Jee-weon Jung

Yifan Peng, Jinchuan Tian, Brian Yan et al. · cmu, meta-ai

Pre-training speech models on large volumes of data has achieved remarkable success. OpenAI Whisper is a multilingual multitask model trained on 680k hours of supervised speech data. It generalizes well to various speech recognition and translation benchmarks even in a zero-shot setup. However, the full pipeline for developing such models (from data collection to training) is not publicly accessible, which makes it difficult for researchers to further improve its performance and address training-related issues such as efficiency, robustness, fairness, and bias. This work presents an Open Whisper-style Speech Model (OWSM), which reproduces Whisper-style training using an open-source toolkit and publicly available data. OWSM even supports more translation directions and can be more efficient to train. We will publicly release all scripts used for data preparation, training, inference, and scoring as well as pre-trained models and training logs to promote open science.

Soumi Maiti, Yifan Peng, Shukjae Choi et al. · nvidia

We propose a decoder-only language model, VoxtLM, that can perform four tasks: speech recognition, speech synthesis, text generation, and speech continuation. VoxtLM integrates text vocabulary with discrete speech tokens from self-supervised speech features and uses special tokens to enable multitask learning. Compared to a single-task model, VoxtLM exhibits a significant improvement in speech synthesis, with improvements in both speech intelligibility from 28.9 to 5.6 and objective quality from 2.68 to 3.90. VoxtLM also improves speech generation and speech recognition performance over the single-task counterpart. Further, VoxtLM is trained with publicly available data and training recipes and model checkpoints are open-sourced to make fully reproducible work.

Samuele Cornell, Jee-weon Jung, Shinji Watanabe et al. · cmu

This paper presents a novel framework for joint speaker diarization (SD) and automatic speech recognition (ASR), named SLIDAR (sliding-window diarization-augmented recognition). SLIDAR can process arbitrary length inputs and can handle any number of speakers, effectively solving ``who spoke what, when'' concurrently. SLIDAR leverages a sliding window approach and consists of an end-to-end diarization-augmented speech transcription (E2E DAST) model which provides, locally, for each window: transcripts, diarization and speaker embeddings. The E2E DAST model is based on an encoder-decoder architecture and leverages recent techniques such as serialized output training and ``Whisper-style" prompting. The local outputs are then combined to get the final SD+ASR result by clustering the speaker embeddings to get global speaker identities. Experiments performed on monaural recordings from the AMI corpus confirm the effectiveness of the method in both close-talk and far-field speech scenarios.

Siddhant Arora, Hayato Futami, Jee-weon Jung et al. · nvidia

Recent studies leverage large language models with multi-tasking capabilities, using natural language prompts to guide the model's behavior and surpassing performance of task-specific models. Motivated by this, we ask: can we build a single model that jointly performs various spoken language understanding (SLU) tasks? We start by adapting a pre-trained automatic speech recognition model to additional tasks using single-token task specifiers. We enhance this approach through instruction tuning, i.e., finetuning by describing the task using natural language instructions followed by the list of label options. Our approach can generalize to new task descriptions for the seen tasks during inference, thereby enhancing its user-friendliness. We demonstrate the efficacy of our single multi-task learning model "UniverSLU" for 12 speech classification and sequence generation task types spanning 17 datasets and 9 languages. On most tasks, UniverSLU achieves competitive performance and often even surpasses task-specific models. Additionally, we assess the zero-shot capabilities, finding that the model generalizes to new datasets and languages for seen task types.

Jaesung Huh, Andrew Brown, Jee-weon Jung et al.

This paper summarises the findings from the VoxCeleb Speaker Recognition Challenge 2022 (VoxSRC-22), which was held in conjunction with INTERSPEECH 2022. The goal of this challenge was to evaluate how well state-of-the-art speaker recognition systems can diarise and recognise speakers from speech obtained "in the wild". The challenge consisted of: (i) the provision of publicly available speaker recognition and diarisation data from YouTube videos together with ground truth annotation and standardised evaluation software; and (ii) a public challenge and hybrid workshop held at INTERSPEECH 2022. We describe the four tracks of our challenge along with the baselines, methods, and results. We conclude with a discussion on the new domain-transfer focus of VoxSRC-22, and on the progression of the challenge from the previous three editions.

Xin Wang, Hector Delgado, Hemlata Tak et al.

ASVspoof 5 is the fifth edition in a series of challenges that promote the study of speech spoofing and deepfake attacks, and the design of detection solutions. Compared to previous challenges, the ASVspoof 5 database is built from crowdsourced data collected from a vastly greater number of speakers in diverse acoustic conditions. Attacks, also crowdsourced, are generated and tested using surrogate detection models, while adversarial attacks are incorporated for the first time. New metrics support the evaluation of spoofing-robust automatic speaker verification (SASV) as well as stand-alone detection solutions, i.e., countermeasures without ASV. We describe the two challenge tracks, the new database, the evaluation metrics, baselines, and the evaluation platform, and present a summary of the results. Attacks significantly compromise the baseline systems, while submissions bring substantial improvements.

Jee-weon Jung, Hee-Soo Heo, Bong-Jin Lee et al.

Speaker embedding extractors (EEs), which map input audio to a speaker discriminant latent space, are of paramount importance in speaker diarisation. However, there are several challenges when adopting EEs for diarisation, from which we tackle two key problems. First, the evaluation is not straightforward because the features required for better performance differ between speaker verification and diarisation. We show that better performance on widely adopted speaker verification evaluation protocols does not lead to better diarisation performance. Second, embedding extractors have not seen utterances in which multiple speakers exist. These inputs are inevitably present in speaker diarisation because of overlapped speech and speaker changes; they degrade the performance. To mitigate the first problem, we generate speaker verification evaluation protocols that mimic the diarisation scenario better. We propose two data augmentation techniques to alleviate the second problem, making embedding extractors aware of overlapped speech or speaker change input. One technique generates overlapped speech segments, and the other generates segments where two speakers utter sequentially. Extensive experimental results using three state-of-the-art speaker embedding extractors demonstrate that both proposed approaches are effective.

