arXiv subjects

Jungwoo Heo

Publications and source records attributed to Jungwoo Heo.

15 recordsLinked to original sources

arXiv2026

SpeakerLLM: A Speaker-Specialized Audio-LLM for Speaker Understanding and Verification Reasoning

As audio-first agents become increasingly common in physical AI, conversational robots, and screenless wearables, audio large language models (audio-LLMs) must integrate speaker-specific understanding to support user authorization, personalization, and context-aware interaction. This requires modeling who is speaking, how the voice sounds, and how recording conditions affect speaker cues. Conventional speaker verification systems provide strong scalar scores but little linguistic evidence, while current audio-LLMs and speaker-aware language models have limited ability to organize speaker information beyond binary labels or descriptive profiles. We present SpeakerLLM, a speaker-specialized audio-LLM framework that unifies single-utterance speaker profiling, recording-condition understanding, utterance-pair speaker comparison, and evidence-organized verification reasoning within a natural-language interface. We construct verification-reasoning targets and a decision-composition policy that separate profile-level evidence from the final same-or-different decision and organize recording condition, profile evidence, and the decision into a structured trace. At its core, SpeakerLLM uses a hierarchical speaker tokenizer designed to capture multiple granularities of speaker evidence. Utterance-level speaker embeddings summarize identity and profile-level cues, whereas frame-level speaker features preserve fine-grained acoustic descriptors. Experiments show that SpeakerLLM-Base improves speaker-profile and recording-condition understanding over general audio-LLMs, while SpeakerLLM-VR preserves strong generated-verdict accuracy and produces decision traces grounded in the supervised verification reasoning schema. We will release the metadata-enriched supervision dataset and target-construction code for reproducibility.

cs.SD↗

arXiv2025

Diffusion-Link: Diffusion Probabilistic Model for Bridging the Audio-Text Modality Gap

Contrastive audio-language pretraining yields powerful joint representations, yet a persistent audio-text modality gap limits the benefits of coupling multimodal encoders with large language models (LLMs). We present Diffusion-Link, a diffusion-based modality-bridging module that generatively maps audio embeddings into the text-embedding distribution. The module is trained at the output embedding from the frozen multimodal encoder and implemented as a lightweight network with three residual MLP blocks. To assess the effect of Diffusion-Link on multimodal encoder-LLM coupling, we evaluate on Automatic Audio Captioning (AAC); to our knowledge, this is the first application of diffusion-based modality bridging to AAC. We report two results. (1) Modality-gap analysis: on similarity and geometric criteria, Diffusion-Link reduces the modality gap the most among prior diffusion-based methods and shows a collective migration of audio embeddings toward the text distribution. (2) Downstream AAC: attaching Diffusion-Link to the same multimodal LLM baseline achieves state-of-the-art on AudioCaps in both zero-shot and fully supervised captioning without external knowledge, with relative gains up to 52.5% and 7.5%, respectively. These findings show that closing the modality gap is pivotal for effective coupling between multimodal encoders and LLMs, and diffusion-based modality bridging offers a promising direction beyond knowledge-retrieval-centric designs. Code will be released upon acceptance https://github.com/DevKiHyun/Diffusion-Link

cs.SD↗

arXiv2025

SV-Mixer: Replacing the Transformer Encoder with Lightweight MLPs for Self-Supervised Model Compression in Speaker Verification

Self-supervised learning (SSL) has pushed speaker verification accuracy close to state-of-the-art levels, but the Transformer backbones used in most SSL encoders hinder on-device and real-time deployment. Prior compression work trims layer depth or width yet still inherits the quadratic cost of self-attention. We propose SV-Mixer, the first fully MLP-based student encoder for SSL distillation. SV-Mixer replaces Transformer with three lightweight modules: Multi-Scale Mixing for multi-resolution temporal features, Local-Global Mixing for frame-to-utterance context, and Group Channel Mixing for spectral subspaces. Distilled from WavLM, SV-Mixer outperforms a Transformer student by 14.6% while cutting parameters and GMACs by over half, and at 75% compression, it closely matches the teacher's performance. Our results show that attention-free SSL students can deliver teacher-level accuracy with hardware-friendly footprints, opening the door to robust on-device speaker verification.

eess.AS↗

arXiv2025

SEED: Speaker Embedding Enhancement Diffusion Model

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

eess.AS↗

arXiv2024

MR-RawNet: Speaker verification system with multiple temporal resolutions for variable duration utterances using raw waveforms

In speaker verification systems, the utilization of short utterances presents a persistent challenge, leading to performance degradation primarily due to insufficient phonetic information to characterize the speakers. To overcome this obstacle, we propose a novel structure, MR-RawNet, designed to enhance the robustness of speaker verification systems against variable duration utterances using raw waveforms. The MR-RawNet extracts time-frequency representations from raw waveforms via a multi-resolution feature extractor that optimally adjusts both temporal and spectral resolutions simultaneously. Furthermore, we apply a multi-resolution attention block that focuses on diverse and extensive temporal contexts, ensuring robustness against changes in utterance length. The experimental results, conducted on VoxCeleb1 dataset, demonstrate that the MR-RawNet exhibits superior performance in handling utterances of variable duration compared to other raw waveform-based systems.

