Juhan Nam

Seungheon Doh, Keunwoo Choi, Juhan Nam

While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited, leaving other simpler yet crucial components such as metadata or attribute filtering underutilized in the system. We propose an LLM-based music recommendation system with tool calling to serve as a unified retrieval-reranking pipeline. Our system positions an LLM as an end-to-end recommendation system that interprets user intent, plans tool invocations, and orchestrates specialized components: boolean filters (SQL), sparse retrieval (BM25), dense retrieval (embedding similarity), and generative retrieval (semantic IDs). Through tool planning, the system predicts which types of tools to use, their execution order, and the arguments needed to find music matching user preferences, supporting diverse modalities while seamlessly integrating multiple database filtering methods. We demonstrate that this unified tool-calling framework achieves competitive performance across diverse recommendation scenarios by selectively employing appropriate retrieval methods based on user queries, envisioning a new paradigm for conversational music recommendation systems.

Seungheon Doh, Keunwoo Choi, Juhan Nam

We present TALKPLAY, a novel multimodal music recommendation system that reformulates recommendation as a token generation problem using large language models (LLMs). By leveraging the instruction-following and natural language generation capabilities of LLMs, our system effectively recommends music from diverse user queries while generating contextually relevant responses. While pretrained LLMs are primarily designed for text modality, TALKPLAY extends their scope through two key innovations: a multimodal music tokenizer that encodes audio features, lyrics, metadata, semantic tags, and playlist co-occurrence signals; and a vocabulary expansion mechanism that enables unified processing and generation of both linguistic and music-relevant tokens. By integrating the recommendation system directly into the LLM architecture, TALKPLAY transforms conventional systems by: (1) unifying previous two-stage conversational recommendation systems (recommendation engines and dialogue managers) into a cohesive end-to-end system, (2) effectively utilizing long conversational context for recommendation while maintaining strong performance in extended multi-turn interactions, and (3) generating natural language responses for seamless user interaction. Our qualitative and quantitative evaluation demonstrates that TALKPLAY significantly outperforms unimodal approaches based solely on text or listening history in both recommendation performance and conversational naturalness.

Ilaria Manco, Benno Weck, SeungHeon Doh et al. · bytedance

We introduce the Song Describer dataset (SDD), a new crowdsourced corpus of high-quality audio-caption pairs, designed for the evaluation of music-and-language models. The dataset consists of 1.1k human-written natural language descriptions of 706 music recordings, all publicly accessible and released under Creative Common licenses. To showcase the use of our dataset, we benchmark popular models on three key music-and-language tasks (music captioning, text-to-music generation and music-language retrieval). Our experiments highlight the importance of cross-dataset evaluation and offer insights into how researchers can use SDD to gain a broader understanding of model performance.

Haven Kim, Kento Watanabe, Masataka Goto et al.

Lyric translation plays a pivotal role in amplifying the global resonance of music, bridging cultural divides, and fostering universal connections. Translating lyrics, unlike conventional translation tasks, requires a delicate balance between singability and semantics. In this paper, we present a computational framework for the quantitative evaluation of singable lyric translation, which seamlessly integrates musical, linguistic, and cultural dimensions of lyrics. Our comprehensive framework consists of four metrics that measure syllable count distance, phoneme repetition similarity, musical structure distance, and semantic similarity. To substantiate the efficacy of our framework, we collected a singable lyrics dataset, which precisely aligns English, Japanese, and Korean lyrics on a line-by-line and section-by-section basis, and conducted a comparative analysis between singable and non-singable lyrics. Our multidisciplinary approach provides insights into the key components that underlie the art of lyric translation and establishes a solid groundwork for the future of computational lyric translation assessment.

Yeonghyeon Lee, Inmo Yeon, Juhan Nam et al.