Jee-weon Jung, Hee-Soo Heo, Bong-Jin Lee et al.

The objective of this work is to develop a speaker recognition model to be used in diverse scenarios. We hypothesise that two components should be adequately configured to build such a model. First, adequate architecture would be required. We explore several recent state-of-the-art models, including ECAPA-TDNN and MFA-Conformer, as well as other baselines. Second, a massive amount of data would be required. We investigate several new training data configurations combining a few existing datasets. The most extensive configuration includes over 87k speakers' 10.22k hours of speech. Four evaluation protocols are adopted to measure how the trained model performs in diverse scenarios. Through experiments, we find that MFA-Conformer with the least inductive bias generalises the best. We also show that training with proposed large data configurations gives better performance. A boost in generalisation is observed, where the average performance on four evaluation protocols improves by more than 20%. In addition, we also demonstrate that these models' performances can improve even further when increasing capacity.

Jee-weon Jung, You Jin Kim, Hee-Soo Heo et al.

In recent years, speaker recognition systems based on raw waveform inputs have received increasing attention. However, the performance of such systems are typically inferior to the state-of-the-art handcrafted feature-based counterparts, which demonstrate equal error rates under 1% on the popular VoxCeleb1 test set. This paper proposes a novel speaker recognition model based on raw waveform inputs. The model incorporates recent advances in machine learning and speaker verification, including the Res2Net backbone module and multi-layer feature aggregation. Our best model achieves an equal error rate of 0.89%, which is competitive with the state-of-the-art models based on handcrafted features, and outperforms the best model based on raw waveform inputs by a large margin. We also explore the application of the proposed model in the context of self-supervised learning framework. Our self-supervised model outperforms single phase-based existing works in this line of research. Finally, we show that self-supervised pre-training is effective for the semi-supervised scenario where we only have a small set of labelled training data, along with a larger set of unlabelled examples.

Jee-weon Jung, Yihan Wu, Xin Wang et al.

This paper introduces SpoofCeleb, a dataset designed for Speech Deepfake Detection (SDD) and Spoofing-robust Automatic Speaker Verification (SASV), utilizing source data from real-world conditions and spoofing attacks generated by Text-To-Speech (TTS) systems also trained on the same real-world data. Robust recognition systems require speech data recorded in varied acoustic environments with different levels of noise to be trained. However, current datasets typically include clean, high-quality recordings (bona fide data) due to the requirements for TTS training; studio-quality or well-recorded read speech is typically necessary to train TTS models. Current SDD datasets also have limited usefulness for training SASV models due to insufficient speaker diversity. SpoofCeleb leverages a fully automated pipeline we developed that processes the VoxCeleb1 dataset, transforming it into a suitable form for TTS training. We subsequently train 23 contemporary TTS systems. SpoofCeleb comprises over 2.5 million utterances from 1,251 unique speakers, collected under natural, real-world conditions. The dataset includes carefully partitioned training, validation, and evaluation sets with well-controlled experimental protocols. We present the baseline results for both SDD and SASV tasks. All data, protocols, and baselines are publicly available at https://jungjee.github.io/spoofceleb.

Sung Hwan Mun, Jee-weon Jung, Min Hyun Han et al.

The majority of recent state-of-the-art speaker verification architectures adopt multi-scale processing and frequency-channel attention mechanisms. Convolutional layers of these models typically have a fixed kernel size, e.g., 3 or 5. In this study, we further contribute to this line of research utilising a selective kernel attention (SKA) mechanism. The SKA mechanism allows each convolutional layer to adaptively select the kernel size in a data-driven fashion. It is based on an attention mechanism which exploits both frequency and channel domain. We first apply existing SKA module to our baseline. Then we propose two SKA variants where the first variant is applied in front of the ECAPA-TDNN model and the other is combined with the Res2net backbone block. Through extensive experiments, we demonstrate that our two proposed SKA variants consistently improves the performance and are complementary when tested on three different evaluation protocols.

Jee-weon Jung, Wangyou Zhang, Soumi Maiti et al.

Traditional Text-to-Speech (TTS) systems rely on studio-quality speech recorded in controlled settings.a Recently, an effort known as noisy-TTS training has emerged, aiming to utilize in-the-wild data. However, the lack of dedicated datasets has been a significant limitation. We introduce the TTS In the Wild (TITW) dataset, which is publicly available, created through a fully automated pipeline applied to the VoxCeleb1 dataset. It comprises two training sets: TITW-Hard, derived from the transcription, segmentation, and selection of raw VoxCeleb1 data, and TITW-Easy, which incorporates additional enhancement and data selection based on DNSMOS. State-of-the-art TTS models achieve over 3.0 UTMOS score with TITW-Easy, while TITW-Hard remains difficult showing UTMOS below 2.8.

Jaesung Huh, Joon Son Chung, Arsha Nagrani et al.

