arXiv subjects

Ji-Hoon Kim

Publications and source records attributed to Ji-Hoon Kim.

At least 19 recordsLinked to original sources

arXiv2026

ProsoCodec: Prosody-Oriented Speech Codec for Voice Conversion

Neural speech codecs efficiently compress speech and have become a foundation for speech generation, but they are typically learned as holistic representations that intertwine linguistic content, speaker identity, and prosody. While this design is effective for zero-shot voice cloning, it hinders downstream tasks that require prosody preservation or transfer, such as voice conversion. To address this, we introduce ProsoCodec, a prosody-oriented speech codec that models prosody as a conditional residual rather than as a disentangled stream. Specifically, by conditioning both the encoder and decoder on text and speaker embeddings as prefix tokens, the discrete bottleneck is encouraged to capture prosodic variation not explained by content and speaker. To further preserve prosody, we use the low-frequency mel band and train the model on paired same-speaker utterances. Experiments on voice conversion show improved prosody preservation and reduced source-timbre leakage.

eess.AS↗

arXiv2026

Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment

Direct Preference Optimization (DPO) and related methods align large language models from pairwise preferences by regularizing updates against a fixed reference policy. As the policy drifts, a static reference, however, can become increasingly miscalibrated, leading to distributional mismatch and amplifying spurious preference signals under noisy supervision. Conversely, reference-free variants avoid mismatch but often suffer from unconstrained reward drift. We propose Geometric Anchor Preference Optimization (GAPO), which replaces the fixed reference with a dynamic, geometry-aware anchor: an adversarial local perturbation of the current policy within a small radius that serves as a pessimistic baseline. This anchor enables an adaptive reweighting mechanism, modulating the importance of each preference pair based on its local sensitivity. We further introduce the Anchor Gap, the reward discrepancy between the policy and its anchor, and show under smoothness conditions that it approximates worst-case local margin degradation. Optimizing a logistic objective weighted by this gap downweights geometrically brittle instances while emphasizing robust preference signals. Across diverse noise settings, GAPO consistently improves robustness while matching or improving performance on standard LLM alignment and reasoning benchmarks.

cs.LG↗

arXiv2026

Probing Cross-modal Information Hubs in Audio-Visual LLMs

Audio-visual large language models (AVLLMs) have recently emerged as a powerful architecture capable of jointly reasoning over audio, visual, and textual modalities. In AVLLMs, the bidirectional interaction between audio and video modalities introduces intricate processing dynamics, necessitating a deeper understanding of their internal mechanisms. However, unlike extensively studied text-only or large vision language models, the internal workings of AVLLMs remain largely unexplored. In this paper, we focus on cross-modal information flow between audio and visual modalities in AVLLMs, investigating where information derived from one modality is encoded within the token representations of the other modality. Through an analysis of multiple recent AVLLMs, we uncover two common findings. First, AVLLMs primarily encode integrated audio-visual information in sink tokens. Second, sink tokens do not uniformly hold cross-modal information. Instead, a distinct subset of sink tokens, which we term cross-modal sink tokens, specializes in storing such information. Based on these findings, we further propose a simple training-free hallucination mitigation method by encouraging reliance on integrated cross-modal information within cross-modal sink tokens. Our code is available at https://github.com/kaistmm/crossmodal-hub.

cs.AI↗

arXiv2026

Hardware-Efficient Softmax and Layer Normalization with Guaranteed Normalization for Edge Devices

In Transformer models, non-GEMM (non-General Matrix Multiplication) operations -- especially Softmax and Layer Normalization (LayerNorm) -- often dominate hardware cost due to their nonlinear nature. To address this, previous approximation studies mainly target rank-oriented tasks, which is acceptable for classification. However, edge Natural Language Processing (NLP) applications and edge generative AI are largely evaluated based on score-oriented tasks, so normalization-guaranteed non-GEMM operations are essential. We propose a hardware-efficient Softmax and LayerNorm with Guaranteed Normalization for Edge devices. Our design employs hardware-efficient approximation methods while preserving the normalization (Softmax: $\sum p = 1$, LayerNorm: $σ= 1$). Our architecture is described in Verilog HDL and synthesized using the Samsung 28nm CMOS process. In accuracy evaluation, we achieve high accuracy with minimal degradation: GLUE +0.07%, SQuAD -0.01%, perplexity -0.09%. Implementation results show that our architecture is small: $942\,μm^2$ for Softmax, $1199\,μm^2$ for LayerNorm. Compared to the state of the art, we achieve up to 11x and 14x reduction in area, respectively.