This paper presents VoiceLDM, a model designed to produce audio that accurately follows two distinct natural language text prompts: the description prompt and the content prompt. The former provides information about the overall environmental context of the audio, while the latter conveys the linguistic content. To achieve this, we adopt a text-to-audio (TTA) model based on latent diffusion models and extend its functionality to incorporate an additional content prompt as a conditional input. By utilizing pretrained contrastive language-audio pretraining (CLAP) and Whisper, VoiceLDM is trained on large amounts of real-world audio without manual annotations or transcriptions. Additionally, we employ dual classifier-free guidance to further enhance the controllability of VoiceLDM. Experimental results demonstrate that VoiceLDM is capable of generating plausible audio that aligns well with both input conditions, even surpassing the speech intelligibility of the ground truth audio on the AudioCaps test set. Furthermore, we explore the text-to-speech (TTS) and zero-shot text-to-audio capabilities of VoiceLDM and show that it achieves competitive results. Demos and code are available at https://voiceldm.github.io.

Sangeon Yong, Li Su, Juhan Nam

Note-level automatic music transcription is one of the most representative music information retrieval (MIR) tasks and has been studied for various instruments to understand music. However, due to the lack of high-quality labeled data, transcription of many instruments is still a challenging task. In particular, in the case of singing, it is difficult to find accurate notes due to its expressiveness in pitch, timbre, and dynamics. In this paper, we propose a method of finding note onsets of singing voice more accurately by leveraging the linguistic characteristics of singing, which are not seen in other instruments. The proposed model uses mel-scaled spectrogram and phonetic posteriorgram (PPG), a frame-wise likelihood of phoneme, as an input of the onset detection network while PPG is generated by the pre-trained network with singing and speech data. To verify how linguistic features affect onset detection, we compare the evaluation results through the dataset with different languages and divide onset types for detailed analysis. Our approach substantially improves the performance of singing transcription and therefore emphasizes the importance of linguistic features in singing analysis.

Eunjin Choi, Yoonjin Chung, Seolhee Lee et al.

Existing multi-instrumental datasets tend to be biased toward pop and classical music. In addition, they generally lack high-level annotations such as emotion tags. In this paper, we propose YM2413-MDB, an 80s FM video game music dataset with multi-label emotion annotations. It includes 669 audio and MIDI files of music from Sega and MSX PC games in the 80s using YM2413, a programmable sound generator based on FM. The collected game music is arranged with a subset of 15 monophonic instruments and one drum instrument. They were converted from binary commands of the YM2413 sound chip. Each song was labeled with 19 emotion tags by two annotators and validated by three verifiers to obtain refined tags. We provide the baseline models and results for emotion recognition and emotion-conditioned symbolic music generation using YM2413-MDB.

Geoffroy Peeters, Zafar Rafii, Magdalena Fuentes et al.

In this paper, we trace the evolution of Music Information Retrieval (MIR) over the past 25 years. While MIR gathers all kinds of research related to music informatics, a large part of it focuses on signal processing techniques for music data, fostering a close relationship with the IEEE Audio and Acoustic Signal Processing Technical Commitee. In this paper, we reflect the main research achievements of MIR along the three EDICS related to music analysis, processing and generation. We then review a set of successful practices that fuel the rapid development of MIR research. One practice is the annual research benchmark, the Music Information Retrieval Evaluation eXchange, where participants compete on a set of research tasks. Another practice is the pursuit of reproducible and open research. The active engagement with industry research and products is another key factor for achieving large societal impacts and motivating younger generations of students to join the field. Last but not the least, the commitment to diversity, equity and inclusion ensures MIR to be a vibrant and open community where various ideas, methodologies, and career pathways collide. We finish by providing future challenges MIR will have to face.

Haven Kim, SeungHeon Doh, Junwon Lee et al.