The VoxCeleb Speaker Recognition Challenges (VoxSRC) were a series of challenges and workshops that ran annually from 2019 to 2023. The challenges primarily evaluated the tasks of speaker recognition and diarisation under various settings including: closed and open training data; as well as supervised, self-supervised, and semi-supervised training for domain adaptation. The challenges also provided publicly available training and evaluation datasets for each task and setting, with new test sets released each year. In this paper, we provide a review of these challenges that covers: what they explored; the methods developed by the challenge participants and how these evolved; and also the current state of the field for speaker verification and diarisation. We chart the progress in performance over the five installments of the challenge on a common evaluation dataset and provide a detailed analysis of how each year's special focus affected participants' performance. This paper is aimed both at researchers who want an overview of the speaker recognition and diarisation field, and also at challenge organisers who want to benefit from the successes and avoid the mistakes of the VoxSRC challenges. We end with a discussion of the current strengths of the field and open challenges. Project page : https://mm.kaist.ac.kr/datasets/voxceleb/voxsrc/workshop.html

Hee-Soo Heo, Youngki Kwon, Bong-Jin Lee et al.

Speaker embedding extractors significantly influence the performance of clustering-based speaker diarisation systems. Conventionally, only one embedding is extracted from each speech segment. However, because of the sliding window approach, a segment easily includes two or more speakers owing to speaker change points. This study proposes a novel embedding extractor architecture, referred to as a high-resolution embedding extractor (HEE), which extracts multiple high-resolution embeddings from each speech segment. Hee consists of a feature-map extractor and an enhancer, where the enhancer with the self-attention mechanism is the key to success. The enhancer of HEE replaces the aggregation process; instead of a global pooling layer, the enhancer combines relative information to each frame via attention leveraging the global context. Extracted dense frame-level embeddings can each represent a speaker. Thus, multiple speakers can be represented by different frame-level features in each segment. We also propose an artificially generating mixture data training framework to train the proposed HEE. Through experiments on five evaluation sets, including four public datasets, the proposed HEE demonstrates at least 10% improvement on each evaluation set, except for one dataset, which we analyse that rapid speaker changes less exist.

Zakaria Aldeneh, Skyler Seto, Maureen de Seyssel et al.

Modern ASR systems are typically trained on large-scale pseudo-labeled, in-the-wild data spanning multiple domains. While such heterogeneous data benefit generalist models designed for broad deployment, they pose challenges for specialist models targeting specific domains: specialist models lack the capacity to learn from all available data, and one must pay closer attention to addressing the mismatch between training and test conditions. In this work, we study targeted data selection as a strategy to address these challenges, selecting relevant subsets from 100k hours of in-the-wild training data to optimize performance on target domains. We represent speech samples using embeddings that capture complementary characteristic--speaker attributes, phonetic content, and semantic meaning--and analyze how relevance and diversity along these axes when performing data selection affect downstream ASR performance. Our experiments with CTC-based Conformer models show that training on a strategically selected 5% subset can exceed the performance of models trained on the full dataset by up to 36.8% relative WER reduction on target domains.

Jee-weon Jung, Soonshin Seo, Hee-Soo Heo et al.

The task of speaker change detection (SCD), which detects points where speakers change in an input, is essential for several applications. Several studies solved the SCD task using audio inputs only and have shown limited performance. Recently, multimodal SCD (MMSCD) models, which utilise text modality in addition to audio, have shown improved performance. In this study, the proposed model are built upon two main proposals, a novel mechanism for modality fusion and the adoption of a encoder-decoder architecture. Different to previous MMSCD works that extract speaker embeddings from extremely short audio segments, aligned to a single word, we use a speaker embedding extracted from 1.5s. A transformer decoder layer further improves the performance of an encoder-only MMSCD model. The proposed model achieves state-of-the-art results among studies that report SCD performance and is also on par with recent work that combines SCD with automatic speech recognition via human transcription.

Yifan Peng, Jinchuan Tian, William Chen et al. · nvidia

Recent studies have highlighted the importance of fully open foundation models. The Open Whisper-style Speech Model (OWSM) is an initial step towards reproducing OpenAI Whisper using public data and open-source toolkits. However, previous versions of OWSM (v1 to v3) are still based on standard Transformer, which might lead to inferior performance compared to state-of-the-art speech encoder architectures. This work aims to improve the performance and efficiency of OWSM without additional data. We present a series of E-Branchformer-based models named OWSM v3.1, ranging from 100M to 1B parameters. OWSM v3.1 outperforms its predecessor, OWSM v3, in most evaluation benchmarks, while showing an improved inference speed of up to 25%. We further reveal the emergent ability of OWSM v3.1 in zero-shot contextual biasing speech recognition. We also provide a model trained on a subset of data with low license restrictions. We will publicly release the code, pre-trained models, and training logs.

Jee-weon Jung, Wangyou Zhang, Jiatong Shi et al.

This paper introduces ESPnet-SPK, a toolkit designed with several objectives for training speaker embedding extractors. First, we provide an open-source platform for researchers in the speaker recognition community to effortlessly build models. We provide several models, ranging from x-vector to recent SKA-TDNN. Through the modularized architecture design, variants can be developed easily. We also aspire to bridge developed models with other domains, facilitating the broad research community to effortlessly incorporate state-of-the-art embedding extractors. Pre-trained embedding extractors can be accessed in an off-the-shelf manner and we demonstrate the toolkit's versatility by showcasing its integration with two tasks. Another goal is to integrate with diverse self-supervised learning features. We release a reproducible recipe that achieves an equal error rate of 0.39% on the Vox1-O evaluation protocol using WavLM-Large with ECAPA-TDNN.

Jee-weon Jung, Roshan Sharma, William Chen et al.

