A Comprehensive Analysis of the Effectiveness of Large Language Models as Automatic Dialogue Evaluators

Automatic evaluation is an integral aspect of dialogue system research. The traditional reference-based NLG metrics are generally found to be unsuitable for dialogue assessment. Consequently, recent studies have suggested various unique, reference-free neural metrics that better align with human evaluations. Notably among them, large language models (LLMs), particularly the instruction-tuned variants like ChatGPT, are shown to be promising substitutes for human judges. Yet, existing works on utilizing LLMs for automatic dialogue evaluation are limited in their scope in terms of the number of meta-evaluation datasets, mode of evaluation, coverage of LLMs, etc. Hence, it remains inconclusive how effective these LLMs are. To this end, we conduct a comprehensive study on the application of LLMs for automatic dialogue evaluation. Specifically, we analyze the multi-dimensional evaluation capability of 30 recently emerged LLMs at both turn and dialogue levels, using a comprehensive set of 12 meta-evaluation datasets. Additionally, we probe the robustness of the LLMs in handling various adversarial perturbations at both turn and dialogue levels. Finally, we explore how model-level and dimension-level ensembles impact the evaluation performance. All resources are available at https://github.com/e0397123/comp-analysis.

Emotion Rendering for Conversational Speech Synthesis with Heterogeneous Graph-Based Context Modeling

Conversational Speech Synthesis (CSS) aims to accurately express an utterance with the appropriate prosody and emotional inflection within a conversational setting. While recognising the significance of CSS task, the prior studies have not thoroughly investigated the emotional expressiveness problems due to the scarcity of emotional conversational datasets and the difficulty of stateful emotion modeling. In this paper, we propose a novel emotional CSS model, termed ECSS, that includes two main components: 1) to enhance emotion understanding, we introduce a heterogeneous graph-based emotional context modeling mechanism, which takes the multi-source dialogue history as input to model the dialogue context and learn the emotion cues from the context; 2) to achieve emotion rendering, we employ a contrastive learning-based emotion renderer module to infer the accurate emotion style for the target utterance. To address the issue of data scarcity, we meticulously create emotional labels in terms of category and intensity, and annotate additional emotional information on the existing conversational dataset (DailyTalk). Both objective and subjective evaluations suggest that our model outperforms the baseline models in understanding and rendering emotions. These evaluations also underscore the importance of comprehensive emotional annotations. Code and audio samples can be found at: https://github.com/walker-hyf/ECSS.

Amphion: An Open-Source Audio, Music and Speech Generation Toolkit

Amphion is a toolkit for Audio, Music, and Speech Generation. Its purpose is to support reproducible research and help junior researchers and engineers get started in the field of audio, music, and speech generation research and development. Amphion offers a unique feature: visualizations of classic models or architectures. We believe that these visualizations are beneficial for junior researchers and engineers who wish to gain a better understanding of the model. The North-Star objective of Amphion is to offer a platform for studying the conversion of any inputs into general audio. Amphion is designed to support individual generation tasks. In addition to the specific generation tasks, Amphion also includes several vocoders and evaluation metrics. A vocoder is an important module for producing high-quality audio signals, while evaluation metrics are critical for ensuring consistent metrics in generation tasks. In this paper, we provide a high-level overview of Amphion.

Golden Gemini is All You Need: Finding the Sweet Spots for Speaker Verification

Previous studies demonstrate the impressive performance of residual neural networks (ResNet) in speaker verification. The ResNet models treat the time and frequency dimensions equally. They follow the default stride configuration designed for image recognition, where the horizontal and vertical axes exhibit similarities. This approach ignores the fact that time and frequency are asymmetric in speech representation. In this paper, we address this issue and look for optimal stride configurations specifically tailored for speaker verification. We represent the stride space on a trellis diagram, and conduct a systematic study on the impact of temporal and frequency resolutions on the performance and further identify two optimal points, namely Golden Gemini, which serves as a guiding principle for designing 2D ResNet-based speaker verification models. By following the principle, a state-of-the-art ResNet baseline model gains a significant performance improvement on VoxCeleb, SITW, and CNCeleb datasets with 7.70%/11.76% average EER/minDCF reductions, respectively, across different network depths (ResNet18, 34, 50, and 101), while reducing the number of parameters by 16.5% and FLOPs by 4.1%. We refer to it as Gemini ResNet. Further investigation reveals the efficacy of the proposed Golden Gemini operating points across various training conditions and architectures. Furthermore, we present a new benchmark, namely the Gemini DF-ResNet, using a cutting-edge model.

HuatuoGPT-II, One-stage Training for Medical Adaption of LLMs

Adapting a language model into a specific domain, a.k.a `domain adaption', is a common practice when specialized knowledge, e.g. medicine, is not encapsulated in a general language model like Llama2. The challenge lies in the heterogeneity of data across the two training stages, as it varies in languages, genres, or formats. To tackle this and simplify the learning protocol, we propose to transform heterogeneous data, from the both pre-training and supervised stages, into a unified, simple input-output pair format. We validate the new protocol in the domains where proprietary LLMs like ChatGPT perform relatively poorly, such as Traditional Chinese Medicine. The developed model, HuatuoGPT-II, has shown state-of-the-art performance in Chinese medicine domain on a number of benchmarks, e.g. medical licensing exams. It even outperforms proprietary models like ChatGPT and GPT-4 in some aspects, especially in Traditional Chinese Medicine. Expert manual evaluations further validate HuatuoGPT-II's advantages over existing LLMs. Notably, HuatuoGPT-II was benchmarked in a fresh Chinese National Medical Licensing Examination where it achieved the best performance, showcasing not only its effectiveness but also its generalization capabilities.