Automatically generating or captioning music playlist titles given a set of tracks is of significant interest in music streaming services as customized playlists are widely used in personalized music recommendation, and well-composed text titles attract users and help their music discovery. We present an encoder-decoder model that generates a playlist title from a sequence of music tracks. While previous work takes track IDs as tokenized input for playlist title generation, we use artist IDs corresponding to the tracks to mitigate the issue from the long-tail distribution of tracks included in the playlist dataset. Also, we introduce a chronological data split method to deal with newly-released tracks in real-world scenarios. Comparing the track IDs and artist IDs as input sequences, we show that the artist-based approach significantly enhances the performance in terms of word overlap, semantic relevance, and diversity.

Yuya Yamamoto, Juhan Nam, Hiroko Terasawa

Professional vocalists modulate their voice timbre or pitch to make their vocal performance more expressive. Such fluctuations are called singing techniques. Automatic detection of singing techniques from audio tracks can be beneficial to understand how each singer expresses the performance, yet it can also be difficult due to the wide variety of the singing techniques. A deep neural network (DNN) model can handle such variety; however, there might be a possibility that considering the characteristics of the data improves the performance of singing technique detection. In this paper, we propose PrimaDNN, a CRNN model with a characteristics-oriented improvement. The features of the model are: 1) input feature representation based on auxiliary pitch information and multi-resolution mel spectrograms, 2) Convolution module based on the Squeeze-and-excitation (SENet) and the Instance normalization. In the results of J-POP singing technique detection, PrimaDNN achieved the best results of 44.9% at the overall macro-F measure, compared to conventional works. We also found that the contribution of each component varies depending on the type of singing technique.

Haven Kim, Jongmin Jung, Dasaem Jeong et al.

Lyric translation, a field studied for over a century, is now attracting computational linguistics researchers. We identified two limitations in previous studies. Firstly, lyric translation studies have predominantly focused on Western genres and languages, with no previous study centering on K-pop despite its popularity. Second, the field of lyric translation suffers from a lack of publicly available datasets; to the best of our knowledge, no such dataset exists. To broaden the scope of genres and languages in lyric translation studies, we introduce a novel singable lyric translation dataset, approximately 89\% of which consists of K-pop song lyrics. This dataset aligns Korean and English lyrics line-by-line and section-by-section. We leveraged this dataset to unveil unique characteristics of K-pop lyric translation, distinguishing it from other extensively studied genres, and to construct a neural lyric translation model, thereby underscoring the importance of a dedicated dataset for singable lyric translations.

Gyubin Lee, Junwon Lee, Juhan Nam

We investigate Counterfactual Video Foley Generation, which aims to adopt a sound-source identity that contradicts the visual evidence while remaining temporally synchronized to a silent video. Existing Video&Text-to-Audio (VT2A) models struggle with this, often remaining anchored to the visually implied sound source when video and text contents disagree. We present ConterFlow, an inference-time dual-phase sampling scheme for pretrained flow-matching VT2A models. Phase 1 builds a video-derived temporal structure while suppressing the visually implied source; Phase 2 drops video conditioning to focus entirely on shaping audio timbre toward the target prompt. ConterFlow substantially improves counterfactual Video Foley generation compared to naive negative prompting and state-of-the-art baselines. To evaluate replacement quality, we propose a metric leveraging a text-audio co-embedding space to measure both target-prompt evidence and residual visually implied source leakage. Video demonstrations and code are available at https://gyubin-lee.github.io/counterflow-demo/

Junwon Lee, Jaekwon Im, Dabin Kim et al.

Foley sound synthesis is crucial for multimedia production, enhancing user experience by synchronizing audio and video both temporally and semantically. Recent studies on automating this labor-intensive process through video-to-sound generation face significant challenges. Systems lacking explicit temporal features suffer from poor alignment and controllability, while timestamp-based models require costly and subjective human annotation. We propose Video-Foley, a video-to-sound system using Root Mean Square (RMS) as an intuitive condition with semantic timbre prompts (audio or text). RMS, a frame-level intensity envelope closely related to audio semantics, acts as a temporal event feature to guide audio generation from video. The annotation-free self-supervised learning framework consists of two stages, Video2RMS and RMS2Sound, incorporating novel ideas including RMS discretization and RMS-ControlNet with a pretrained text-to-audio model. Our extensive evaluation shows that Video-Foley achieves state-of-the-art performance in audio-visual alignment and controllability for sound timing, intensity, timbre, and nuance. Source code, model weights and demos are available on our companion website. (https://jnwnlee.github.io/video-foley-demo)

Yoonjin Chung, Pilsun Eu, Junwon Lee et al.