Abstractive speech summarization (SSUM) aims to generate human-like summaries from speech. Given variations in information captured and phrasing, recordings can be summarized in multiple ways. Therefore, it is more reasonable to consider a probabilistic distribution of all potential summaries rather than a single summary. However, conventional SSUM models are mostly trained and evaluated with a single ground-truth (GT) human-annotated deterministic summary for every recording. Generating multiple human references would be ideal to better represent the distribution statistically, but is impractical because annotation is expensive. We tackle this challenge by proposing AugSumm, a method to leverage large language models (LLMs) as a proxy for human annotators to generate augmented summaries for training and evaluation. First, we explore prompting strategies to generate synthetic summaries from ChatGPT. We validate the quality of synthetic summaries using multiple metrics including human evaluation, where we find that summaries generated using AugSumm are perceived as more valid to humans. Second, we develop methods to utilize synthetic summaries in training and evaluation. Experiments on How2 demonstrate that pre-training on synthetic summaries and fine-tuning on GT summaries improves ROUGE-L by 1 point on both GT and AugSumm-based test sets. AugSumm summaries are available at https://github.com/Jungjee/AugSumm.

KiHyun Nam, Jungwoo Heo, Jee-weon Jung et al.

A primary challenge when deploying speaker recognition systems in real-world applications is performance degradation caused by environmental mismatch. We propose a diffusion-based method that takes speaker embeddings extracted from a pre-trained speaker recognition model and generates refined embeddings. For training, our approach progressively adds Gaussian noise to both clean and noisy speaker embeddings extracted from clean and noisy speech, respectively, via forward process of a diffusion model, and then reconstructs them to clean embeddings in the reverse process. While inferencing, all embeddings are regenerated via diffusion process. Our method needs neither speaker label nor any modification to the existing speaker recognition pipeline. Experiments on evaluation sets simulating environment mismatch scenarios show that our method can improve recognition accuracy by up to 19.6% over baseline models while retaining performance on conventional scenarios. We publish our code here https://github.com/kaistmm/seed-pytorch

Siddhant Arora, Ankita Pasad, Chung-Ming Chien et al.

The Spoken Language Understanding Evaluation (SLUE) suite of benchmark tasks was recently introduced to address the need for open resources and benchmarking of complex spoken language understanding (SLU) tasks, including both classification and sequence generation tasks, on natural speech. The benchmark has demonstrated preliminary success in using pre-trained speech foundation models (SFM) for these SLU tasks. However, the community still lacks a fine-grained understanding of the comparative utility of different SFMs. Inspired by this, we ask: which SFMs offer the most benefits for these complex SLU tasks, and what is the most effective approach for incorporating these SFMs? To answer this, we perform an extensive evaluation of multiple supervised and self-supervised SFMs using several evaluation protocols: (i) frozen SFMs with a lightweight prediction head, (ii) frozen SFMs with a complex prediction head, and (iii) fine-tuned SFMs with a lightweight prediction head. Although the supervised SFMs are pre-trained on much more speech recognition data (with labels), they do not always outperform self-supervised SFMs; the latter tend to perform at least as well as, and sometimes better than, supervised SFMs, especially on the sequence generation tasks in SLUE. While there is no universally optimal way of incorporating SFMs, the complex prediction head gives the best performance for most tasks, although it increases the inference time. We also introduce an open-source toolkit and performance leaderboard, SLUE-PERB, for these tasks and modeling strategies.

Hemlata Tak, Jee-weon Jung, Jose Patino et al.

Artefacts that serve to distinguish bona fide speech from spoofed or deepfake speech are known to reside in specific subbands and temporal segments. Various approaches can be used to capture and model such artefacts, however, none works well across a spectrum of diverse spoofing attacks. Reliable detection then often depends upon the fusion of multiple detection systems, each tuned to detect different forms of attack. In this paper we show that better performance can be achieved when the fusion is performed within the model itself and when the representation is learned automatically from raw waveform inputs. The principal contribution is a spectro-temporal graph attention network (GAT) which learns the relationship between cues spanning different sub-bands and temporal intervals. Using a model-level graph fusion of spectral (S) and temporal (T) sub-graphs and a graph pooling strategy to improve discrimination, the proposed RawGAT-ST model achieves an equal error rate of 1.06 % for the ASVspoof 2019 logical access database. This is one of the best results reported to date and is reproducible using an open source implementation.

Hye-jin Shim, Jee-weon Jung, Tomi Kinnunen et al.

Spoofing detection is today a mainstream research topic. Standard metrics can be applied to evaluate the performance of isolated spoofing detection solutions and others have been proposed to support their evaluation when they are combined with speaker detection. These either have well-known deficiencies or restrict the architectural approach to combine speaker and spoof detectors. In this paper, we propose an architecture-agnostic detection cost function (a-DCF). A generalisation of the original DCF used widely for the assessment of automatic speaker verification (ASV), the a-DCF is designed for the evaluation of spoofing-robust ASV. Like the DCF, the a-DCF reflects the cost of decisions in a Bayes risk sense, with explicitly defined class priors and detection cost model. We demonstrate the merit of the a-DCF through the benchmarking evaluation of architecturally-heterogeneous spoofing-robust ASV solutions.

Minsu Kim, Jee-weon Jung, Hyeongseop Rha et al.

The capability to jointly process multi-modal information is becoming an essential task. However, the limited number of paired multi-modal data and the large computational requirements in multi-modal learning hinder the development. We propose a novel Tri-Modal Translation (TMT) model that translates between arbitrary modalities spanning speech, image, and text. We introduce a novel viewpoint, where we interpret different modalities as different languages, and treat multi-modal translation as a well-established machine translation problem. To this end, we tokenize speech and image data into discrete tokens, which provide a unified interface across modalities and significantly decrease the computational cost. In the proposed TMT, a multi-modal encoder-decoder conducts the core translation, whereas modality-specific processing is conducted only within the tokenization and detokenization stages. We evaluate the proposed TMT on all six modality translation tasks. TMT outperforms single model counterparts consistently, demonstrating that unifying tasks is beneficial not only for practicality but also for performance.

