocal inputs, listeners expressed a significant preference for instrumentals generated by SingSong compared to those from a strong retrieval baseline. Sound examples at https://g.co/magenta/singsong

Hakan Erdogan, Scott Wisdom, Xuankai Chang et al. · deepmind

We present TokenSplit, a speech separation model that acts on discrete token sequences. The model is trained on multiple tasks simultaneously: separate and transcribe each speech source, and generate speech from text. The model operates on transcripts and audio token sequences and achieves multiple tasks through masking of inputs. The model is a sequence-to-sequence encoder-decoder model that uses the Transformer architecture. We also present a "refinement" version of the model that predicts enhanced audio tokens from the audio tokens of speech separated by a conventional separation model. Using both objective metrics and subjective MUSHRA listening tests, we show that our model achieves excellent performance in terms of separation, both with or without transcript conditioning. We also measure the automatic speech recognition (ASR) performance and provide audio samples of speech synthesis to demonstrate the additional utility of our model.

Paul K. Rubenstein, Chulayuth Asawaroengchai, Duc Dung Nguyen et al.

We introduce AudioPaLM, a large language model for speech understanding and generation. AudioPaLM fuses text-based and speech-based language models, PaLM-2 [Anil et al., 2023] and AudioLM [Borsos et al., 2022], into a unified multimodal architecture that can process and generate text and speech with applications including speech recognition and speech-to-speech translation. AudioPaLM inherits the capability to preserve paralinguistic information such as speaker identity and intonation from AudioLM and the linguistic knowledge present only in text large language models such as PaLM-2. We demonstrate that initializing AudioPaLM with the weights of a text-only large language model improves speech processing, successfully leveraging the larger quantity of text training data used in pretraining to assist with the speech tasks. The resulting model significantly outperforms existing systems for speech translation tasks and has the ability to perform zero-shot speech-to-text translation for many languages for which input/target language combinations were not seen in training. AudioPaLM also demonstrates features of audio language models, such as transferring a voice across languages based on a short spoken prompt. We release examples of our method at https://google-research.github.io/seanet/audiopalm/examples

Zalán Borsos, Raphaël Marinier, Damien Vincent et al.

We introduce AudioLM, a framework for high-quality audio generation with long-term consistency. AudioLM maps the input audio to a sequence of discrete tokens and casts audio generation as a language modeling task in this representation space. We show how existing audio tokenizers provide different trade-offs between reconstruction quality and long-term structure, and we propose a hybrid tokenization scheme to achieve both objectives. Namely, we leverage the discretized activations of a masked language model pre-trained on audio to capture long-term structure and the discrete codes produced by a neural audio codec to achieve high-quality synthesis. By training on large corpora of raw audio waveforms, AudioLM learns to generate natural and coherent continuations given short prompts. When trained on speech, and without any transcript or annotation, AudioLM generates syntactically and semantically plausible speech continuations while also maintaining speaker identity and prosody for unseen speakers. Furthermore, we demonstrate how our approach extends beyond speech by generating coherent piano music continuations, despite being trained without any symbolic representation of music.

Andrea Agostinelli, Timo I. Denk, Zalán Borsos et al.

We introduce MusicLM, a model generating high-fidelity music from text descriptions such as "a calming violin melody backed by a distorted guitar riff". MusicLM casts the process of conditional music generation as a hierarchical sequence-to-sequence modeling task, and it generates music at 24 kHz that remains consistent over several minutes. Our experiments show that MusicLM outperforms previous systems both in audio quality and adherence to the text description. Moreover, we demonstrate that MusicLM can be conditioned on both text and a melody in that it can transform whistled and hummed melodies according to the style described in a text caption. To support future research, we publicly release MusicCaps, a dataset composed of 5.5k music-text pairs, with rich text descriptions provided by human experts.

Alexandre Défossez, Laurent Mazaré, Manu Orsini et al.

We introduce Moshi, a speech-text foundation model and full-duplex spoken dialogue framework. Current systems for spoken dialogue rely on pipelines of independent components, namely voice activity detection, speech recognition, textual dialogue and text-to-speech. Such frameworks cannot emulate the experience of real conversations. First, their complexity induces a latency of several seconds between interactions. Second, text being the intermediate modality for dialogue, non-linguistic information that modifies meaning -- such as emotion or non-speech sounds -- is lost in the interaction. Finally, they rely on a segmentation into speaker turns, which does not take into account overlapping speech, interruptions and interjections. Moshi solves these independent issues altogether by casting spoken dialogue as speech-to-speech generation. Starting from a text language model backbone, Moshi generates speech as tokens from the residual quantizer of a neural audio codec, while modeling separately its own speech and that of the user into parallel streams. This allows for the removal of explicit speaker turns, and the modeling of arbitrary conversational dynamics. We moreover extend the hierarchical semantic-to-acoustic token generation of previous work to first predict time-aligned text tokens as a prefix to audio tokens. Not only this "Inner Monologue" method significantly improves the linguistic quality of generated speech, but we also illustrate how it can provide streaming speech recognition and text-to-speech. Our resulting model is the first real-time full-duplex spoken large language model, with a theoretical latency of 160ms, 200ms in practice, and is available at https://github.com/kyutai-labs/moshi.

