分享

As deep neural networks continue to revolutionize various application domains, there is increasing interest in making these powerful models more understandable and interpretable, and narrowing down the causes of good and bad predictions. We focus on recurrent neural networks (RNNs), state of the art models in speech recognition and translation. Our approach to increasing interpretability is by combining an RNN with a hidden Markov model (HMM), a simpler and more transparent model. We explore various combinations of RNNs and HMMs: an HMM trained on LSTM states; a hybrid model where an HMM is trained first, then a small LSTM is given HMM state distributions and trained to fill in gaps in the HMM's performance; and a jointly trained hybrid model. We find that the LSTM and HMM learn complementary information about the features in the text.

Deep Learning-based end-to-end Automatic Speech Recognition (ASR) has made significant strides but still struggles with performance on out-of-domain samples due to domain shifts in real-world scenarios. Test-Time Adaptation (TTA) methods address this issue by adapting models using test samples at inference time. However, current ASR TTA methods have largely focused on non-continual TTA, which limits cross-sample knowledge learning compared to continual TTA. In this work, we first propose a Fast-slow TTA framework for ASR that leverages the advantage of continual and non-continual TTA. Following this framework, we introduce Dynamic SUTA (DSUTA), an entropy-minimization-based continual TTA method for ASR. To enhance DSUTA robustness for time-varying data, we design a dynamic reset strategy to automatically detect domain shifts and reset the model, making it more effective at handling multi-domain data. Our method demonstrates superior performance on various noisy ASR datasets, outperforming both non-continual and continual TTA baselines while maintaining robustness to domain changes without requiring domain boundary information.

Synthesizing speech across different accents while preserving the speaker identity is essential for various real-world customer applications. However, the individual and accurate modeling of accents and speakers in a text-to-speech (TTS) system is challenging due to the complexity of accent variations and the intrinsic entanglement between the accent and speaker identity. In this paper, we present a novel approach for multi-speaker multi-accent TTS synthesis, which aims to synthesize voices of multiple speakers, each with various accents. Our proposed approach employs a multi-scale accent modeling strategy to address accent variations at different levels. Specifically, we introduce both global (utterance level) and local (phoneme level) accent modeling, supervised by individual accent classifiers to capture the overall variation within accented utterances and fine-grained variations between phonemes, respectively. To control accents and speakers separately, speaker-independent accent modeling is necessary, which is achieved by adversarial training with speaker classifiers to disentangle speaker identity within the multi-scale accent modeling. Consequently, we obtain speaker-independent and accent-discriminative multi-scale embeddings as comprehensive accent features. Additionally, we propose a local accent prediction model that allows to generate accented speech directly from phoneme inputs. Extensive experiments are conducted on an accented English speech corpus. Both objective and subjective evaluations show the superiority of our proposed system compared to baselines systems. Detailed component analysis demonstrates the effectiveness of global and local accent modeling, and speaker disentanglement on multi-speaker multi-accent speech synthesis.

Keyword Spotting (KWS) is essential in edge computing requiring rapid and energy-efficient responses. Spiking Neural Networks (SNNs) are well-suited for KWS for their efficiency and temporal capacity for speech. To further reduce the latency and energy consumption, this study introduces ED-sKWS, an SNN-based KWS model with an early-decision mechanism that can stop speech processing and output the result before the end of speech utterance. Furthermore, we introduce a Cumulative Temporal (CT) loss that can enhance prediction accuracy at both the intermediate and final timesteps. To evaluate early-decision performance, we present the SC-100 dataset including 100 speech commands with beginning and end timestamp annotation. Experiments on the Google Speech Commands v2 and our SC-100 datasets show that ED-sKWS maintains competitive accuracy with 61% timesteps and 52% energy consumption compared to SNN models without early-decision mechanism, ensuring rapid response and energy efficiency.

This paper introduces the idea of applying signal processing inside a Large Language Model (LLM). With the recent explosion of generative AI, our work can help bridge two fields together, namely the field of signal processing and large language models. We draw parallels between classical Fourier-Transforms and Fourier Transform-like learnable time-frequency representations for every intermediate activation signal of an LLM. Once we decompose every activation signal across tokens into a time-frequency representation, we learn how to filter and reconstruct them, with all components learned from scratch, to predict the next token given the previous context. We show that for GPT-like architectures, our work achieves faster convergence and significantly increases performance by adding a minuscule number of extra parameters when trained for the same epochs. We hope this work paves the way for algorithms exploring signal processing inside the signals found in neural architectures like LLMs and beyond.

Accurate load forecasting is crucial for maintaining the power balance between generators and consumers,particularly with the increasing integration of renewable energy sources, which introduce significant intermittent volatility. With the advancement of data-driven methods, machine learning and deep learning models have become the predominant approaches for load forecasting tasks. In recent years, pre-trained large language models (LLMs) have achieved significant progress, demonstrating superior performance across various fields. This paper proposes a load forecasting method based on LLMs, offering not only precise predictive capabilities but also broad and flexible applicability. Additionally, a data modeling method is introduced to effectively transform load sequence data into natural language suitable for LLM training. Furthermore, a data enhancement strategy is designed to mitigate the impact of LLM hallucinations on forecasting results. The effectiveness of the proposed method is validated using two real-world datasets. Compared to existing methods, our approach demonstrates state-of-the-art performance across all validation metrics.