eess.AS↗

arXiv2023

Diff-SV: A Unified Hierarchical Framework for Noise-Robust Speaker Verification Using Score-Based Diffusion Probabilistic Models

Background noise considerably reduces the accuracy and reliability of speaker verification (SV) systems. These challenges can be addressed using a speech enhancement system as a front-end module. Recently, diffusion probabilistic models (DPMs) have exhibited remarkable noise-compensation capabilities in the speech enhancement domain. Building on this success, we propose Diff-SV, a noise-robust SV framework that leverages DPM. Diff-SV unifies a DPM-based speech enhancement system with a speaker embedding extractor, and yields a discriminative and noise-tolerable speaker representation through a hierarchical structure. The proposed model was evaluated under both in-domain and out-of-domain noisy conditions using the VoxCeleb1 test set, an external noise source, and the VOiCES corpus. The obtained experimental results demonstrate that Diff-SV achieves state-of-the-art performance, outperforming recently proposed noise-robust SV systems.

eess.AS↗

arXiv2023

HM-Conformer: A Conformer-based audio deepfake detection system with hierarchical pooling and multi-level classification token aggregation methods

Audio deepfake detection (ADD) is the task of detecting spoofing attacks generated by text-to-speech or voice conversion systems. Spoofing evidence, which helps to distinguish between spoofed and bona-fide utterances, might exist either locally or globally in the input features. To capture these, the Conformer, which consists of Transformers and CNN, possesses a suitable structure. However, since the Conformer was designed for sequence-to-sequence tasks, its direct application to ADD tasks may be sub-optimal. To tackle this limitation, we propose HM-Conformer by adopting two components: (1) Hierarchical pooling method progressively reducing the sequence length to eliminate duplicated information (2) Multi-level classification token aggregation method utilizing classification tokens to gather information from different blocks. Owing to these components, HM-Conformer can efficiently detect spoofing evidence by processing various sequence lengths and aggregating them. In experimental results on the ASVspoof 2021 Deepfake dataset, HM-Conformer achieved a 15.71% EER, showing competitive performance compared to recent systems.

cs.SD↗

arXiv2023

PAS: Partial Additive Speech Data Augmentation Method for Noise Robust Speaker Verification

Background noise reduces speech intelligibility and quality, making speaker verification (SV) in noisy environments a challenging task. To improve the noise robustness of SV systems, additive noise data augmentation method has been commonly used. In this paper, we propose a new additive noise method, partial additive speech (PAS), which aims to train SV systems to be less affected by noisy environments. The experimental results demonstrate that PAS outperforms traditional additive noise in terms of equal error rates (EER), with relative improvements of 4.64% and 5.01% observed in SE-ResNet34 and ECAPA-TDNN. We also show the effectiveness of proposed method by analyzing attention modules and visualizing speaker embeddings.

eess.AS↗

arXiv2023

One-Step Knowledge Distillation and Fine-Tuning in Using Large Pre-Trained Self-Supervised Learning Models for Speaker Verification

The application of speech self-supervised learning (SSL) models has achieved remarkable performance in speaker verification (SV). However, there is a computational cost hurdle in employing them, which makes development and deployment difficult. Several studies have simply compressed SSL models through knowledge distillation (KD) without considering the target task. Consequently, these methods could not extract SV-tailored features. This paper suggests One-Step Knowledge Distillation and Fine-Tuning (OS-KDFT), which incorporates KD and fine-tuning (FT). We optimize a student model for SV during KD training to avert the distillation of inappropriate information for the SV. OS-KDFT could downsize Wav2Vec 2.0 based ECAPA-TDNN size by approximately 76.2%, and reduce the SSL model's inference time by 79% while presenting an EER of 0.98%. The proposed OS-KDFT is validated across VoxCeleb1 and VoxCeleb2 datasets and W2V2 and HuBERT SSL models. Experiments are available on our GitHub.

eess.AS↗

arXiv2023

Integrated Parameter-Efficient Tuning for General-Purpose Audio Models

The advent of hyper-scale and general-purpose pre-trained models is shifting the paradigm of building task-specific models for target tasks. In the field of audio research, task-agnostic pre-trained models with high transferability and adaptability have achieved state-of-the-art performances through fine-tuning for downstream tasks. Nevertheless, re-training all the parameters of these massive models entails an enormous amount of time and cost, along with a huge carbon footprint. To overcome these limitations, the present study explores and applies efficient transfer learning methods in the audio domain. We also propose an integrated parameter-efficient tuning (IPET) framework by aggregating the embedding prompt (a prompt-based learning approach), and the adapter (an effective transfer learning method). We demonstrate the efficacy of the proposed framework using two backbone pre-trained audio models with different characteristics: the audio spectrogram transformer and wav2vec 2.0. The proposed IPET framework exhibits remarkable performance compared to fine-tuning method with fewer trainable parameters in four downstream tasks: sound event classification, music genre classification, keyword spotting, and speaker verification. Furthermore, the authors identify and analyze the shortcomings of the IPET framework, providing lessons and research directions for parameter efficient tuning in the audio domain.