Kwanghee Choi, Jee-weon Jung, Shinji Watanabe

With the success of self-supervised representations, researchers seek a better understanding of the information encapsulated within a representation. Among various interpretability methods, we focus on classification-based linear probing. We aim to foster a solid understanding and provide guidelines for linear probing by constructing a novel mathematical framework leveraging information theory. First, we connect probing with the variational bounds of mutual information (MI) to relax the probe design, equating linear probing with fine-tuning. Then, we investigate empirical behaviors and practices of probing through our mathematical framework. We analyze the layer-wise performance curve being convex, which seemingly violates the data processing inequality. However, we show that the intermediate representations can have the biggest MI estimate because of the tradeoff between better separability and decreasing MI. We further suggest that the margin of linearly separable representations can be a criterion for measuring the "goodness of representation." We also compare accuracy with MI as the measuring criteria. Finally, we empirically validate our claims by observing the self-supervised speech models on retaining word and phoneme information.

Siddhant Arora, Jinchuan Tian, Hayato Futami et al.

Unlike traditional cascaded pipelines, end-to-end (E2E) spoken dialogue systems preserve full differentiability and capture non-phonemic information, making them well-suited for modeling spoken interactions. However, existing E2E approaches often require large-scale training data and generates responses lacking semantic coherence. We propose a simple yet effective strategy leveraging a chain-of-thought (CoT) formulation, ensuring that training on conversational data remains closely aligned with the multimodal language model (LM)'s pre-training on speech recognition~(ASR), text-to-speech synthesis (TTS), and text LM tasks. Our method achieves over 1.5 ROUGE-1 improvement over the baseline, successfully training spoken dialogue systems on publicly available human-human conversation datasets, while being compute-efficient enough to train on just 300 hours of public human-human conversation data, such as the Switchboard. We will publicly release our models and training code.

Yihan Wu, Jee-weon Jung, Hye-jin Shim et al.

The WildSpoof Challenge aims to advance the use of in-the-wild data in two intertwined speech processing tasks. It consists of two parallel tracks: (1) Text-to-Speech (TTS) synthesis for generating spoofed speech, and (2) Spoofing-robust Automatic Speaker Verification (SASV) for detecting spoofed speech. While the organizers coordinate both tracks and define the data protocols, participants treat them as separate and independent tasks. The primary objectives of the challenge are: (i) to promote the use of in-the-wild data for both TTS and SASV, moving beyond conventional clean and controlled datasets and considering real-world scenarios; and (ii) to encourage interdisciplinary collaboration between the spoofing generation (TTS) and spoofing detection (SASV) communities, thereby fostering the development of more integrated, robust, and realistic systems.

Hye-jin Shim, Md Sahidullah, Jee-weon Jung et al.

Current trends in audio anti-spoofing detection research strive to improve models' ability to generalize across unseen attacks by learning to identify a variety of spoofing artifacts. This emphasis has primarily focused on the spoof class. Recently, several studies have noted that the distribution of silence differs between the two classes, which can serve as a shortcut. In this paper, we extend class-wise interpretations beyond silence. We employ loss analysis and asymmetric methodologies to move away from traditional attack-focused and result-oriented evaluations towards a deeper examination of model behaviors. Our investigations highlight the significant differences in training dynamics between the two classes, emphasizing the need for future research to focus on robust modeling of the bonafide class.

Jee-weon Jung, Xin Wang, Nicholas Evans et al.

The current automatic speaker verification (ASV) task involves making binary decisions on two types of trials: target and non-target. However, emerging advancements in speech generation technology pose significant threats to the reliability of ASV systems. This study investigates whether ASV effortlessly acquires robustness against spoofing attacks (i.e., zero-shot capability) by systematically exploring diverse ASV systems and spoofing attacks, ranging from traditional to cutting-edge techniques. Through extensive analyses conducted on eight distinct ASV systems and 29 spoofing attack systems, we demonstrate that the evolution of ASV inherently incorporates defense mechanisms against spoofing attacks. Nevertheless, our findings also underscore that the advancement of spoofing attacks far outpaces that of ASV systems, hence necessitating further research on spoofing-robust ASV methodologies.

Hye-jin Shim, Jee-weon Jung, Tomi Kinnunen

Audio anti-spoofing for automatic speaker verification aims to safeguard users' identities from spoofing attacks. Although state-of-the-art spoofing countermeasure(CM) models perform well on specific datasets, they lack generalization when evaluated with different datasets. To address this limitation, previous studies have explored large pre-trained models, which require significant resources and time. We aim to develop a compact but well-generalizing CM model that can compete with large pre-trained models. Our approach involves multi-dataset co-training and sharpness-aware minimization, which has not been investigated in this domain. Extensive experiments reveal that proposed method yield competitive results across various datasets while utilizing 4,000 times less parameters than the large pre-trained models.

Sung Hwan Mun, Hye-jin Shim, Hemlata Tak et al.

This study aims to develop a single integrated spoofing-aware speaker verification (SASV) embeddings that satisfy two aspects. First, rejecting non-target speakers' input as well as target speakers' spoofed inputs should be addressed. Second, competitive performance should be demonstrated compared to the fusion of automatic speaker verification (ASV) and countermeasure (CM) embeddings, which outperformed single embedding solutions by a large margin in the SASV2022 challenge. We analyze that the inferior performance of single SASV embeddings comes from insufficient amount of training data and distinct nature of ASV and CM tasks. To this end, we propose a novel framework that includes multi-stage training and a combination of loss functions. Copy synthesis, combined with several vocoders, is also exploited to address the lack of spoofed data. Experimental results show dramatic improvements, achieving a SASV-EER of 1.06% on the evaluation protocol of the SASV2022 challenge.