Although being widely adopted for evaluating generated audio signals, the Fréchet Audio Distance (FAD) suffers from significant limitations, including reliance on Gaussian assumptions, sensitivity to sample size, and high computational complexity. As an alternative, we introduce the Kernel Audio Distance (KAD), a novel, distribution-free, unbiased, and computationally efficient metric based on Maximum Mean Discrepancy (MMD). Through analysis and empirical validation, we demonstrate KAD's advantages: (1) faster convergence with smaller sample sizes, enabling reliable evaluation with limited data; (2) lower computational cost, with scalable GPU acceleration; and (3) stronger alignment with human perceptual judgments. By leveraging advanced embeddings and characteristic kernels, KAD captures nuanced differences between real and generated audio. Open-sourced in the kadtk toolkit, KAD provides an efficient, reliable, and perceptually aligned benchmark for evaluating generative audio models.

Joonhyung Bae, Kirak Kim, Hyeyoon Cho et al.

Synthesizing realistic piano hand motions requires both precision and naturalness. Physics-based methods achieve precision but produce stiff motions; data-driven models learn natural dynamics but struggle with positional accuracy. Piano motion exhibits a natural hierarchy: fingertip positions are nearly deterministic given piano geometry and fingering, while wrist and intermediate joints offer stylistic freedom. We present [OURS], a four-stage framework exploiting this hierarchy: (1) statistics-based fingertip positioning, (2) FiLM-conditioned trajectory refinement, (3) wrist estimation, and (4) STGCN-based pose synthesis. We contribute expert-annotated fingerings for the FürElise dataset (153 pieces, ~10 hours). Experiments demonstrate F1 = 0.910, substantially outperforming diffusion baselines (F1 = 0.121), with user study (N=41) confirming quality approaching motion capture. Expert evaluation by professional pianists (N=5) identified anticipatory motion as the key remaining gap, providing concrete directions for future improvement.

Joonhyung Bae, Kirak Kim, Hyeyoon Cho et al.

Piano fingering shapes how a passage can be played, yet it is difficult to label after a performance. An annotator must decide which finger produced each note while reconciling the score, timing, video, and hand motion. We present PiAnnotate, a web-based pipeline for adding expert fingering annotations to the FurElise performance dataset. The tool brings together a piano-roll view, performance video, and a 3D MANO hand mesh so that reviewers can inspect each assignment in musical and physical context. Rather than storing only the final answer, PiAnnotate keeps paired rule-based and human-edited fingering tracks. These paired tracks make the annotation history auditable by showing where a geometric rule was sufficient, where experts intervened, and how labels changed across review passes. As a final diagnostic, we train a small Transformer probe on the paired tracks. The probe improves on the rule baseline on held-out pieces while remaining conservative about changing labels that were already correct, suggesting that the edited labels contain learnable structure rather than only isolated fixes.

Eunjin Choi, Hounsu Kim, Hayeon Bang et al.

Generating piano accompaniments in the symbolic music domain is a challenging task that requires producing a complete piece of piano music from given melody and chord constraints, such as those provided by a lead sheet. In this paper, we propose a discrete diffusion-based piano accompaniment generation model, D3PIA, leveraging local alignment between lead sheet and accompaniment in piano-roll representation. D3PIA incorporates Neighborhood Attention (NA) to both encode the lead sheet and condition it for predicting note states in the piano accompaniment. This design enhances local contextual modeling by efficiently attending to nearby melody and chord conditions. We evaluate our model using the POP909 dataset, a widely used benchmark for piano accompaniment generation. Objective evaluation results demonstrate that D3PIA preserves chord conditions more faithfully compared to continuous diffusion-based and Transformer-based baselines. Furthermore, a subjective listening test indicates that D3PIA generates more musically coherent accompaniments than the comparison models.