Curtis Hawthorne, Ian Simon, Adam Roberts et al.

An ideal music synthesizer should be both interactive and expressive, generating high-fidelity audio in realtime for arbitrary combinations of instruments and notes. Recent neural synthesizers have exhibited a tradeoff between domain-specific models that offer detailed control of only specific instruments, or raw waveform models that can train on any music but with minimal control and slow generation. In this work, we focus on a middle ground of neural synthesizers that can generate audio from MIDI sequences with arbitrary combinations of instruments in realtime. This enables training on a wide range of transcription datasets with a single model, which in turn offers note-level control of composition and instrumentation across a wide range of instruments. We use a simple two-stage process: MIDI to spectrograms with an encoder-decoder Transformer, then spectrograms to audio with a generative adversarial network (GAN) spectrogram inverter. We compare training the decoder as an autoregressive model and as a Denoising Diffusion Probabilistic Model (DDPM) and find that the DDPM approach is superior both qualitatively and as measured by audio reconstruction and Fréchet distance metrics. Given the interactivity and generality of this approach, we find this to be a promising first step towards interactive and expressive neural synthesis for arbitrary combinations of instruments and notes.

Ahmed Omran, Neil Zeghidour, Zalán Borsos et al.

We present a method to separate speech signals from noisy environments in the embedding space of a neural audio codec. We introduce a new training procedure that allows our model to produce structured encodings of audio waveforms given by embedding vectors, where one part of the embedding vector represents the speech signal, and the rest represent the environment. We achieve this by partitioning the embeddings of different input waveforms and training the model to faithfully reconstruct audio from mixed partitions, thereby ensuring each partition encodes a separate audio attribute. As use cases, we demonstrate the separation of speech from background noise or from reverberation characteristics. Our method also allows for targeted adjustments of the audio output characteristics.

Sarthak Yadav, Neil Zeghidour

Deep audio classification, traditionally cast as training a deep neural network on top of mel-filterbanks in a supervised fashion, has recently benefited from two independent lines of work. The first one explores "learnable frontends", i.e., neural modules that produce a learnable time-frequency representation, to overcome limitations of fixed features. The second one uses self-supervised learning to leverage unprecedented scales of pre-training data. In this work, we study the feasibility of combining both approaches, i.e., pre-training learnable frontend jointly with the main architecture for downstream classification. First, we show that pretraining two previously proposed frontends (SincNet and LEAF) on Audioset drastically improves linear-probe performance over fixed mel-filterbanks, suggesting that learnable time-frequency representations can benefit self-supervised pre-training even more than supervised training. Surprisingly, randomly initialized learnable filterbanks outperform mel-scaled initialization in the self-supervised setting, a counter-intuitive result that questions the appropriateness of strong priors when designing learnable filters. Through exploratory analysis of the learned frontend components, we uncover crucial differences in properties of these frontends when used in a supervised and self-supervised setting, especially the affinity of self-supervised filters to diverge significantly from the mel-scale to model a broader range of frequencies.

David W. Romero, Neil Zeghidour

We present Differentiable Neural Architectures (DNArch), a method that jointly learns the weights and the architecture of Convolutional Neural Networks (CNNs) by backpropagation. In particular, DNArch allows learning (i) the size of convolutional kernels at each layer, (ii) the number of channels at each layer, (iii) the position and values of downsampling layers, and (iv) the depth of the network. To this end, DNArch views neural architectures as continuous multidimensional entities, and uses learnable differentiable masks along each dimension to control their size. Unlike existing methods, DNArch is not limited to a predefined set of possible neural components, but instead it is able to discover entire CNN architectures across all feasible combinations of kernel sizes, widths, depths and downsampling. Empirically, DNArch finds performant CNN architectures for several classification and dense prediction tasks on sequential and image data. When combined with a loss term that controls the network complexity, DNArch constrains its search to architectures that respect a predefined computational budget during training.

Tsai-Ning Wang, Lin-Lin Chen, Neil Zeghidour et al.