cs.AR↗

arXiv2026

MamTra: A Hybrid Mamba-Transformer Backbone for Speech Synthesis

Despite the remarkable quality of LLM-based text-to-speech systems, their reliance on autoregressive Transformers leads to quadratic computational complexity, which severely limits practical applications. Linear-time alternatives, notably Mamba, offer a potential remedy; however, they often sacrifice the global context essential for expressive synthesis. In this paper, we propose MamTra, an interleaved Mamba-Transformer framework designed to leverage the advantages of Mamba's efficiency and Transformers' modeling capability. We also introduce novel knowledge transfer strategies to distill insights from a pretrained Transformer into our hybrid architecture, thereby bypassing the prohibitive costs of training from scratch. Systematic experiments identify the optimal hybrid configuration, and demonstrate that MamTra reduces inference VRAM usage by up to 34% without compromising speech fidelity - even trained on only 2% of the original training dataset. Audio samples are available at https://mamtratts.github.io.

eess.AS↗

arXiv2026

SPADE: Structured Pruning and Adaptive Distillation for Efficient LLM-TTS

The goal of this paper is to introduce SPADE, a framework for Structured Pruning and Adaptive Distillation for Efficient Large Language Model-based text-to-speech (LLM-TTS). Recent LLM-TTS systems achieve strong controllability and zero-shot generalization, but their large parameter counts and high latency limit real-world deployment. SPADE addresses this by combining (i) a pruning step guided by a word-error-rate-based layer importance index to remove non-essential Transformer layers, with (ii) multi-level knowledge distillation to restore autoregressive coherence. On zero-shot benchmarks, SPADE preserves near-parity perceptual quality while halving Transformer depth, reducing VRAM usage by up to 20%, and achieving up to 1.7x faster real-time factor with less than 5% of the original training data. These results show that compact LLM-TTS models can maintain naturalness and speaker similarity while enabling practical real-time speech generation. Audio samples are available at https://mm.kaist.ac.kr/projects/SPADE/.

eess.AS↗

arXiv2026

UNMIXX: Untangling Highly Correlated Singing Voices Mixtures

We introduce UNMIXX, a novel framework for multiple singing voices separation (MSVS). While related to speech separation, MSVS faces unique challenges: data scarcity and the highly correlated nature of singing voices mixture. To address these issues, we propose UNMIXX with three key components: (1) musically informed mixing strategy to construct highly correlated, music-like mixtures, (2) cross-source attention that drives representations of two singers apart via reverse attention, and (3) magnitude penalty loss penalizing erroneously assigned interfering energy. UNMIXX not only addresses data scarcity by simulating realistic training data, but also excels at separating highly correlated mixtures through cross-source interactions at both the architectural and loss levels. Our extensive experiments demonstrate that UNMIXX greatly enhances performance, with SDRi gains exceeding 2.2 dB over prior work.

cs.SD↗

arXiv2025

TAVID: Text-Driven Audio-Visual Interactive Dialogue Generation

The objective of this paper is to jointly synthesize interactive videos and conversational speech from text and reference images. With the ultimate goal of building human-like conversational systems, recent studies have explored talking or listening head generation as well as conversational speech generation. However, these works are typically studied in isolation, overlooking the multimodal nature of human conversation, which involves tightly coupled audio-visual interactions. In this paper, we introduce TAVID, a unified framework that generates both interactive faces and conversational speech in a synchronized manner. TAVID integrates face and speech generation pipelines through two cross-modal mappers (i.e., a motion mapper and a speaker mapper), which enable bidirectional exchange of complementary information between the audio and visual modalities. We evaluate our system across four dimensions: talking face realism, listening head responsiveness, dyadic interaction fluency, and speech quality. Extensive experiments demonstrate the effectiveness of our approach across all these aspects.

cs.CV↗

arXiv2025

AlignDiT: Multimodal Aligned Diffusion Transformer for Synchronized Speech Generation

