分享

Speaker diarization provides the answer to the question "who spoke when?" for an audio file. This information can be used to complete audio transcripts for further processing steps. Most speaker diarization systems assume that the audio file is available as a whole. However, there are scenarios in which the speaker labels are needed immediately after the arrival of an audio segment. Speaker diarization with a correspondingly low latency is referred to as online speaker diarization. This paper provides an overview. First the history of online speaker diarization is briefly presented. Next a taxonomy and datasets for training and evaluation are given. In the sections that follow, online diarization methods and systems are discussed in detail. This paper concludes with the presentation of challenges that still need to be solved by future research in the field of online speaker diarization.

Language models like BERT excel at sentence classification tasks due to extensive pre-training on general data, but their robustness to parameter corruption is unexplored. To understand this better, we look at what happens if a language model is "broken", in the sense that some of its parameters are corrupted and then recovered by fine-tuning. Strategically corrupting BERT variants at different levels, we find corrupted models struggle to fully recover their original performance, with higher corruption causing more severe degradation. Notably, bottom-layer corruption affecting fundamental linguistic features is more detrimental than top-layer corruption. Our insights contribute to understanding language model robustness and adaptability under adverse conditions, informing strategies for developing resilient NLP systems against parameter perturbations.

With the rising popularity of Transformer-based large language models (LLMs), reducing their high inference costs has become a significant research focus. One effective approach is to compress the long input contexts. Existing methods typically leverage the self-attention mechanism of the LLM itself for context compression. While these methods have achieved notable results, the compression process still involves quadratic time complexity, which limits their applicability. To mitigate this limitation, we propose the In-Context Former (IC-Former). Unlike previous methods, IC-Former does not depend on the target LLMs. Instead, it leverages the cross-attention mechanism and a small number of learnable digest tokens to directly condense information from the contextual word embeddings. This approach significantly reduces inference time, which achieves linear growth in time complexity within the compression range. Experimental results indicate that our method requires only 1/32 of the floating-point operations of the baseline during compression and improves processing speed by 68 to 112 times while achieving over 90% of the baseline performance on evaluation metrics. Overall, our model effectively reduces compression costs and makes real-time compression scenarios feasible.

In this paper, we tackle the task of distractor generation (DG) for multiple-choice questions. Our study introduces two key designs. First, we propose \textit{retrieval augmented pretraining}, which involves refining the language model pretraining to align it more closely with the downstream task of DG. Second, we explore the integration of knowledge graphs to enhance the performance of DG. Through experiments with benchmarking datasets, we show that our models significantly outperform the state-of-the-art results. Our best-performing model advances the F1@3 score from 14.80 to 16.47 in MCQ dataset and from 15.92 to 16.50 in Sciq dataset.

This study addresses the widening gap in Automatic Speech Recognition (ASR) research between high resource and extremely low resource languages, with a particular focus on Manchu, a critically endangered language. Manchu exemplifies the challenges faced by marginalized linguistic communities in accessing state-of-the-art technologies. In a pioneering effort, we introduce the first-ever Manchu ASR model ManWav, leveraging Wav2Vec2-XLSR-53. The results of the first Manchu ASR is promising, especially when trained with our augmented data. Wav2Vec2-XLSR-53 fine-tuned with augmented data demonstrates a 0.02 drop in CER and 0.13 drop in WER compared to the same base model fine-tuned with original data.

Children's speech recognition is considered a low-resource task mainly due to the lack of publicly available data. There are several reasons for such data scarcity, including expensive data collection and annotation processes, and data privacy, among others. Transforming speech signals into discrete tokens that do not carry sensitive information but capture both linguistic and acoustic information could be a solution for privacy concerns. In this study, we investigate the integration of discrete speech tokens into children's speech recognition systems as input without significantly degrading the ASR performance. Additionally, we explored single-view and multi-view strategies for creating these discrete labels. Furthermore, we tested the models for generalization capabilities with unseen domain and nativity dataset. Results reveal that the discrete token ASR for children achieves nearly equivalent performance with an approximate 83% reduction in parameters.

By leveraging the power of Large Language Models(LLMs) and speech foundation models, state of the art speech-text bimodal works can achieve challenging tasks like spoken translation(ST) and question answering(SQA) altogether with much simpler architectures. In this paper, we utilize the capability of Whisper encoder and pre-trained Yi-6B. Empirical results reveal that modal alignment can be achieved with one layer module and hundred hours of speech-text multitask corpus. We further swap the Yi-6B with human preferences aligned version of Yi-6B-Chat during inference, and discover that the alignment capability is applicable as well. In addition, the alignment subspace revealed by singular value decomposition(SVD) also implies linear alignment subspace is sparse, which leaves the possibility to concatenate other features like voice-print or video to expand modality.