Pre-trained audio models excel at detecting acoustic patterns in auscultation sounds but often fail to grasp their clinical significance, limiting their use and performance in diagnostic tasks. To bridge this gap, we introduce AcuLa (Audio-Clinical Understanding via Language Alignment), a lightweight post-training framework that instills semantic understanding into any audio encoder by aligning it with a medical language model, which acts as a "semantic teacher." To enable alignment at scale, we construct a large-scale dataset by leveraging off-the-shelf large language models to translate the rich, structured metadata accompanying existing audio recordings into coherent clinical reports. Our alignment strategy combines a representation-level contrastive objective with a self-supervised modeling, ensuring that the model learns clinical semantics while preserving fine-grained temporal cues. AcuLa achieves state-of-the-art results across 18 diverse cardio-respiratory tasks from 10 different datasets, improving the mean AUROC on classification benchmarks from 0.68 to 0.79 and, on the most challenging COVID-19 cough detection task, boosting the AUROC from 0.55 to 0.89. Our work demonstrates that this audio-language alignment transforms purely acoustic models into clinically-aware diagnostic tools, establishing a novel paradigm for enhancing physiological understanding in audio-based health monitoring.

Chung-Ming Chien, Manu Orsini, Eugene Kharitonov et al.

Speech-to-speech language models have recently emerged to enhance the naturalness of conversational AI. In particular, full-duplex models are distinguished by their real-time interactivity, including handling of pauses, interruptions, and backchannels. However, improving their factuality remains an open challenge. While scaling the model size could address this gap, it would make real-time inference prohibitively expensive. In this work, we propose MoshiRAG, a modular approach that combines a compact full-duplex interface with selective retrieval to access more powerful knowledge sources. Our asynchronous framework enables the model to identify knowledge-demanding queries and ground its responses in external information. By leveraging the natural temporal gap between response onset and the delivery of core information, the retrieval process can be completed while maintaining a natural conversation flow. With this approach, MoshiRAG achieves factuality comparable to the best publicly released non-duplex speech language models while preserving the interactivity inherent to full-duplex systems. Moreover, our flexible design supports plug-and-play retrieval methods without retraining and demonstrates strong performance on out-of-domain mathematical reasoning tasks.

Neil Zeghidour, Eugene Kharitonov, Manu Orsini et al.

We introduce Delayed Streams Modeling (DSM), a flexible formulation for streaming, multimodal sequence-to-sequence learning. Sequence-to-sequence generation is often cast in an offline manner, where the model consumes the complete input sequence before generating the first output timestep. Alternatively, streaming sequence-to-sequence rely on learning a policy for choosing when to advance on the input stream, or write to the output stream. DSM instead models already time-aligned streams with a decoder-only language model. By moving the alignment to a pre-processing step,and introducing appropriate delays between streams, DSM provides streaming inference of arbitrary output sequences, from any input combination, making it applicable to many sequence-to-sequence problems. In particular, given text and audio streams, automatic speech recognition (ASR) corresponds to the text stream being delayed, while the opposite gives a text-to-speech (TTS) model. We perform extensive experiments for these two major sequence-to-sequence tasks, showing that DSM provides state-of-the-art performance and latency while supporting arbitrary long sequences, being even competitive with offline baselines. Code, samples and demos are available at https://github.com/kyutai-labs/delayed-streams-modeling

Tsai-Ning Wang, Herman Teun den Dekker, Lin-Lin Chen et al.

Automated respiratory audio analysis promises scalable, non-invasive disease screening, yet progress is limited by scarce labeled data and costly expert annotation. Zero-shot inference eliminates task-specific supervision, but existing methods apply uniform computation to every input regardless of difficulty. We introduce TRIAGE, a tiered zero-shot framework that adaptively scales test-time compute by routing each audio sample through progressively richer reasoning stages: fast label-cosine scoring in a joint audio-text embedding space (Tier-L), structured matching with clinician-style descriptors (Tier-M), and retrieval-augmented large language model reasoning (Tier-H). A confidence-based router finalizes easy predictions early while allocating additional computation to ambiguous inputs, enabling nearly half of all samples to exit at the cheapest tier. Across nine respiratory classification tasks without task-specific training, TRIAGE achieves a mean AUROC of 0.744, outperforming prior zero-shot methods and matching or exceeding supervised baselines on multiple tasks. Our analysis show that test-time scaling concentrates gains where they matter: uncertain cases see up to 19% relative improvement while confident predictions remain unchanged at minimal cost.

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.

Geoffrey Cideron, Sertan Girgin, Mauro Verzetti et al.

We propose MusicRL, the first music generation system finetuned from human feedback. Appreciation of text-to-music models is particularly subjective since the concept of musicality as well as the specific intention behind a caption are user-dependent (e.g. a caption such as "upbeat work-out music" can map to a retro guitar solo or a techno pop beat). Not only this makes supervised training of such models challenging, but it also calls for integrating continuous human feedback in their post-deployment finetuning. MusicRL is a pretrained autoregressive MusicLM (Agostinelli et al., 2023) model of discrete audio tokens finetuned with reinforcement learning to maximise sequence-level rewards. We design reward functions related specifically to text-adherence and audio quality with the help from selected raters, and use those to finetune MusicLM into MusicRL-R. We deploy MusicLM to users and collect a substantial dataset comprising 300,000 pairwise preferences. Using Reinforcement Learning from Human Feedback (RLHF), we train MusicRL-U, the first text-to-music model that incorporates human feedback at scale. Human evaluations show that both MusicRL-R and MusicRL-U are preferred to the baseline. Ultimately, MusicRL-RU combines the two approaches and results in the best model according to human raters. Ablation studies shed light on the musical attributes influencing human preferences, indicating that text adherence and quality only account for a part of it. This underscores the prevalence of subjectivity in musical appreciation and calls for further involvement of human listeners in the finetuning of music generation models.