eess.AS↗

arXiv2022

Convolution channel separation and frequency sub-bands aggregation for music genre classification

In music, short-term features such as pitch and tempo constitute long-term semantic features such as melody and narrative. A music genre classification (MGC) system should be able to analyze these features. In this research, we propose a novel framework that can extract and aggregate both short- and long-term features hierarchically. Our framework is based on ECAPA-TDNN, where all the layers that extract short-term features are affected by the layers that extract long-term features because of the back-propagation training. To prevent the distortion of short-term features, we devised the convolution channel separation technique that separates short-term features from long-term feature extraction paths. To extract more diverse features from our framework, we incorporated the frequency sub-bands aggregation method, which divides the input spectrogram along frequency bandwidths and processes each segment. We evaluated our framework using the Melon Playlist dataset which is a large-scale dataset containing 600 times more data than GTZAN which is a widely used dataset in MGC studies. As the result, our framework achieved 70.4% accuracy, which was improved by 16.9% compared to a conventional framework.

eess.AS↗

arXiv2022

Two Methods for Spoofing-Aware Speaker Verification: Multi-Layer Perceptron Score Fusion Model and Integrated Embedding Projector

The use of deep neural networks (DNN) has dramatically elevated the performance of automatic speaker verification (ASV) over the last decade. However, ASV systems can be easily neutralized by spoofing attacks. Therefore, the Spoofing-Aware Speaker Verification (SASV) challenge is designed and held to promote development of systems that can perform ASV considering spoofing attacks by integrating ASV and spoofing countermeasure (CM) systems. In this paper, we propose two back-end systems: multi-layer perceptron score fusion model (MSFM) and integrated embedding projector (IEP). The MSFM, score fusion back-end system, derived SASV score utilizing ASV and CM scores and embeddings. On the other hand,IEP combines ASV and CM embeddings into SASV embedding and calculates final SASV score based on the cosine similarity. We effectively integrated ASV and CM systems through proposed MSFM and IEP and achieved the SASV equal error rates 0.56%, 1.32% on the official evaluation trials of the SASV 2022 challenge.

eess.AS↗

arXiv2022

RawNeXt: Speaker verification system for variable-duration utterances with deep layer aggregation and extended dynamic scaling policies

Despite achieving satisfactory performance in speaker verification using deep neural networks, variable-duration utterances remain a challenge that threatens the robustness of systems. To deal with this issue, we propose a speaker verification system called RawNeXt that can handle input raw waveforms of arbitrary length by employing the following two components: (1) A deep layer aggregation strategy enhances speaker information by iteratively and hierarchically aggregating features of various time scales and spectral channels output from blocks. (2) An extended dynamic scaling policy flexibly processes features according to the length of the utterance by selectively merging the activations of different resolution branches in each block. Owing to these two components, our proposed model can extract speaker embeddings rich in time-spectral information and operate dynamically on length variations. Experimental results on the VoxCeleb1 test set consisting of various duration utterances demonstrate that RawNeXt achieves state-of-the-art performance compared to the recently proposed systems. Our code and trained model weights are available at https://github.com/wngh1187/RawNeXt.

eess.AS↗

arXiv2022

Extended U-Net for Speaker Verification in Noisy Environments

Background noise is a well-known factor that deteriorates the accuracy and reliability of speaker verification (SV) systems by blurring speech intelligibility. Various studies have used separate pretrained enhancement models as the front-end module of the SV system in noisy environments, and these methods effectively remove noises. However, the denoising process of independent enhancement models not tailored to the SV task can also distort the speaker information included in utterances. We argue that the enhancement network and speaker embedding extractor should be fully jointly trained for SV tasks under noisy conditions to alleviate this issue. Therefore, we proposed a U-Net-based integrated framework that simultaneously optimizes speaker identification and feature enhancement losses. Moreover, we analyzed the structural limitations of using U-Net directly for noise SV tasks and further proposed Extended U-Net to reduce these drawbacks. We evaluated the models on the noise-synthesized VoxCeleb1 test set and VOiCES development set recorded in various noisy scenarios. The experimental results demonstrate that the U-Net-based fully joint training framework is more effective than the baseline, and the extended U-Net exhibited state-of-the-art performance versus the recently proposed compensation systems.

eess.AS↗

arXiv2021

Graph attentive feature aggregation for text-independent speaker verification

The objective of this paper is to combine multiple frame-level features into a single utterance-level representation considering pairwise relationship. For this purpose, we propose a novel graph attentive feature aggregation module by interpreting each frame-level feature as a node of a graph. The inter-relationship between all possible pairs of features, typically exploited indirectly, can be directly modeled using a graph. The module comprises a graph attention layer and a graph pooling layer followed by a readout operation. The graph attention layer first models the non-Euclidean data manifold between different nodes. Then, the graph pooling layer discards less informative nodes considering the significance of the nodes. Finally, the readout operation combines the remaining nodes into a single representation. We employ two recent systems, SE-ResNet and RawNet2, with different input features and architectures and demonstrate that the proposed feature aggregation module consistently shows a relative improvement over 10%, compared to the baseline.

cs.SD↗