Hounsu Kim, Taegyun Kwon, Juhan Nam

Diffusion models have been widely used in the generative domain due to their convincing performance in modeling complex data distributions. Moreover, they have shown competitive results on discriminative tasks, such as image segmentation. While diffusion models have also been explored for automatic music transcription, their performance has yet to reach a competitive level. In this paper, we focus on discrete diffusion model's refinement capabilities and present a novel architecture for piano transcription. Our model utilizes Neighborhood Attention layers as the denoising module, gradually predicting the target high-resolution piano roll, conditioned on the finetuned features of a pretrained acoustic model. To further enhance refinement, we devise a novel strategy which applies distinct transition states during training and inference stage of discrete diffusion models. Experiments on the MAESTRO dataset show that our approach outperforms previous diffusion-based piano transcription models and the baseline model in terms of F1 score. Our code is available in https://github.com/hanshounsu/d3rm.

Daeyong Kwon, SeungHeon Doh, Juhan Nam

Recent advances in large language models (LLMs) have transformed open-domain question answering, yet their effectiveness in music-related reasoning remains limited due to sparse music knowledge in pretraining data. While music information retrieval and computational musicology have explored structured and multimodal understanding, few resources support factual and contextual music question answering (MQA) grounded in artist metadata or historical context. We introduce MusWikiDB, a vector database of 3.2M passages from 144K music-related Wikipedia pages, and ArtistMus, a benchmark of 1,000 questions on 500 diverse artists with metadata such as genre, debut year, and topic. These resources enable systematic evaluation of retrieval-augmented generation (RAG) for MQA. Experiments show that RAG markedly improves factual accuracy; open-source models gain up to +56.8 percentage points (for example, Qwen3 8B improves from 35.0 to 91.8), approaching proprietary model performance. RAG-style fine-tuning further boosts both factual recall and contextual reasoning, improving results on both in-domain and out-of-domain benchmarks. MusWikiDB also yields approximately 6 percentage points higher accuracy and 40% faster retrieval than a general-purpose Wikipedia corpus. We release MusWikiDB and ArtistMus to advance research in music information retrieval and domain-specific question answering, establishing a foundation for retrieval-augmented reasoning in culturally rich domains such as music.

Junwon Lee, Juhan Nam, Jiyoung Lee

This work introduces a new task, text-conditioned selective video-to-audio (V2A) generation, which produces only the user-intended sound from a multi-object video. This capability is especially crucial in multimedia production, where audio tracks are handled individually for each sound source for precise editing, mixing, and creative control. However, current approaches generate single source-mixed sounds at once, largely because visual features are entangled, and region cues or prompts often fail to specify the source. We propose SelVA, a novel text-conditioned V2A model that treats the text prompt as an explicit selector of target source and modulates video encoder to distinctly extract prompt-relevant video features. The proposed supplementary tokens promote cross-attention by suppressing text-irrelevant activations with efficient parameter tuning, yielding robust semantic and temporal grounding. SelVA further employs a self-augmentation scheme to overcome the lack of mono audio track supervision. We evaluate SelVA on VGG-MONOAUDIO, a curated benchmark of clean single-source videos for such a task. Extensive experiments and ablations consistently verify its effectiveness across audio quality, semantic alignment, and temporal synchronization. Code and demo are available at https://jnwnlee.github.io/selva-demo/.

Gyehun Go, Satbyul Han, Ahyeon Choi et al.