Tom Labiausse, Laurent Mazaré, Edouard Grave et al.

We introduce Hibiki, a decoder-only model for simultaneous speech translation. Hibiki leverages a multistream language model to synchronously process source and target speech, and jointly produces text and audio tokens to perform speech-to-text and speech-to-speech translation. We furthermore address the fundamental challenge of simultaneous interpretation, which unlike its consecutive counterpart, where one waits for the end of the source utterance to start translating, adapts its flow to accumulate just enough context to produce a correct translation in real-time, chunk by chunk. To do so, we introduce a weakly-supervised method that leverages the perplexity of an off-the-shelf text translation system to identify optimal delays on a per-word basis and create aligned synthetic data. After supervised training, Hibiki performs adaptive, simultaneous speech translation with vanilla temperature sampling. On a French-English simultaneous speech translation task, Hibiki demonstrates state-of-the-art performance in translation quality, speaker fidelity and naturalness. Moreover, the simplicity of its inference process makes it compatible with batched translation and even real-time on-device deployment. We provide examples as well as models and inference code.

Anne Wu, Laurent Mazaré, Neil Zeghidour et al.

We propose a novel preference alignment framework for improving spoken dialogue models on real-time conversations from user interactions. Current preference learning methods primarily focus on text-based language models, and are not directly suited to the complexities of real-time speech interactions, with richer dynamics (e.g. interruption, interjection) and no explicit segmentation between speaker turns.We create a large-scale dataset of more than 150,000 preference pairs from raw multi-turn speech conversations, annotated with AI feedback, to cover preferences over both linguistic content and temporal context variations. We leverage offline alignment methods to finetune a full-duplex autoregressive speech-to-speech model. Extensive experiments demonstrate that feedback on generic conversations can be consistently effective in improving spoken dialogue models to produce more factual, safer and more contextually aligned interactions. We deploy the finetuned model and conduct holistic human evaluations to assess the impact beyond single-turn conversations. Our findings shed light on the importance of a well-calibrated balance among various dynamics, crucial for natural real-time speech dialogue systems.

Matthieu Futeral, Andrea Agostinelli, Marco Tagliasacchi et al.

Generative spoken language models produce speech in a wide range of voices, prosody, and recording conditions, seemingly approaching the diversity of natural speech. However, the extent to which generated speech is acoustically diverse remains unclear due to a lack of appropriate metrics. We address this gap by developing lightweight metrics of acoustic diversity, which we collectively refer to as MAD Speech. We focus on measuring five facets of acoustic diversity: voice, gender, emotion, accent, and background noise. We construct the metrics as a composition of specialized, per-facet embedding models and an aggregation function that measures diversity within the embedding space. Next, we build a series of datasets with a priori known diversity preferences for each facet. Using these datasets, we demonstrate that our proposed metrics achieve a stronger agreement with the ground-truth diversity than baselines. Finally, we showcase the applicability of our proposed metrics across several real-life evaluation scenarios. MAD Speech is made publicly accessible.

Tsai-Ning Wang, Lin-Lin Chen, Neil Zeghidour et al.

Medical audio signals, such as heart and lung sounds, play a crucial role in clinical diagnosis. However, analyzing these signals remains challenging: traditional methods rely on handcrafted features or supervised deep learning models that demand extensive labeled datasets, limiting their scalability and applicability. To address these issues, we propose CaReAQA, an audio-language model that integrates a foundation audio model with the reasoning capabilities of large language models, enabling clinically relevant, open-ended diagnostic responses. Alongside CaReAQA, we introduce CaReSound, a benchmark dataset of annotated medical audio recordings enriched with metadata and paired question-answer examples, intended to drive progress in diagnostic reasoning research. Evaluation results show that CaReAQA achieves 86.2% accuracy on open-ended diagnostic reasoning tasks, outperforming baseline models. It also generalizes well to closed-ended classification tasks, achieving an average accuracy of 56.9% on unseen datasets. Our findings show how audio-language integration and reasoning advances medical diagnostics, enabling efficient AI systems for clinical decision support.

Amélie Royer, Moritz Böhle, Gabriel de Marmiesse et al.

