Text-Speech Language Models (TSLMs) -- language models trained to jointly process and generate text and speech -- aim to enable cross-modal knowledge transfer to overcome the scaling limitations of unimodal speech LMs. The predominant approach to TSLM training expands the vocabulary of a pre-trained text LM by appending new embeddings and linear projections for speech, followed by fine-tuning on speech data. We hypothesize that this method limits cross-modal transfer by neglecting feature compositionality, preventing text-learned functions from being fully leveraged at appropriate abstraction levels. To address this, we propose augmenting vocabulary expansion with modules that better align abstraction levels across layers. Our models, \textsc{SmolTolk}, rival or surpass state-of-the-art TSLMs trained with orders of magnitude more compute. Representation analyses and improved multimodal performance suggest our method enhances cross-modal transfer.

Speech Language Models (SLMs) aim to learn language from raw audio, without textual resources. Despite significant advances, our current models exhibit weak syntax and semantic abilities. However, if the scaling properties of neural language models hold for the speech modality, these abilities will improve as the amount of compute used for training increases. In this paper, we use models of this scaling behavior to estimate the scale at which our current methods will yield a SLM with the English proficiency of text-based Large Language Models (LLMs). We establish a strong correlation between pre-training loss and downstream syntactic and semantic performance in SLMs and LLMs, which results in predictable scaling of linguistic performance. We show that the linguistic performance of SLMs scales up to three orders of magnitude more slowly than that of text-based LLMs. Additionally, we study the benefits of synthetic data designed to boost semantic understanding and the effects of coarser speech tokenization.

Macroscopic intelligibility models predict the expected human word-error-rate for a given speech-in-noise stimulus. In contrast, microscopic intelligibility models aim to make fine-grained predictions about listeners' perception, e.g. predicting phonetic or lexical responses. State-of-the-art macroscopic models use transfer learning from large scale deep learning models for speech processing, whereas such methods have rarely been used for microscopic modeling. In this paper, we study the use of transfer learning from Whisper, a state-of-the-art deep learning model for automatic speech recognition, for microscopic intelligibility prediction at the level of lexical responses. Our method outperforms the considered baselines, even in a zero-shot setup, and yields a relative improvement of up to 66\% when fine-tuned to predict listeners' responses. Our results showcase the promise of large scale deep learning based methods for microscopic intelligibility prediction.

Speech foundation models (SFMs) have been benchmarked on many speech processing tasks, often achieving state-of-the-art performance with minimal adaptation. However, the SFM paradigm has been significantly less explored for applications of interest to the speech perception community. In this paper we present a systematic evaluation of 10 SFMs on one such application: Speech intelligibility prediction. We focus on the non-intrusive setup of the Clarity Prediction Challenge 2 (CPC2), where the task is to predict the percentage of words correctly perceived by hearing-impaired listeners from speech-in-noise recordings. We propose a simple method that learns a lightweight specialized prediction head on top of frozen SFMs to approach the problem. Our results reveal statistically significant differences in performance across SFMs. Our method resulted in the winning submission in the CPC2, demonstrating its promise for speech perception applications.

Visual localization plays an important role in the positioning and navigation of robotics systems within previously visited environments. When visits occur over long periods of time, changes in the environment related to seasons or day-night cycles present a major challenge. Under water, the sources of variability are due to other factors such as water conditions or growth of marine organisms. Yet it remains a major obstacle and a much less studied one, partly due to the lack of data. This paper presents a new deep-sea dataset to benchmark underwater long-term visual localization. The dataset is composed of images from four visits to the same hydrothermal vent edifice over the course of five years. Camera poses and a common geometry of the scene were estimated using navigation data and Structure-from-Motion. This serves as a reference when evaluating visual localization techniques. An analysis of the data provides insights about the major changes observed throughout the years. Furthermore, several well-established visual localization methods are evaluated on the dataset, showing there is still room for improvement in underwater long-term visual localization. The data is made publicly available at https://www.seanoe.org/data/00810/92226/.

The success of deep learning comes from its ability to capture the hierarchical structure of data by learning high-level representations defined in terms of low-level ones. In this paper we explore self-supervised learning of hierarchical representations of speech by applying multiple levels of Contrastive Predictive Coding (CPC). We observe that simply stacking two CPC models does not yield significant improvements over single-level architectures. Inspired by the fact that speech is often described as a sequence of discrete units unevenly distributed in time, we propose a model in which the output of a low-level CPC module is non-uniformly downsampled to directly minimize the loss of a high-level CPC module. The latter is designed to also enforce a prior of separability and discreteness in its representations by enforcing dissimilarity of successive high-level representations through focused negative sampling, and by quantization of the prediction targets. Accounting for the structure of the speech signal improves upon single-level CPC features and enhances the disentanglement of the learned representations, as measured by downstream speech recognition tasks, while resulting in a meaningful segmentation of the signal that closely resembles phone boundaries.

In this paper we demonstrate methods for reliable and efficient training of discrete representation using Vector-Quantized Variational Auto-Encoder models (VQ-VAEs). Discrete latent variable models have been shown to learn nontrivial representations of speech, applicable to unsupervised voice conversion and reaching state-of-the-art performance on unit discovery tasks. For unsupervised representation learning, they became viable alternatives to continuous latent variable models such as the Variational Auto-Encoder (VAE). However, training deep discrete variable models is challenging, due to the inherent non-differentiability of the discretization operation. In this paper we focus on VQ-VAE, a state-of-the-art discrete bottleneck model shown to perform on par with its continuous counterparts. It quantizes encoder outputs with on-line $k$-means clustering. We show that the codebook learning can suffer from poor initialization and non-stationarity of clustered encoder outputs. We demonstrate that these can be successfully overcome by increasing the learning rate for the codebook and periodic date-dependent codeword re-initialization. As a result, we achieve more robust training across different tasks, and significantly increase the usage of latent codewords even for large codebooks. This has practical benefit, for instance, in unsupervised representation learning, where large codebooks may lead to disentanglement of latent representations.

A system is presented that segments, clusters and predicts musical audio in an unsupervised manner, adjusting the number of (timbre) clusters instantaneously to the audio input. A sequence learning algorithm adapts its structure to a dynamically changing clustering tree. The flow of the system is as follows: 1) segmentation by onset detection, 2) timbre representation of each segment by Mel frequency cepstrum coefficients, 3) discretization by incremental clustering, yielding a tree of different sound classes (e.g. instruments) that can grow or shrink on the fly driven by the instantaneous sound events, resulting in a discrete symbol sequence, 4) extraction of statistical regularities of the symbol sequence, using hierarchical N-grams and the newly introduced conceptual Boltzmann machine, and 5) prediction of the next sound event in the sequence. The system's robustness is assessed with respect to complexity and noisiness of the signal. Clustering in isolation yields an adjusted Rand index (ARI) of 82.7% / 85.7% for data sets of singing voice and drums. Onset detection jointly with clustering achieve an ARI of 81.3% / 76.3% and the prediction of the entire system yields an ARI of 27.2% / 39.2%.