Hemlata Tak, Massimiliano Todisco, Xin Wang et al.

The performance of spoofing countermeasure systems depends fundamentally upon the use of sufficiently representative training data. With this usually being limited, current solutions typically lack generalisation to attacks encountered in the wild. Strategies to improve reliability in the face of uncontrolled, unpredictable attacks are hence needed. We report in this paper our efforts to use self-supervised learning in the form of a wav2vec 2.0 front-end with fine tuning. Despite initial base representations being learned using only bona fide data and no spoofed data, we obtain the lowest equal error rates reported in the literature for both the ASVspoof 2021 Logical Access and Deepfake databases. When combined with data augmentation,these results correspond to an improvement of almost 90% relative to our baseline system.

Jee-weon Jung, Hemlata Tak, Hye-jin Shim et al.

ASV (automatic speaker verification) systems are intrinsically required to reject both non-target (e.g., voice uttered by different speaker) and spoofed (e.g., synthesised or converted) inputs. However, there is little consideration for how ASV systems themselves should be adapted when they are expected to encounter spoofing attacks, nor when they operate in tandem with CMs (spoofing countermeasures), much less how both systems should be jointly optimised. The goal of the first SASV (spoofing-aware speaker verification) challenge, a special sesscion in ISCA INTERSPEECH 2022, is to promote development of integrated systems that can perform ASV and CM simultaneously.

You Jin Kim, Hee-Soo Heo, Jee-weon Jung et al.

The objective of this work is to train noise-robust speaker embeddings adapted for speaker diarisation. Speaker embeddings play a crucial role in the performance of diarisation systems, but they often capture spurious information such as noise, adversely affecting performance. Our previous work has proposed an auto-encoder-based dimensionality reduction module to help remove the redundant information. However, they do not explicitly separate such information and have also been found to be sensitive to hyper-parameter values. To this end, we propose two contributions to overcome these issues: (i) a novel dimensionality reduction framework that can disentangle spurious information from the speaker embeddings; (ii) the use of speech activity vector to prevent the speaker code from representing the background noise. Through a range of experiments conducted on four datasets, our approach consistently demonstrates the state-of-the-art performance among models without system fusion.

Youngki Kwon, Hee-Soo Heo, Jee-weon Jung et al.

The objective of this work is effective speaker diarisation using multi-scale speaker embeddings. Typically, there is a trade-off between the ability to recognise short speaker segments and the discriminative power of the embedding, according to the segment length used for embedding extraction. To this end, recent works have proposed the use of multi-scale embeddings where segments with varying lengths are used. However, the scores are combined using a weighted summation scheme where the weights are fixed after the training phase, whereas the importance of segment lengths can differ with in a single session. To address this issue, we present three key contributions in this paper: (1) we propose graph attention networks for multi-scale speaker diarisation; (2) we design scale indicators to utilise scale information of each embedding; (3) we adapt the attention-based aggregation to utilise a pre-computed affinity matrix from multi-scale embeddings. We demonstrate the effectiveness of our method in various datasets where the speaker confusion which constitutes the primary metric drops over 10% in average relative compared to the baseline.

Jee-weon Jung, Hee-Soo Heo, Hemlata Tak et al.

Artefacts that differentiate spoofed from bona-fide utterances can reside in spectral or temporal domains. Their reliable detection usually depends upon computationally demanding ensemble systems where each subsystem is tuned to some specific artefacts. We seek to develop an efficient, single system that can detect a broad range of different spoofing attacks without score-level ensembles. We propose a novel heterogeneous stacking graph attention layer which models artefacts spanning heterogeneous temporal and spectral domains with a heterogeneous attention mechanism and a stack node. With a new max graph operation that involves a competitive mechanism and an extended readout scheme, our approach, named AASIST, outperforms the current state-of-the-art by 20% relative. Even a lightweight variant, AASIST-L, with only 85K parameters, outperforms all competing systems.

Hye-jin Shim, Jee-weon Jung, Ju-ho Kim et al.

Various attention mechanisms are being widely applied to acoustic scene classification. However, we empirically found that the attention mechanism can excessively discard potentially valuable information, despite improving performance. We propose the attentive max feature map that combines two effective techniques, attention and a max feature map, to further elaborate the attention mechanism and mitigate the above-mentioned phenomenon. We also explore various joint training methods, including multi-task learning, that allocate additional abstract labels for each audio recording. Our proposed system demonstrates state-of-the-art performance for single systems on Subtask A of the DCASE 2020 challenge by applying the two proposed techniques using relatively fewer parameters. Furthermore, adopting the proposed attentive max feature map, our team placed fourth in the recent DCASE 2021 challenge.

Ju-ho Kim, Hye-jin Shim, Jee-weon Jung et al.

Most speaker verification tasks are studied as an open-set evaluation scenario considering the real-world condition. Thus, the generalization power to unseen speakers is of paramount important to the performance of the speaker verification system. We propose to apply \textit {Mean Teacher}, a temporal averaging model, to extract speaker embeddings with small intra-class variance and large inter-class variance. The mean teacher network refers to the temporal averaging of deep neural network parameters; it can produces more accurate and stable representations than using weights after the training finished. By learning the reliable intermediate representation of the mean teacher network, we expect that the proposed method can explore more discriminatory embedding spaces and improve the generalization performance of the speaker verification system. Experimental results on the VoxCeleb1 test set demonstrate that the proposed method relatively improves performance by 11.61\%, compared to a baseline system.

Hemlata Tak, Jee-weon Jung, Jose Patino et al.