Speaker embedding extractors are typically trained using a classification loss over the training speakers. During the last few years, the standard softmax/cross-entropy loss has been replaced by the margin-based losses, yielding significant improvements in speaker recognition accuracy. Motivated by the fact that the margin merely reduces the logit of the target speaker during training, we consider a probabilistic framework that has a similar effect. The variational information bottleneck provides a principled mechanism for making deterministic nodes stochastic, resulting in an implicit reduction of the posterior of the target speaker. We experiment with a wide range of speaker recognition benchmarks and scoring methods and report competitive results to those obtained with the state-of-the-art Additive Angular Margin loss.

While automatic speech recognition (ASR) greatly benefits from data augmentation, the augmentation recipes themselves tend to be heuristic. In this paper, we address one of the heuristic approach associated with balancing the right amount of augmented data in ASR training by introducing a reinforcement learning (RL) based dynamic adjustment of original-to-augmented data ratio (OAR). Unlike the fixed OAR approach in conventional data augmentation, our proposed method employs a deep Q-network (DQN) as the RL mechanism to learn the optimal dynamics of OAR throughout the wav2vec2.0 based ASR training. We conduct experiments using the LibriSpeech dataset with varying amounts of training data, specifically, the 10Min, 1H, 10H, and 100H splits to evaluate the efficacy of the proposed method under different data conditions. Our proposed method, on average, achieves a relative improvement of 4.96% over the open-source wav2vec2.0 base model on standard LibriSpeech test sets.

Advances in artificial intelligence (AI) raise important questions about whether people view moral evaluations by AI systems similarly to human-generated moral evaluations. We conducted a modified Moral Turing Test (m-MTT), inspired by Allen and colleagues' (2000) proposal, by asking people to distinguish real human moral evaluations from those made by a popular advanced AI language model: GPT-4. A representative sample of 299 U.S. adults first rated the quality of moral evaluations when blinded to their source. Remarkably, they rated the AI's moral reasoning as superior in quality to humans' along almost all dimensions, including virtuousness, intelligence, and trustworthiness, consistent with passing what Allen and colleagues call the comparative MTT. Next, when tasked with identifying the source of each evaluation (human or computer), people performed significantly above chance levels. Although the AI did not pass this test, this was not because of its inferior moral reasoning but, potentially, its perceived superiority, among other possible explanations. The emergence of language models capable of producing moral responses perceived as superior in quality to humans' raises concerns that people may uncritically accept potentially harmful moral guidance from AI. This possibility highlights the need for safeguards around generative language models in matters of morality.

To be included into chatbot systems, Large language models (LLMs) must be aligned with human conversational conventions. However, being trained mainly on web-scraped data gives existing LLMs a voice closer to informational text than actual human speech. In this paper, we examine the effect of decoding methods on the alignment between LLM-generated and human conversations, including Beam Search, Top K Sampling, and Nucleus Sampling. We present new measures of alignment in substance, style, and psychometric orientation, and experiment with two conversation datasets. Our results provide subtle insights: better alignment is attributed to fewer beams in Beam Search and lower values of P in Nucleus Sampling. We also find that task-oriented and open-ended datasets perform differently in terms of alignment, indicating the significance of taking into account the context of the interaction.

Human emotions are often not expressed directly, but regulated according to internal processes and social display rules. For affective computing systems, an understanding of how users regulate their emotions can be highly useful, for example to provide feedback in job interview training, or in psychotherapeutic scenarios. However, at present no method to automatically classify different emotion regulation strategies in a cross-user scenario exists. At the same time, recent studies showed that instruction-tuned Large Language Models (LLMs) can reach impressive performance across a variety of affect recognition tasks such as categorical emotion recognition or sentiment analysis. While these results are promising, it remains unclear to what extent the representational power of LLMs can be utilized in the more subtle task of classifying users' internal emotion regulation strategy. To close this gap, we make use of the recently introduced \textsc{Deep} corpus for modeling the social display of the emotion shame, where each point in time is annotated with one of seven different emotion regulation classes. We fine-tune Llama2-7B as well as the recently introduced Gemma model using Low-rank Optimization on prompts generated from different sources of information on the \textsc{Deep} corpus. These include verbal and nonverbal behavior, person factors, as well as the results of an in-depth interview after the interaction. Our results show, that a fine-tuned Llama2-7B LLM is able to classify the utilized emotion regulation strategy with high accuracy (0.84) without needing access to data from post-interaction interviews. This represents a significant improvement over previous approaches based on Bayesian Networks and highlights the importance of modeling verbal behavior in emotion regulation.

This study investigates how BERT processes and represents Argument Structure Constructions (ASCs), extending previous LSTM analyses. Using a dataset of 2000 sentences across four ASC types (transitive, ditransitive, caused-motion, resultative), we analyzed BERT's token embeddings across 12 layers. Visualizations with MDS and t-SNE and clustering quantified by Generalized Discrimination Value (GDV) were used. Feedforward classifiers (probes) predicted construction categories from embeddings. CLS token embeddings clustered best in layers 2-4, decreased in intermediate layers, and slightly increased in final layers. DET and SUBJ embeddings showed consistent clustering in intermediate layers, VERB embeddings increased in clustering from layer 1 to 12, and OBJ embeddings peaked in layer 10. Probe accuracies indicated low construction information in layer 1, with over 90 percent accuracy from layer 2 onward, revealing latent construction information beyond GDV clustering. Fisher Discriminant Ratio (FDR) analysis of attention weights showed OBJ tokens were crucial for differentiating ASCs, followed by VERB and DET tokens. SUBJ, CLS, and SEP tokens had insignificant FDR scores. This study highlights BERT's layered processing of linguistic constructions and its differences from LSTMs. Future research will compare these findings with neuroimaging data to understand the neural correlates of ASC processing. This research underscores neural language models' potential to mirror linguistic processing in the human brain, offering insights into the computational and neural mechanisms underlying language understanding.