In this paper, we address the task of multimodal-to-speech generation, which aims to synthesize high-quality speech from multiple input modalities: text, video, and reference audio. This task has gained increasing attention due to its wide range of applications, such as film production, dubbing, and virtual avatars. Despite recent progress, existing methods still suffer from limitations in speech intelligibility, audio-video synchronization, speech naturalness, and voice similarity to the reference speaker. To address these challenges, we propose AlignDiT, a multimodal Aligned Diffusion Transformer that generates accurate, synchronized, and natural-sounding speech from aligned multimodal inputs. Built upon the in-context learning capability of the DiT architecture, AlignDiT explores three effective strategies to align multimodal representations. Furthermore, we introduce a novel multimodal classifier-free guidance mechanism that allows the model to adaptively balance information from each modality during speech synthesis. Extensive experiments demonstrate that AlignDiT significantly outperforms existing methods across multiple benchmarks in terms of quality, synchronization, and speaker similarity. Moreover, AlignDiT exhibits strong generalization capability across various multimodal tasks, such as video-to-speech synthesis and visual forced alignment, consistently achieving state-of-the-art performance. The demo page is available at https://mm.kaist.ac.kr/projects/AlignDiT.

cs.CV↗

arXiv2025

InfiniteAudio: Infinite-Length Audio Generation with Consistency

This paper presents InfiniteAudio, a simple yet effective strategy for generating infinite-length audio using diffusion-based text-to-audio methods. Current approaches face memory constraints because the output size increases with input length, making long duration generation challenging. A common workaround is to concatenate short audio segments, but this often leads to inconsistencies due to the lack of shared temporal context. To address this, InfiniteAudio integrates seamlessly into existing pipelines without additional training. It introduces two key techniques: FIFO sampling, a first-in, first-out inference strategy with fixed-size inputs, and curved denoising, which selectively prioritizes key diffusion steps for efficiency. Experiments show that InfiniteAudio achieves comparable or superior performance across all metrics. Audio samples are available on our project page.

eess.AS↗

arXiv2025

Text-To-Speech Synthesis In The Wild

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.

eess.AS↗

arXiv2025

V2SFlow: Video-to-Speech Generation with Speech Decomposition and Rectified Flow

In this paper, we introduce V2SFlow, a novel Video-to-Speech (V2S) framework designed to generate natural and intelligible speech directly from silent talking face videos. While recent V2S systems have shown promising results on constrained datasets with limited speakers and vocabularies, their performance often degrades on real-world, unconstrained datasets due to the inherent variability and complexity of speech signals. To address these challenges, we decompose the speech signal into manageable subspaces (content, pitch, and speaker information), each representing distinct speech attributes, and predict them directly from the visual input. To generate coherent and realistic speech from these predicted attributes, we employ a rectified flow matching decoder built on a Transformer architecture, which models efficient probabilistic pathways from random noise to the target speech distribution. Extensive experiments demonstrate that V2SFlow significantly outperforms state-of-the-art methods, even surpassing the naturalness of ground truth utterances. Code and models are available at: https://github.com/kaistmm/V2SFlow

cs.CV↗

arXiv2025

Accelerating Diffusion-based Text-to-Speech Model Training with Dual Modality Alignment

The goal of this paper is to optimize the training process of diffusion-based text-to-speech models. While recent studies have achieved remarkable advancements, their training demands substantial time and computational costs, largely due to the implicit guidance of diffusion models in learning complex intermediate representations. To address this, we propose A-DMA, an effective strategy for Accelerating training with Dual Modality Alignment. Our method introduces a novel alignment pipeline leveraging both text and speech modalities: text-guided alignment, which incorporates contextual representations, and speech-guided alignment, which refines semantic representations. By aligning hidden states with discriminative features, our training scheme reduces the reliance on diffusion models for learning complex representations. Extensive experiments demonstrate that A-DMA doubles the convergence speed while achieving superior performance over baselines. Code and demo samples are available at: https://github.com/ZhikangNiu/A-DMA

eess.AS↗

arXiv2025

SpoofCeleb: Speech Deepfake Detection and SASV In The Wild

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.

cs.SD↗

arXiv2025

SCRec: A Scalable Computational Storage System with Statistical Sharding and Tensor-train Decomposition for Recommendation Models