The cues needed to detect spoofing attacks against automatic speaker verification are often located in specific spectral sub-bands or temporal segments. Previous works show the potential to learn these using either spectral or temporal self-attention mechanisms but not the relationships between neighbouring sub-bands or segments. This paper reports our use of graph attention networks (GATs) to model these relationships and to improve spoofing detection performance. GATs leverage a self-attention mechanism over graph structured data to model the data manifold and the relationships between nodes. Our graph is constructed from representations produced by a ResNet. Nodes in the graph represent information either in specific sub-bands or temporal segments. Experiments performed on the ASVspoof 2019 logical access database show that our GAT-based model with temporal attention outperforms all of our baseline single systems. Furthermore, GAT-based systems are complementary to a set of existing systems. The fusion of GAT-based models with more conventional countermeasures delivers a 47% relative improvement in performance compared to the best performing single GAT system.

Youngki Kwon, Jee-weon Jung, Hee-Soo Heo et al.

The goal of this paper is to adapt speaker embeddings for solving the problem of speaker diarisation. The quality of speaker embeddings is paramount to the performance of speaker diarisation systems. Despite this, prior works in the field have directly used embeddings designed only to be effective on the speaker verification task. In this paper, we propose three techniques that can be used to better adapt the speaker embeddings for diarisation: dimensionality reduction, attention-based embedding aggregation, and non-speech clustering. A wide range of experiments is performed on various challenging datasets. The results demonstrate that all three techniques contribute positively to the performance of the diarisation system achieving an average relative improvement of 25.07% in terms of diarisation error rate over the baseline.

Jee-weon Jung, Hee-Soo Heo, Youngki Kwon et al.

In this work, we propose an overlapped speech detection system trained as a three-class classifier. Unlike conventional systems that perform binary classification as to whether or not a frame contains overlapped speech, the proposed approach classifies into three classes: non-speech, single speaker speech, and overlapped speech. By training a network with the more detailed label definition, the model can learn a better notion on deciding the number of speakers included in a given frame. A convolutional recurrent neural network architecture is explored to benefit from both convolutional layer's capability to model local patterns and recurrent layer's ability to model sequential information. The proposed overlapped speech detection model establishes a state-of-the-art performance with a precision of 0.6648 and a recall of 0.3222 on the DIHARD II evaluation set, showing a 20% increase in recall along with higher precision. In addition, we also introduce a simple approach to utilize the proposed overlapped speech detection model for speaker diarization which ranked third place in the Track 1 of the DIHARD III challenge.

Jee-weon Jung, Hee-Soo Heo, Ha-Jin Yu et al.

This work presents a novel back-end framework for speaker verification using graph attention networks. Segment-wise speaker embeddings extracted from multiple crops within an utterance are interpreted as node representations of a graph. The proposed framework inputs segment-wise speaker embeddings from an enrollment and a test utterance and directly outputs a similarity score. We first construct a graph using segment-wise speaker embeddings and then input these to graph attention networks. After a few graph attention layers with residual connections, each node is projected into a one-dimensional space using affine transform, followed by a readout operation resulting in a scalar similarity score. To enable successful adaptation for speaker verification, we propose techniques such as separating trainable weights for attention map calculations between segment-wise speaker embeddings from different utterances. The effectiveness of the proposed framework is validated using three different speaker embedding extractors trained with different architectures and objective functions. Experimental results demonstrate consistent improvement over various baseline back-end classifiers, with an average equal error rate improvement of 20% over the cosine similarity back-end without test time augmentation.

Hye-jin Shim, Jee-weon Jung, Ju-ho Kim et al.

Frequently misclassified pairs of classes that share many common acoustic properties exist in acoustic scene classification (ASC). To distinguish such pairs of classes, trivial details scattered throughout the data could be vital clues. However, these details are less noticeable and are easily removed using conventional non-linear activations (e.g. ReLU). Furthermore, making design choices to emphasize trivial details can easily lead to overfitting if the system is not sufficiently generalized. In this study, based on the analysis of the ASC task's characteristics, we investigate various methods to capture discriminative information and simultaneously mitigate the overfitting problem. We adopt a max feature map method to replace conventional non-linear activations in a deep neural network, and therefore, we apply an element-wise comparison between different filters of a convolution layer's output. Two data augment methods and two deep architecture modules are further explored to reduce overfitting and sustain the system's discriminative power. Various experiments are conducted using the detection and classification of acoustic scenes and events 2020 task1-a dataset to validate the proposed methods. Our results show that the proposed system consistently outperforms the baseline, where the single best performing system has an accuracy of 70.4% compared to 65.1% of the baseline.

Hye-jin Shim, Jee-weon Jung, Ju-ho Kim et al.

A number of studies have successfully developed speaker verification or presentation attack detection systems. However, studies integrating the two tasks remain in the preliminary stages. In this paper, we propose two approaches for building an integrated system of speaker verification and presentation attack detection: an end-to-end monolithic approach and a back-end modular approach. The first approach simultaneously trains speaker identification, presentation attack detection, and the integrated system using multi-task learning using a common feature. However, through experiments, we hypothesize that the information required for performing speaker verification and presentation attack detection might differ because speaker verification systems try to remove device-specific information from speaker embeddings, while presentation attack detection systems exploit such information. Therefore, we propose a back-end modular approach using a separate deep neural network (DNN) for speaker verification and presentation attack detection. This approach has thee input components: two speaker embeddings (for enrollment and test each) and prediction of presentation attacks. Experiments are conducted using the ASVspoof 2017-v2 dataset, which includes official trials on the integration of speaker verification and presentation attack detection. The proposed back-end approach demonstrates a relative improvement of 21.77% in terms of the equal error rate for integrated trials compared to a conventional speaker verification system.

Seung-bin Kim, Jee-weon Jung, Hye-jin Shim et al.