The recent successes of Vision-Language models raise the question of how to equivalently imbue a pretrained speech model with vision understanding, an important milestone towards building a multimodal speech model able to freely converse about images. Building such a conversational Vision-Speech model brings its unique challenges: (i) paired image-speech datasets are much scarcer than their image-text counterparts, (ii) ensuring real-time latency at inference is crucial thus bringing compute and memory constraints, and (iii) the model should preserve prosodic features (e.g., speaker tone) which cannot be inferred from text alone. In this work, we introduce MoshiVis, augmenting a recent dialogue speech LLM, Moshi, with visual inputs through lightweight adaptation modules. An additional dynamic gating mechanism enables the model to more easily switch between the visual inputs and unrelated conversation topics. To reduce training costs, we design a simple one-stage, parameter-efficient fine-tuning pipeline in which we leverage a mixture of image-text (i.e., "speechless") and image-speech samples. We evaluate the model on downstream visual understanding tasks with both audio and text prompts, and report qualitative samples of interactions with MoshiVis. Our inference code will be made available, as well as the image-speech data used for audio evaluation.

Zalán Borsos, Matt Sharifi, Damien Vincent et al.

We present SoundStorm, a model for efficient, non-autoregressive audio generation. SoundStorm receives as input the semantic tokens of AudioLM, and relies on bidirectional attention and confidence-based parallel decoding to generate the tokens of a neural audio codec. Compared to the autoregressive generation approach of AudioLM, our model produces audio of the same quality and with higher consistency in voice and acoustic conditions, while being two orders of magnitude faster. SoundStorm generates 30 seconds of audio in 0.5 seconds on a TPU-v4. We demonstrate the ability of our model to scale audio generation to longer sequences by synthesizing high-quality, natural dialogue segments, given a transcript annotated with speaker turns and a short prompt with the speakers' voices.

Curtis Hawthorne, Andrew Jaegle, Cătălina Cangea et al.

Real-world data is high-dimensional: a book, image, or musical performance can easily contain hundreds of thousands of elements even after compression. However, the most commonly used autoregressive models, Transformers, are prohibitively expensive to scale to the number of inputs and layers needed to capture this long-range structure. We develop Perceiver AR, an autoregressive, modality-agnostic architecture which uses cross-attention to map long-range inputs to a small number of latents while also maintaining end-to-end causal masking. Perceiver AR can directly attend to over a hundred thousand tokens, enabling practical long-context density estimation without the need for hand-crafted sparsity patterns or memory mechanisms. When trained on images or music, Perceiver AR generates outputs with clear long-term coherence and structure. Our architecture also obtains state-of-the-art likelihood on long-sequence benchmarks, including 64 x 64 ImageNet images and PG-19 books.

Rachid Riad, Olivier Teboul, David Grangier et al.

Convolutional neural networks typically contain several downsampling operators, such as strided convolutions or pooling layers, that progressively reduce the resolution of intermediate representations. This provides some shift-invariance while reducing the computational complexity of the whole architecture. A critical hyperparameter of such layers is their stride: the integer factor of downsampling. As strides are not differentiable, finding the best configuration either requires cross-validation or discrete optimization (e.g. architecture search), which rapidly become prohibitive as the search space grows exponentially with the number of downsampling layers. Hence, exploring this search space by gradient descent would allow finding better configurations at a lower computational cost. This work introduces DiffStride, the first downsampling layer with learnable strides. Our layer learns the size of a cropping mask in the Fourier domain, that effectively performs resizing in a differentiable way. Experiments on audio and image classification show the generality and effectiveness of our solution: we use DiffStride as a drop-in replacement to standard downsampling layers and outperform them. In particular, we show that introducing our layer into a ResNet-18 architecture allows keeping consistent high performance on CIFAR10, CIFAR100 and ImageNet even when training starts from poor random stride configurations. Moreover, formulating strides as learnable variables allows us to introduce a regularization term that controls the computational complexity of the architecture. We show how this regularization allows trading off accuracy for efficiency on ImageNet.

Neil Zeghidour, Alejandro Luebs, Ahmed Omran et al.

We present SoundStream, a novel neural audio codec that can efficiently compress speech, music and general audio at bitrates normally targeted by speech-tailored codecs. SoundStream relies on a model architecture composed by a fully convolutional encoder/decoder network and a residual vector quantizer, which are trained jointly end-to-end. Training leverages recent advances in text-to-speech and speech enhancement, which combine adversarial and reconstruction losses to allow the generation of high-quality audio content from quantized embeddings. By training with structured dropout applied to quantizer layers, a single model can operate across variable bitrates from 3kbps to 18kbps, with a negligible quality loss when compared with models trained at fixed bitrates. In addition, the model is amenable to a low latency implementation, which supports streamable inference and runs in real time on a smartphone CPU. In subjective evaluations using audio at 24kHz sampling rate, SoundStream at 3kbps outperforms Opus at 12kbps and approaches EVS at 9.6kbps. Moreover, we are able to perform joint compression and enhancement either at the encoder or at the decoder side with no additional latency, which we demonstrate through background noise suppression for speech.