Recent advances in text-to-music (TTM) generation have enabled controllable and expressive music creation using natural language prompts. However, the emotional fidelity of TTM systems remains largely underexplored compared to human preference or text alignment. In this study, we introduce AImoclips, a benchmark for evaluating how well TTM systems convey intended emotions to human listeners, covering both open-source and commercial models. We selected 12 emotion intents spanning four quadrants of the valence-arousal space, and used six state-of-the-art TTM systems to generate over 1,000 music clips. A total of 111 participants rated the perceived valence and arousal of each clip on a 9-point Likert scale. Our results show that commercial systems tend to produce music perceived as more pleasant than intended, while open-source systems tend to perform the opposite. Emotions are more accurately conveyed under high-arousal conditions across all models. Additionally, all systems exhibit a bias toward emotional neutrality, highlighting a key limitation in affective controllability. This benchmark offers valuable insights into model-specific emotion rendering characteristics and supports future development of emotionally aligned TTM systems.

Yoonjin Chung, Junwon Lee, Juhan Nam

Foley sound, audio content inserted synchronously with videos, plays a critical role in the user experience of multimedia content. Recently, there has been active research in Foley sound synthesis, leveraging the advancements in deep generative models. However, such works mainly focus on replicating a single sound class or a textual sound description, neglecting temporal information, which is crucial in the practical applications of Foley sound. We present T-Foley, a Temporal-event-guided waveform generation model for Foley sound synthesis. T-Foley generates high-quality audio using two conditions: the sound class and temporal event feature. For temporal conditioning, we devise a temporal event feature and a novel conditioning technique named Block-FiLM. T-Foley achieves superior performance in both objective and subjective evaluation metrics and generates Foley sound well-synchronized with the temporal events. Additionally, we showcase T-Foley's practical applications, particularly in scenarios involving vocal mimicry for temporal event control. We show the demo on our companion website.

Taegyun Kwon, Dasaem Jeong, Juhan Nam

In recent years, advancements in neural network designs and the availability of large-scale labeled datasets have led to significant improvements in the accuracy of piano transcription models. However, most previous work focused on high-performance offline transcription, neglecting deliberate consideration of model size. The goal of this work is to implement real-time inference for piano transcription while ensuring both high performance and lightweight. To this end, we propose novel architectures for convolutional recurrent neural networks, redesigning an existing autoregressive piano transcription model. First, we extend the acoustic module by adding a frequency-conditioned FiLM layer to the CNN module to adapt the convolutional filters on the frequency axis. Second, we improve note-state sequence modeling by using a pitchwise LSTM that focuses on note-state transitions within a note. In addition, we augment the autoregressive connection with an enhanced recursive context. Using these components, we propose two types of models; one for high performance and the other for high compactness. Through extensive experiments, we show that the proposed models are comparable to state-of-the-art models in terms of note accuracy on the MAESTRO dataset. We also investigate the effective model size and real-time inference latency by gradually streamlining the architecture. Finally, we conduct cross-data evaluation on unseen piano datasets and in-depth analysis to elucidate the effect of the proposed components in the view of note length and pitch range.

Jiyun Park, Sangeon Yong, Taegyun Kwon et al.

The goal of real-time lyrics alignment is to take live singing audio as input and to pinpoint the exact position within given lyrics on the fly. The task can benefit real-world applications such as the automatic subtitling of live concerts or operas. However, designing a real-time model poses a great challenge due to the constraints of only using past input and operating within a minimal latency. Furthermore, due to the lack of datasets for real-time models for lyrics alignment, previous studies have mostly evaluated with private in-house datasets, resulting in a lack of standard evaluation methods. This paper presents a real-time lyrics alignment system for classical vocal performances with two contributions. First, we improve the lyrics alignment algorithm by finding an optimal combination of chromagram and phonetic posteriorgram (PPG) that capture melodic and phonetics features of the singing voice, respectively. Second, we recast the Schubert Winterreise Dataset (SWD) which contains multiple performance renditions of the same pieces as an evaluation set for the real-time lyrics alignment.