How Well Do Text Embedding Models Understand Syntax?

Text embedding models have significantly contributed to advancements in natural language processing by adeptly capturing semantic properties of textual data. However, the ability of these models to generalize across a wide range of syntactic contexts remains under-explored. In this paper, we first develop an evaluation set, named \textbf{SR}, to scrutinize the capability for syntax understanding of text embedding models from two crucial syntactic aspects: Structural heuristics, and Relational understanding among concepts, as revealed by the performance gaps in previous studies. Our findings reveal that existing text embedding models have not sufficiently addressed these syntactic understanding challenges, and such ineffectiveness becomes even more apparent when evaluated against existing benchmark datasets. Furthermore, we conduct rigorous analysis to unearth factors that lead to such limitations and examine why previous evaluations fail to detect such ineffectiveness. Lastly, we propose strategies to augment the generalization ability of text embedding models in diverse syntactic scenarios. This study serves to highlight the hurdles associated with syntactic generalization and provides pragmatic guidance for boosting model performance across varied syntactic contexts.

Selective HuBERT: Self-Supervised Pre-Training for Target Speaker in Clean and Mixture Speech

Self-supervised pre-trained speech models were shown effective for various downstream speech processing tasks. Since they are mainly pre-trained to map input speech to pseudo-labels, the resulting representations are only effective for the type of pre-train data used, either clean or mixture speech. With the idea of selective auditory attention, we propose a novel pre-training solution called Selective-HuBERT, or SHuBERT, which learns the selective extraction of target speech representations from either clean or mixture speech. Specifically, SHuBERT is trained to predict pseudo labels of a target speaker, conditioned on an enrolled speech from the target speaker. By doing so, SHuBERT is expected to selectively attend to the target speaker in a complex acoustic environment, thus benefiting various downstream tasks. We further introduce a dual-path training strategy and use the cross-correlation constraint between the two branches to encourage the model to generate noise-invariant representation. Experiments on SUPERB benchmark and LibriMix dataset demonstrate the universality and noise-robustness of SHuBERT. Furthermore, we find that our high-quality representation can be easily integrated with conventional supervised learning methods to achieve significant performance, even under extremely low-resource labeled data.

Prompt-driven Target Speech Diarization

We introduce a novel task named `target speech diarization', which seeks to determine `when target event occurred' within an audio signal. We devise a neural architecture called Prompt-driven Target Speech Diarization (PTSD), that works with diverse prompts that specify the target speech events of interest. We train and evaluate PTSD using sim2spk, sim3spk and sim4spk datasets, which are derived from the Librispeech. We show that the proposed framework accurately localizes target speech events. Furthermore, our framework exhibits versatility through its impressive performance in three diarization-related tasks: target speaker voice activity detection, overlapped speech detection and gender diarization. In particular, PTSD achieves comparable performance to specialized models across these tasks on both real and simulated data. This work serves as a reference benchmark and provides valuable insights into prompt-driven target speech processing.

LC-TTFS: Towards Lossless Network Conversion for Spiking Neural Networks with TTFS Coding

The biological neurons use precise spike times, in addition to the spike firing rate, to communicate with each other. The time-to-first-spike (TTFS) coding is inspired by such biological observation. However, there is a lack of effective solutions for training TTFS-based spiking neural network (SNN). In this paper, we put forward a simple yet effective network conversion algorithm, which is referred to as LC-TTFS, by addressing two main problems that hinder an effective conversion from a high-performance artificial neural network (ANN) to a TTFS-based SNN. We show that our algorithm can achieve a near-perfect mapping between the activation values of an ANN and the spike times of an SNN on a number of challenging AI tasks, including image classification, image reconstruction, and speech enhancement. With TTFS coding, we can achieve up to orders of magnitude saving in computation over ANN and other rate-based SNNs. The study, therefore, paves the way for deploying ultra-low-power TTFS-based SNNs on power-constrained edge computing platforms.

Quantify Health-Related Atomic Knowledge in Chinese Medical Large Language Models: A Computational Analysis

Large Language Models (LLMs) have the potential to revolutionize the way users self-diagnose through search engines by offering direct and efficient suggestions. Recent studies primarily focused on the quality of LLMs evaluated by GPT-4 or their ability to pass medical exams, no studies have quantified the extent of health-related atomic knowledge stored in LLMs' memory, which is the basis of LLMs to provide more factual suggestions. In this paper, we first constructed a benchmark, including the most common types of atomic knowledge in user self-diagnosis queries, with 17 atomic types and a total of 14, 048 pieces of atomic knowledge. Then, we evaluated both generic and specialized LLMs on the benchmark. The experimental results showcased that generic LLMs perform better than specialized LLMs in terms of atomic knowledge and instruction-following ability. Error analysis revealed that both generic and specialized LLMs are sycophantic, e.g., always catering to users' claims when it comes to unknown knowledge. Besides, generic LLMs showed stronger safety, which can be learned by specialized LLMs through distilled data. We further explored different types of data commonly adopted for fine-tuning specialized LLMs, i.e., real-world, semi-distilled, and distilled data, and found that distilled data can benefit LLMs most.