Neil Zeghidour, Olivier Teboul, David Grangier

We introduce DIVE, an end-to-end speaker diarization algorithm. Our neural algorithm presents the diarization task as an iterative process: it repeatedly builds a representation for each speaker before predicting the voice activity of each speaker conditioned on the extracted representations. This strategy intrinsically resolves the speaker ordering ambiguity without requiring the classical permutation invariant training loss. In contrast with prior work, our model does not rely on pretrained speaker representations and optimizes all parameters of the system with a multi-speaker voice activity loss. Importantly, our loss explicitly excludes unreliable speaker turn boundaries from training, which is adapted to the standard collar-based Diarization Error Rate (DER) evaluation. Overall, these contributions yield a system redefining the state-of-the-art on the standard CALLHOME benchmark, with 6.7% DER compared to 7.8% for the best alternative.

Andrew N Carr, Quentin Berthet, Mathieu Blondel et al.

Self-supervised pre-training using so-called "pretext" tasks has recently shown impressive performance across a wide range of modalities. In this work, we advance self-supervised learning from permutations, by pre-training a model to reorder shuffled parts of the spectrogram of an audio signal, to improve downstream classification performance. We make two main contributions. First, we overcome the main challenges of integrating permutation inversions into an end-to-end training scheme, using recent advances in differentiable ranking. This was heretofore sidestepped by casting the reordering task as classification, fundamentally reducing the space of permutations that can be exploited. Our experiments validate that learning from all possible permutations improves the quality of the pre-trained representations over using a limited, fixed set. Second, we show that inverting permutations is a meaningful pretext task for learning audio representations in an unsupervised fashion. In particular, we improve instrument classification and pitch estimation of musical notes by reordering spectrogram patches in the time-frequency space.

Neil Zeghidour, Olivier Teboul, Félix de Chaumont Quitry et al.

Mel-filterbanks are fixed, engineered audio features which emulate human perception and have been used through the history of audio understanding up to today. However, their undeniable qualities are counterbalanced by the fundamental limitations of handmade representations. In this work we show that we can train a single learnable frontend that outperforms mel-filterbanks on a wide range of audio signals, including speech, music, audio events and animal sounds, providing a general-purpose learned frontend for audio classification. To do so, we introduce a new principled, lightweight, fully learnable architecture that can be used as a drop-in replacement of mel-filterbanks. Our system learns all operations of audio features extraction, from filtering to pooling, compression and normalization, and can be integrated into any neural network at a negligible parameter cost. We perform multi-task training on eight diverse audio classification tasks, and show consistent improvements of our model over mel-filterbanks and previous learnable alternatives. Moreover, our system outperforms the current state-of-the-art learnable frontend on Audioset, with orders of magnitude fewer parameters.

Aaqib Saeed, David Grangier, Olivier Pietquin et al.

We propose CHARM, a method for training a single neural network across inconsistent input channels. Our work is motivated by Electroencephalography (EEG), where data collection protocols from different headsets result in varying channel ordering and number, which limits the feasibility of transferring trained systems across datasets. Our approach builds upon attention mechanisms to estimate a latent reordering matrix from each input signal and map input channels to a canonical order. CHARM is differentiable and can be composed further with architectures expecting a consistent channel ordering to build end-to-end trainable classifiers. We perform experiments on four EEG classification datasets and demonstrate the efficacy of CHARM via simulated shuffling and masking of input channels. Moreover, our method improves the transfer of pre-trained representations between datasets collected with different protocols.

Aaqib Saeed, David Grangier, Neil Zeghidour

We introduce COLA, a self-supervised pre-training approach for learning a general-purpose representation of audio. Our approach is based on contrastive learning: it learns a representation which assigns high similarity to audio segments extracted from the same recording while assigning lower similarity to segments from different recordings. We build on top of recent advances in contrastive learning for computer vision and reinforcement learning to design a lightweight, easy-to-implement self-supervised model of audio. We pre-train embeddings on the large-scale Audioset database and transfer these representations to 9 diverse classification tasks, including speech, music, animal sounds, and acoustic scenes. We show that despite its simplicity, our method significantly outperforms previous self-supervised systems. We furthermore conduct ablation studies to identify key design choices and release a library to pre-train and fine-tune COLA models.