Keunwoo Choi, Seungheon Doh, Juhan Nam

We present TalkPlayData 2, a synthetic dataset for multimodal conversational music recommendation generated by an agentic data pipeline. In the proposed pipeline, multiple large language model (LLM) agents are created under various roles with specialized prompts and access to different parts of information, and the chat data is acquired by logging the conversation between the Listener LLM and the Recsys LLM. To cover various conversation scenarios, for each conversation, the Listener LLM is conditioned on a finetuned conversation goal. Finally, all the LLMs are multimodal with audio and images, allowing a simulation of multimodal recommendation and conversation. In the LLM-as-a-judge and subjective evaluation experiments, TalkPlayData 2 achieved the proposed goal in various aspects related to training a generative recommendation model for music. TalkPlayData 2 and its generation code are released at https://talkpl.ai/talkplaydata2.

Hayeon Bang, Eunjin Choi, Megan Finch et al.

While piano music has become a significant area of study in Music Information Retrieval (MIR), there is a notable lack of datasets for piano solo music with text labels. To address this gap, we present PIAST (PIano dataset with Audio, Symbolic, and Text), a piano music dataset. Utilizing a piano-specific taxonomy of semantic tags, we collected 9,673 tracks from YouTube and added human annotations for 2,023 tracks by music experts, resulting in two subsets: PIAST-YT and PIAST-AT. Both include audio, text, tag annotations, and transcribed MIDI utilizing state-of-the-art piano transcription and beat tracking models. Among many possible tasks with the multi-modal dataset, we conduct music tagging and retrieval using both audio and MIDI data and report baseline performances to demonstrate its potential as a valuable resource for MIR research.

Yonghyun Kim, Junhyung Park, Joonhyung Bae et al.

The multimodal nature of music performance has driven increasing interest in data beyond the audio domain within the music information retrieval (MIR) community. This paper introduces PianoVAM, a comprehensive piano performance dataset that includes videos, audio, MIDI, hand landmarks, fingering labels, and rich metadata. The dataset was recorded using a Disklavier piano, capturing audio and MIDI from amateur pianists during their daily practice sessions, alongside synchronized top-view videos in realistic and varied performance conditions. Hand landmarks and fingering labels were extracted using a pretrained hand pose estimation model and a semi-automated fingering annotation algorithm. We discuss the challenges encountered during data collection and the alignment process across different modalities. Additionally, we describe our fingering annotation method based on hand landmarks extracted from videos. Finally, we present benchmarking results for both audio-only and audio-visual piano transcription using the PianoVAM dataset and discuss additional potential applications.

Daeyong Kwon, SeungHeon Doh, Juhan Nam

Intent classification is a text understanding task that identifies user needs from input text queries. While intent classification has been extensively studied in various domains, it has not received much attention in the music domain. In this paper, we investigate intent classification models for music discovery conversation, focusing on pre-trained language models. Rather than only predicting functional needs: intent classification, we also include a task for classifying musical needs: musical attribute classification. Additionally, we propose a method of concatenating previous chat history with just single-turn user queries in the input text, allowing the model to understand the overall conversation context better. Our proposed model significantly improves the F1 score for both user intent and musical attribute classification, and surpasses the zero-shot and few-shot performance of the pretrained Llama 3 model.

Eunjin Choi, Hyerin Kim, Jiwoo Ryu et al.

A large-scale dataset is essential for training a well-generalized deep-learning model. Most such datasets are collected via scraping from various internet sources, inevitably introducing duplicated data. In the symbolic music domain, these duplicates often come from multiple user arrangements and metadata changes after simple editing. However, despite critical issues such as unreliable training evaluation from data leakage during random splitting, dataset duplication has not been extensively addressed in the MIR community. This study investigates the dataset duplication issues regarding Lakh MIDI Dataset (LMD), one of the largest publicly available sources in the symbolic music domain. To find and evaluate the best retrieval method for duplicated data, we employed the Clean MIDI subset of the LMD as a benchmark test set, in which different versions of the same songs are grouped together. We first evaluated rule-based approaches and previous symbolic music retrieval models for de-duplication and also investigated with a contrastive learning-based BERT model with various augmentations to find duplicate files. As a result, we propose three different versions of the filtered list of LMD, which filters out at least 38,134 samples in the most conservative settings among 178,561 files.