Deep Learning Recommendation Models (DLRMs) play a crucial role in delivering personalized content across web applications such as social networking and video streaming. However, with improvements in performance, the parameter size of DLRMs has grown to terabyte (TB) scales, accompanied by memory bandwidth demands exceeding TB/s levels. Furthermore, the workload intensity within the model varies based on the target mechanism, making it difficult to build an optimized recommendation system. In this paper, we propose SCRec, a scalable computational storage recommendation system that can handle TB-scale industrial DLRMs while guaranteeing high bandwidth requirements. SCRec utilizes a software framework that features a mixed-integer programming (MIP)-based cost model, efficiently fetching data based on data access patterns and adaptively configuring memory-centric and compute-centric cores. Additionally, SCRec integrates hardware acceleration cores to enhance DLRM computations, particularly allowing for the high-performance reconstruction of approximated embedding vectors from extremely compressed tensor-train (TT) format. By combining its software framework and hardware accelerators, while eliminating data communication overhead by being implemented on a single server, SCRec achieves substantial improvements in DLRM inference performance. It delivers up to 55.77$\times$ speedup compared to a CPU-DRAM system with no loss in accuracy and up to 13.35$\times$ energy efficiency gains over a multi-GPU system.

cs.AR↗

arXiv2025

From Faces to Voices: Learning Hierarchical Representations for High-quality Video-to-Speech

The objective of this study is to generate high-quality speech from silent talking face videos, a task also known as video-to-speech synthesis. A significant challenge in video-to-speech synthesis lies in the substantial modality gap between silent video and multi-faceted speech. In this paper, we propose a novel video-to-speech system that effectively bridges this modality gap, significantly enhancing the quality of synthesized speech. This is achieved by learning of hierarchical representations from video to speech. Specifically, we gradually transform silent video into acoustic feature spaces through three sequential stages -- content, timbre, and prosody modeling. In each stage, we align visual factors -- lip movements, face identity, and facial expressions -- with corresponding acoustic counterparts to ensure the seamless transformation. Additionally, to generate realistic and coherent speech from the visual representations, we employ a flow matching model that estimates direct trajectories from a simple prior distribution to the target speech distribution. Extensive experiments demonstrate that our method achieves exceptional generation quality comparable to real utterances, outperforming existing methods by a significant margin.

eess.AS↗

arXiv2025

Optimized Memory System Architecture for VESA VDC-M Decoder with Multi-Slice Support

Video compression plays a pivotal role in managing and transmitting large-scale display data, particularly given the growing demand for higher resolutions and improved video quality. This paper proposes an optimized memory system architecture for Video Electronics Standards Association (VESA) Display Compression-M (VDC-M) decoder, characterized by its substantial on-chip buffer requirements. We design and analyze three architectures categorized by optimization levels and management complexity. Our strategy focuses on enhancing line buffer access scheduling and minimizing reconstruction buffer, targeting prediction and multi-slice operation that are the major resource consumers in the decoder. By adjusting line delay and segmenting SRAM bank alongside reconstructed block forwarding, we achieve a 33.3% size reduction in the line buffer and 77.3% in the reconstruction buffer compared to Baseline VDC-M decoder. Synthesized using a 28 nm CMOS process, the proposed architecture achieves a 31.5% reduction in gate count of the decoder backend hardware, supporting real-time performance with up to 96.45 fps for 4K UHD resolution at 200 MHz operating frequency and a throughput of 4 pixels per cycle.

cs.AR↗

arXiv2025

AdaptVC: High Quality Voice Conversion with Adaptive Learning

The goal of voice conversion is to transform the speech of a source speaker to sound like that of a reference speaker while preserving the original content. A key challenge is to extract disentangled linguistic content from the source and voice style from the reference. While existing approaches leverage various methods to isolate the two, a generalization still requires further attention, especially for robustness in zero-shot scenarios. In this paper, we achieve successful disentanglement of content and speaker features by tuning self-supervised speech features with adapters. The adapters are trained to dynamically encode nuanced features from rich self-supervised features, and the decoder fuses them to produce speech that accurately resembles the reference with minimal loss of content. Moreover, we leverage a conditional flow matching decoder with cross-attention speaker conditioning to further boost the synthesis quality and efficiency. Subjective and objective evaluations in a zero-shot scenario demonstrate that the proposed method outperforms existing models in speech quality and similarity to the reference speech.

cs.SD↗