Neil Zeghidour, David Grangier

We introduce Wavesplit, an end-to-end source separation system. From a single mixture, the model infers a representation for each source and then estimates each source signal given the inferred representations. The model is trained to jointly perform both tasks from the raw waveform. Wavesplit infers a set of source representations via clustering, which addresses the fundamental permutation problem of separation. For speech separation, our sequence-wide speaker representations provide a more robust separation of long, challenging recordings compared to prior work. Wavesplit redefines the state-of-the-art on clean mixtures of 2 or 3 speakers (WSJ0-2/3mix), as well as in noisy and reverberated settings (WHAM/WHAMR). We also set a new benchmark on the recent LibriMix dataset. Finally, we show that Wavesplit is also applicable to other domains, by separating fetal and maternal heart rates from a single abdominal electrocardiogram.

Gabriel Dulac-Arnold, Neil Zeghidour, Marco Cuturi et al.

We propose a learning algorithm capable of learning from label proportions instead of direct data labels. In this scenario, our data are arranged into various bags of a certain size, and only the proportions of each label within a given bag are known. This is a common situation in cases where per-data labeling is lengthy, but a more general label is easily accessible. Several approaches have been proposed to learn in this setting with linear models in the multiclass setting, or with nonlinear models in the binary classification setting. Here we investigate the more general nonlinear multiclass setting, and compare two differentiable loss functions to train end-to-end deep neural networks from bags with label proportions. We illustrate the relevance of our methods on an image classification benchmark, and demonstrate the possibility to learn accurate image classifiers from bags of images.

Neil Zeghidour, Qiantong Xu, Vitaliy Liptchinsky et al.

Current state-of-the-art speech recognition systems build on recurrent neural networks for acoustic and/or language modeling, and rely on feature extraction pipelines to extract mel-filterbanks or cepstral coefficients. In this paper we present an alternative approach based solely on convolutional neural networks, leveraging recent advances in acoustic models from the raw waveform and language modeling. This fully convolutional approach is trained end-to-end to predict characters from the raw waveform, removing the feature extraction step altogether. An external convolutional language model is used to decode words. On Wall Street Journal, our model matches the current state-of-the-art. On Librispeech, we report state-of-the-art performance among end-to-end models, including Deep Speech 2 trained with 12 times more acoustic data and significantly more linguistic data.

Yossi Adi, Neil Zeghidour, Ronan Collobert et al.

Transcribed datasets typically contain speaker identity for each instance in the data. We investigate two ways to incorporate this information during training: Multi-Task Learning and Adversarial Learning. In multi-task learning, the goal is speaker prediction; we expect a performance improvement with this joint training if the two tasks of speech recognition and speaker recognition share a common set of underlying features. In contrast, adversarial learning is a means to learn representations invariant to the speaker. We then expect better performance if this learnt invariance helps generalizing to new speakers. While the two approaches seem natural in the context of speech recognition, they are incompatible because they correspond to opposite gradients back-propagated to the model. In order to better understand the effect of these approaches in terms of error rates, we compare both strategies in controlled settings. Moreover, we explore the use of additional untranscribed data in a semi-supervised, adversarial learning manner to improve error rates. Our results show that deep models trained on big datasets already develop invariant representations to speakers without any auxiliary loss. When considering adversarial learning and multi-task learning, the impact on the acoustic model seems minor. However, models trained in a semi-supervised manner can improve error-rates.

Juliette Millet, Neil Zeghidour

Speech classifiers of paralinguistic traits traditionally learn from diverse hand-crafted low-level features, by selecting the relevant information for the task at hand. We explore an alternative to this selection, by learning jointly the classifier, and the feature extraction. Recent work on speech recognition has shown improved performance over speech features by learning from the waveform. We extend this approach to paralinguistic classification and propose a neural network that can learn a filterbank, a normalization factor and a compression power from the raw speech, jointly with the rest of the architecture. We apply this model to dysarthria detection from sentence-level audio recordings. Starting from a strong attention-based baseline on which mel-filterbanks outperform standard low-level descriptors, we show that learning the filters or the normalization and compression improves over fixed features by 10% absolute accuracy. We also observe a gain over OpenSmile features by learning jointly the feature extraction, the normalization, and the compression factor with the architecture. This constitutes a first attempt at learning jointly all these operations from raw audio for a speech classification task.

Alexandre Défossez, Neil Zeghidour, Nicolas Usunier et al.

Recent progress in deep learning for audio synthesis opens the way to models that directly produce the waveform, shifting away from the traditional paradigm of relying on vocoders or MIDI synthesizers for speech or music generation. Despite their successes, current state-of-the-art neural audio synthesizers such as WaveNet and SampleRNN suffer from prohibitive training and inference times because they are based on autoregressive models that generate audio samples one at a time at a rate of 16kHz. In this work, we study the more computationally efficient alternative of generating the waveform frame-by-frame with large strides. We present SING, a lightweight neural audio synthesizer for the original task of generating musical notes given desired instrument, pitch and velocity. Our model is trained end-to-end to generate notes from nearly 1000 instruments with a single decoder, thanks to a new loss function that minimizes the distances between the log spectrograms of the generated and target waveforms. On the generalization task of synthesizing notes for pairs of pitch and instrument not seen during training, SING produces audio with significantly improved perceptual quality compared to a state-of-the-art autoencoder based on WaveNet as measured by a Mean Opinion Score (MOS), and is about 32 times faster for training and 2, 500 times faster for inference.