Most studies on speaker verification systems focus on long-duration utterances, which are composed of sufficient phonetic information. However, the performances of these systems are known to degrade when short-duration utterances are inputted due to the lack of phonetic information as compared to the long utterances. In this paper, we propose a method that compensates for the performance degradation of speaker verification for short utterances, referred to as "segment aggregation". The proposed method adopts an ensemble-based design to improve the stability and accuracy of speaker verification systems. The proposed method segments an input utterance into several short utterances and then aggregates the segment embeddings extracted from the segmented inputs to compose a speaker embedding. Then, this method simultaneously trains the segment embeddings and the aggregated speaker embedding. In addition, we also modified the teacher-student learning method for the proposed method. Experimental results on different input duration using the VoxCeleb1 test set demonstrate that the proposed technique improves speaker verification performance by about 45.37% relatively compared to the baseline system with 1-second test utterance condition.

Jee-weon Jung, Seung-bin Kim, Hye-jin Shim et al.

Recent advances in deep learning have facilitated the design of speaker verification systems that directly input raw waveforms. For example, RawNet extracts speaker embeddings from raw waveforms, which simplifies the process pipeline and demonstrates competitive performance. In this study, we improve RawNet by scaling feature maps using various methods. The proposed mechanism utilizes a scale vector that adopts a sigmoid non-linear function. It refers to a vector with dimensionality equal to the number of filters in a given feature map. Using a scale vector, we propose to scale the feature map multiplicatively, additively, or both. In addition, we investigate replacing the first convolution layer with the sinc-convolution layer of SincNet. Experiments performed on the VoxCeleb1 evaluation dataset demonstrate the effectiveness of the proposed methods, and the best performing system reduces the equal error rate by half compared to the original RawNet. Expanded evaluation results obtained using the VoxCeleb1-E and VoxCeleb-H protocols marginally outperform existing state-of-the-art systems.

Jee-weon Jung, Hye-jin Shim, Hee-Soo Heo et al.

In this study, we analyze the role of various categories of subsidiary information in conducting replay attack spoofing detection: `Room Size', `Reverberation', `Speaker-to-ASV distance, `Attacker-to-Speaker distance', and `Replay Device Quality'. As a means of analyzing subsidiary information, we use two frameworks to either subtract or include a category of subsidiary information to the code extracted from a deep neural network. For subtraction, we utilize an adversarial process framework which makes the code orthogonal to the basis vectors of the subsidiary information. For addition, we utilize the multi-task learning framework to include subsidiary information to the code. All experiments are conducted using the ASVspoof 2019 physical access scenario with the provided meta data. Through the analysis of the result of the two approaches, we conclude that various categories of subsidiary information does not reside enough in the code when the deep neural network is trained for binary classification. Explicitly including various categories of subsidiary information through the multi-task learning framework can help improve performance in closed set condition.

Hye-jin Shim, Hee-Soo Heo, Jee-weon Jung et al.

Constructing a dataset for replay spoofing detection requires a physical process of playing an utterance and re-recording it, presenting a challenge to the collection of large-scale datasets. In this study, we propose a self-supervised framework for pretraining acoustic configurations using datasets published for other tasks, such as speaker verification. Here, acoustic configurations refer to the environmental factors generated during the process of voice recording but not the voice itself, including microphone types, place and ambient noise levels. Specifically, we select pairs of segments from utterances and train deep neural networks to determine whether the acoustic configurations of the two segments are identical. We validate the effectiveness of the proposed method based on the ASVspoof 2019 physical access dataset utilizing two well-performing systems. The experimental results demonstrate that the proposed method outperforms the baseline approach by 30%.

Hee-Soo Heo, Jee-weon Jung, Hye-jin Shim et al.

An adversarial process between two deep neural networks is a promising approach to train a robust model. In this paper, we propose an adversarial process using cosine similarity, whereas conventional adversarial processes are based on inverted categorical cross entropy (CCE). When used for training an identification model, the adversarial process induces the competition of two discriminative models; one for a primary task such as speaker identification or image recognition, the other one for a subsidiary task such as channel identification or domain identification. In particular, the adversarial process degrades the performance of the subsidiary model by eliminating the subsidiary information in the input which, in assumption, may degrade the performance of the primary model. The conventional adversarial processes maximize the CCE of the subsidiary model to degrade the performance. We have studied a framework for training robust discriminative models by eliminating channel or domain information (subsidiary information) by applying such an adversarial process. However, we found through experiments that using the process of maximizing the CCE does not guarantee the performance degradation of the subsidiary model. In the proposed adversarial process using cosine similarity, on the contrary, the performance of the subsidiary model can be degraded more efficiently by searching feature space orthogonal to the subsidiary model. The experiments on speaker identification and image recognition show that we found features that make the outputs of the subsidiary models independent of the input, and the performances of the primary models are improved.

Hee-Soo Heo, Jee-weon Jung, Hye-jin Shim et al.

Acoustic scene classification identifies an input segment into one of the pre-defined classes using spectral information. The spectral information of acoustic scenes may not be mutually exclusive due to common acoustic properties across different classes, such as babble noises included in both airports and shopping malls. However, conventional training procedure based on one-hot labels does not consider the similarities between different acoustic scenes. We exploit teacher-student learning with the purpose to derive soft-labels that consider common acoustic properties among different acoustic scenes. In teacher-student learning, the teacher network produces soft-labels, based on which the student network is trained. We investigate various methods to extract soft-labels that better represent similarities across different scenes. Such attempts include extracting soft-labels from multiple audio segments that are defined as an identical acoustic scene. Experimental results demonstrate the potential of our approach, showing a classification accuracy of 77.36 % on the DCASE 2018 task 1 validation set.