Calibrating language models (LMs) aligns their generation confidence with the actual likelihood of answer correctness, which can inform users about LMs' reliability and mitigate hallucinated content. However, prior calibration methods, such as self-consistency-based and logit-based approaches, are either limited in inference-time efficiency or fall short of providing informative signals. Moreover, simply filtering out low-confidence responses reduces the LM's helpfulness when the answers are correct. Therefore, effectively using calibration techniques to enhance an LM's factuality remains an unsolved challenge. In this paper, we first propose an activation-based calibration method, ActCab, which trains a linear layer on top of the LM's last-layer activations that can better capture the representations of knowledge. Built on top of ActCab, we further propose CoDec, a confidence-guided decoding strategy to elicit truthful answers with high confidence from LMs. By evaluating on five popular QA benchmarks, ActCab achieves superior calibration performance than all competitive baselines, e.g., by reducing the average expected calibration error (ECE) score by up to 39%. Further experiments on CoDec show consistent improvements in several LMs' factuality on challenging QA datasets, such as TruthfulQA, highlighting the value of confidence signals in enhancing factuality.

Diffusion-based generative models (DGMs) have recently attracted attention in speech enhancement research (SE) as previous works showed a remarkable generalization capability. However, DGMs are also computationally intensive, as they usually require many iterations in the reverse diffusion process (RDP), making them impractical for streaming SE systems. In this paper, we propose to use discriminative scores from discriminative models in the first steps of the RDP. These discriminative scores require only one forward pass with the discriminative model for multiple RDP steps, thus greatly reducing computations. This approach also allows for performance improvements. We show that we can trade off between generative and discriminative capabilities as the number of steps with the discriminative score increases. Furthermore, we propose a novel streamable time-domain generative model with an algorithmic latency of 50 ms, which has no significant performance degradation compared to offline models.

Audio-driven talking face video generation has attracted increasing attention due to its huge industrial potential. Some previous methods focus on learning a direct mapping from audio to visual content. Despite progress, they often struggle with the ambiguity of the mapping process, leading to flawed results. An alternative strategy involves facial structural representations (e.g., facial landmarks) as intermediaries. This multi-stage approach better preserves the appearance details but suffers from error accumulation due to the independent optimization of different stages. Moreover, most previous methods rely on generative adversarial networks, prone to training instability and mode collapse. To address these challenges, our study proposes a novel landmark-based diffusion model for talking face generation, which leverages facial landmarks as intermediate representations while enabling end-to-end optimization. Specifically, we first establish the less ambiguous mapping from audio to landmark motion of lip and jaw. Then, we introduce an innovative conditioning module called TalkFormer to align the synthesized motion with the motion represented by landmarks via differentiable cross-attention, which enables end-to-end optimization for improved lip synchronization. Besides, TalkFormer employs implicit feature warping to align the reference image features with the target motion for preserving more appearance details. Extensive experiments demonstrate that our approach can synthesize high-fidelity and lip-synced talking face videos, preserving more subject appearance details from the reference image.

With the advancement of Self-supervised Learning (SSL) in speech-related tasks, there has been growing interest in utilizing discrete tokens generated by SSL for automatic speech recognition (ASR), as they offer faster processing techniques. However, previous studies primarily focused on multilingual ASR with Fbank features or English ASR with discrete tokens, leaving a gap in adapting discrete tokens for multilingual ASR scenarios. This study presents a comprehensive comparison of discrete tokens generated by various leading SSL models across multiple language domains. We aim to explore the performance and efficiency of speech discrete tokens across multiple language domains for both monolingual and multilingual ASR scenarios. Experimental results demonstrate that discrete tokens achieve comparable results against systems trained on Fbank features in ASR tasks across seven language domains with an average word error rate (WER) reduction of 0.31% and 1.76% absolute (2.80% and 15.70% relative) on dev and test sets respectively, with particularly WER reduction of 6.82% absolute (41.48% relative) on the Polish test set.

The use of synthetic speech as data augmentation is gaining increasing popularity in fields such as automatic speech recognition and speech classification tasks. Despite novel text-to-speech systems with voice cloning capabilities, that allow the usage of a larger amount of voices based on short audio segments, it is known that these systems tend to hallucinate and oftentimes produce bad data that will most likely have a negative impact on the downstream task. In the present work, we conduct a set of experiments around zero-shot learning with synthetic speech data for the specific task of speech commands classification. Our results on the Google Speech Commands dataset show that a simple ASR-based filtering method can have a big impact in the quality of the generated data, translating to a better performance. Furthermore, despite the good quality of the generated speech data, we also show that synthetic and real speech can still be easily distinguishable when using self-supervised (WavLM) features, an aspect further explored with a CycleGAN to bridge the gap between the two types of speech material.