Neil Zeghidour, Nicolas Usunier, Gabriel Synnaeve et al.

State-of-the-art speech recognition systems rely on fixed, hand-crafted features such as mel-filterbanks to preprocess the waveform before the training pipeline. In this paper, we study end-to-end systems trained directly from the raw waveform, building on two alternatives for trainable replacements of mel-filterbanks that use a convolutional architecture. The first one is inspired by gammatone filterbanks (Hoshen et al., 2015; Sainath et al, 2015), and the second one by the scattering transform (Zeghidour et al., 2017). We propose two modifications to these architectures and systematically compare them to mel-filterbanks, on the Wall Street Journal dataset. The first modification is the addition of an instance normalization layer, which greatly improves on the gammatone-based trainable filterbanks and speeds up the training of the scattering-based filterbanks. The second one relates to the low-pass filter used in these approaches. These modifications consistently improve performances for both approaches, and remove the need for a careful initialization in scattering-based trainable filterbanks. In particular, we show a consistent improvement in word error rate of the trainable filterbanks relatively to comparable mel-filterbanks. It is the first time end-to-end models trained from the raw signal significantly outperform mel-filterbanks on a large vocabulary task under clean recording conditions.

Rachid Riad, Corentin Dancette, Julien Karadayi et al.

Recent studies have investigated siamese network architectures for learning invariant speech representations using same-different side information at the word level. Here we investigate systematically an often ignored component of siamese networks: the sampling procedure (how pairs of same vs. different tokens are selected). We show that sampling strategies taking into account Zipf's Law, the distribution of speakers and the proportions of same and different pairs of words significantly impact the performance of the network. In particular, we show that word frequency compression improves learning across a large range of variations in number of training pairs. This effect does not apply to the same extent to the fully unsupervised setting, where the pairs of same-different words are obtained by spoken term discovery. We apply these results to pairs of words discovered using an unsupervised algorithm and show an improvement on state-of-the-art in unsupervised representation learning using siamese networks.

Neil Zeghidour, Nicolas Usunier, Iasonas Kokkinos et al.

We train a bank of complex filters that operates on the raw waveform and is fed into a convolutional neural network for end-to-end phone recognition. These time-domain filterbanks (TD-filterbanks) are initialized as an approximation of mel-filterbanks, and then fine-tuned jointly with the remaining convolutional architecture. We perform phone recognition experiments on TIMIT and show that for several architectures, models trained on TD-filterbanks consistently outperform their counterparts trained on comparable mel-filterbanks. We get our best performance by learning all front-end steps, from pre-emphasis up to averaging. Finally, we observe that the filters at convergence have an asymmetric impulse response, and that some of them remain almost analytic.

Guillaume Lample, Neil Zeghidour, Nicolas Usunier et al.

This paper introduces a new encoder-decoder architecture that is trained to reconstruct images by disentangling the salient information of the image and the values of attributes directly in the latent space. As a result, after training, our model can generate different realistic versions of an input image by varying the attribute values. By using continuous attribute values, we can choose how much a specific attribute is perceivable in the generated image. This property could allow for applications where users can modify an image using sliding knobs, like faders on a mixing console, to change the facial expression of a portrait, or to update the color of some objects. Compared to the state-of-the-art which mostly relies on training adversarial networks in pixel space by altering attribute values at train time, our approach results in much simpler training schemes and nicely scales to multiple attributes. We present evidence that our model can significantly change the perceived value of the attributes while preserving the naturalness of images.

Rahma Chaabouni, Ewan Dunbar, Neil Zeghidour et al.

Recent works have explored deep architectures for learning multimodal speech representation (e.g. audio and images, articulation and audio) in a supervised way. Here we investigate the role of combining different speech modalities, i.e. audio and visual information representing the lips movements, in a weakly supervised way using Siamese networks and lexical same-different side information. In particular, we ask whether one modality can benefit from the other to provide a richer representation for phone recognition in a weakly supervised setting. We introduce mono-task and multi-task methods for merging speech and visual modalities for phone recognition. The mono-task learning consists in applying a Siamese network on the concatenation of the two modalities, while the multi-task learning receives several different combinations of modalities at train time. We show that multi-task learning enhances discriminability for visual and multimodal inputs while minimally impacting auditory inputs. Furthermore, we present a qualitative analysis of the obtained phone embeddings, and show that cross-modal visual input can improve the discriminability of phonological features which are visually discernable (rounding, open/close, labial place of articulation), resulting in representations that are closer to abstract linguistic features than those based on audio only.