We introduce EmphAssess, a prosodic benchmark designed to evaluate the capability of speech-to-speech models to encode and reproduce prosodic emphasis. We apply this to two tasks: speech resynthesis and speech-to-speech translation. In both cases, the benchmark evaluates the ability of the model to encode emphasis in the speech input and accurately reproduce it in the output, potentially across a change of speaker and language. As part of the evaluation pipeline, we introduce EmphaClass, a new model that classifies emphasis at the frame or word level.
The transformer architecture from Vaswani et al. (2017) is now ubiquitous across application domains, from natural language processing to speech processing and image understanding. We propose DenseFormer, a simple modification to the standard architecture that improves the perplexity of the model without increasing its size -- adding a few thousand parameters for large-scale models in the 100B parameters range. Our approach relies on an additional averaging step after each transformer block, which computes a weighted average of current and past representations -- we refer to this operation as Depth-Weighted-Average (DWA). The learned DWA weights exhibit coherent patterns of information flow, revealing the strong and structured reuse of activations from distant layers. Experiments demonstrate that DenseFormer is more data efficient, reaching the same perplexity of much deeper transformer models, and that for the same perplexity, these new models outperform transformer baselines in terms of memory efficiency and inference time.
Multichannel convolutive blind speech source separation refers to the problem of separating different speech sources from the observed multichannel mixtures without much a priori information about the mixing system. Multichannel nonnegative matrix factorization (MNMF) has been proven to be one of the most powerful separation frameworks and the representative algorithms such as MNMF and the independent low-rank matrix analysis (ILRMA) have demonstrated great performance. However, the sparseness properties of speech source signals are not fully taken into account in such a framework. It is well known that speech signals are sparse in nature, which is considered in this work to improve the separation performance. Specifically, we utilize the Bingham and Laplace distributions to formulate a disjoint constraint regularizer, which is subsequently incorporated into both MNMF and ILRMA. We then derive majorization-minimization rules for updating parameters related to the source model, resulting in the development of two enhanced algorithms: s-MNMF and s-ILRMA. Comprehensive simulations are conducted, and the results unequivocally demonstrate the efficacy of our proposed methodologies.
In multilingual speech recognition systems, a situation can often arise when the language is not known in advance, but the signal has already been received and is being processed. For such cases, some generalized model is needed that will be able to respond to phonetic differences and, depending on them, correctly recog-nize speech in the desired language. To build such a model, it is necessary to set the values of phonetic parameters, and then compare similar sounds, establishing significant differences.
Current talking avatars mostly generate co-speech gestures based on audio and text of the utterance, without considering the non-speaking motion of the speaker. Furthermore, previous works on co-speech gesture generation have designed network structures based on individual gesture datasets, which results in limited data volume, compromised generalizability, and restricted speaker movements. To tackle these issues, we introduce FreeTalker, which, to the best of our knowledge, is the first framework for the generation of both spontaneous (e.g., co-speech gesture) and non-spontaneous (e.g., moving around the podium) speaker motions. Specifically, we train a diffusion-based model for speaker motion generation that employs unified representations of both speech-driven gestures and text-driven motions, utilizing heterogeneous data sourced from various motion datasets. During inference, we utilize classifier-free guidance to highly control the style in the clips. Additionally, to create smooth transitions between clips, we utilize DoubleTake, a method that leverages a generative prior and ensures seamless motion blending. Extensive experiments show that our method generates natural and controllable speaker movements. Our code, model, and demo are are available at \url{https://youngseng.github.io/FreeTalker/}.
Non-active adaptive sampling is a way of building machine learning models from a training data base which are supposed to dynamically and automatically derive guaranteed sample size. In this context and regardless of the strategy used in both scheduling and generating of weak predictors, a proposal for calculating absolute convergence and error thresholds is described. We not only make it possible to establish when the quality of the model no longer increases, but also supplies a proximity condition to estimate in absolute terms how close it is to achieving such a goal, thus supporting decision making for fine-tuning learning parameters in model selection. The technique proves its correctness and completeness with respect to our working hypotheses, in addition to strengthening the robustness of the sampling scheme. Tests meet our expectations and illustrate the proposal in the domain of natural language processing, taking the generation of part-of-speech taggers as case study.
This paper presents an innovative framework that integrates Large Language Models (LLMs) with an external Thinker module to enhance the reasoning capabilities of LLM-based agents. Unlike augmenting LLMs with prompt engineering, Thinker directly harnesses knowledge from databases and employs various optimization techniques. The framework forms a reasoning hierarchy where LLMs handle intuitive System-1 tasks such as natural language processing, while the Thinker focuses on cognitive System-2 tasks that require complex logical analysis and domain-specific knowledge. Our framework is presented using a 9-player Werewolf game that demands dual-system reasoning. We introduce a communication protocol between LLMs and the Thinker, and train the Thinker using data from 18800 human sessions and reinforcement learning. Experiments demonstrate the framework's effectiveness in deductive reasoning, speech generation, and online game evaluation. Additionally, we fine-tune a 6B LLM to surpass GPT4 when integrated with the Thinker. This paper also contributes the largest dataset for social deduction games to date.
We present a novel model designed for resource-efficient multichannel speech enhancement in the time domain, with a focus on low latency, lightweight, and low computational requirements. The proposed model incorporates explicit spatial and temporal processing within deep neural network (DNN) layers. Inspired by frequency-dependent multichannel filtering, our spatial filtering process applies multiple trainable filters to each hidden unit across the spatial dimension, resulting in a multichannel output. The temporal processing is applied over a single-channel output stream from the spatial processing using a Long Short-Term Memory (LSTM) network. The output from the temporal processing stage is then further integrated into the spatial dimension through elementwise multiplication. This explicit separation of spatial and temporal processing results in a resource-efficient network design. Empirical findings from our experiments show that our proposed model significantly outperforms robust baseline models while demanding far fewer parameters and computations, while achieving an ultra-low algorithmic latency of just 2 milliseconds.
In this paper, we present a novel bilevel optimization-based training approach to training acoustic models for automatic speech recognition (ASR) tasks that we term {bi-level joint unsupervised and supervised training (BL-JUST)}. {BL-JUST employs a lower and upper level optimization with an unsupervised loss and a supervised loss respectively, leveraging recent advances in penalty-based bilevel optimization to solve this challenging ASR problem with affordable complexity and rigorous convergence guarantees.} To evaluate BL-JUST, extensive experiments on the LibriSpeech and TED-LIUM v2 datasets have been conducted. BL-JUST achieves superior performance over the commonly used pre-training followed by fine-tuning strategy.
In this work, we explore the task of hierarchical distance-based speech separation defined on a hyperbolic manifold. Based on the recent advent of audio-related tasks performed in non-Euclidean spaces, we propose to make use of the Poincar\'e ball to effectively unveil the inherent hierarchical structure found in complex speaker mixtures. We design two sets of experiments in which the distance-based parent sound classes, namely "near" and "far", can contain up to two or three speakers (i.e., children) each. We show that our hyperbolic approach is suitable for unveiling hierarchical structure from the problem definition, resulting in improved child-level separation. We further show that a clear correlation emerges between the notion of hyperbolic certainty (i.e., the distance to the ball's origin) and acoustic semantics such as speaker density, inter-source location, and microphone-to-speaker distance.