Junhyung Park, Yonghyun Kim, Joonhyung Bae et al.

Piano performance is a multimodal activity that intrinsically combines physical actions with the acoustic rendition. Despite growing research interest in analyzing the multimodal nature of piano performance, the laborious process of acquiring large-scale multimodal data remains a significant bottleneck, hindering further progress in this field. To overcome this barrier, we present an integrated web toolkit comprising two graphical user interfaces (GUIs): (i) PiaRec, which supports the synchronized acquisition of audio, video, MIDI, and performance metadata. (ii) ASDF, which enables the efficient annotation of performer fingering from the visual data. Collectively, this system can streamline the acquisition of multimodal piano performance datasets.

Daeyong Kwon, SeungHeon Doh, Juhan Nam

Recent advancements in Large language models (LLMs) have demonstrated remarkable capabilities across diverse domains. While they exhibit strong zero-shot performance on various tasks, LLMs' effectiveness in music-related applications remains limited due to the relatively small proportion of music-specific knowledge in their training data. To address this limitation, we propose MusT-RAG, a comprehensive framework based on Retrieval Augmented Generation (RAG) to adapt general-purpose LLMs for text-only music question answering (MQA) tasks. RAG is a technique that provides external knowledge to LLMs by retrieving relevant context information when generating answers to questions. To optimize RAG for the music domain, we (1) propose MusWikiDB, a music-specialized vector database for the retrieval stage, and (2) utilizes context information during both inference and fine-tuning processes to effectively transform general-purpose LLMs into music-specific models. Our experiment demonstrates that MusT-RAG significantly outperforms traditional fine-tuning approaches in enhancing LLMs' music domain adaptation capabilities, showing consistent improvements across both in-domain and out-of-domain MQA benchmarks. Additionally, our MusWikiDB proves substantially more effective than general Wikipedia corpora, delivering superior performance and computational efficiency.

Hayeon Bang, Taegyun Kwon, Juhan Nam

This paper presents <Dialogue in Resonance>, an interactive music piece for a human pianist and a computer-controlled piano that integrates real-time automatic music transcription into a score-driven framework. Unlike previous approaches that primarily focus on improvisation-based interactions, our work establishes a balanced framework that combines composed structure with dynamic interaction. Through real-time automatic transcription as its core mechanism, the computer interprets and responds to the human performer's input in real time, creating a musical dialogue that balances compositional intent with live interaction while incorporating elements of unpredictability. In this paper, we present the development process from composition to premiere performance, including technical implementation, rehearsal process, and performance considerations.

Gyubin Lee, Hounsu Kim, Junwon Lee et al.

Deep learning models have seen widespread use in modelling LFO-driven audio effects, such as phaser and flanger. Although existing neural architectures exhibit high-quality emulation of individual effects, they do not possess the capability to manipulate the output via control parameters. To address this issue, we introduce Controllable Neural Frame-based Modulation Effects (CONMOD), a single black-box model which emulates various LFO-driven effects in a frame-wise manner, offering control over LFO frequency and feedback parameters. Additionally, the model is capable of learning the continuous embedding space of two distinct phaser effects, enabling us to steer between effects and achieve creative outputs. Our model outperforms previous work while possessing both controllability and universality, presenting opportunities to enhance creativity in modern LFO-driven audio effects.

Hounsu Kim, Soonbeom Choi, Juhan Nam

Synthesizing performing guitar sound is a highly challenging task due to the polyphony and high variability in expression. Recently, deep generative models have shown promising results in synthesizing expressive polyphonic instrument sounds from music scores, often using a generic MIDI input. In this work, we propose an expressive acoustic guitar sound synthesis model with a customized input representation to the instrument, which we call guitarroll. We implement the proposed approach using diffusion-based outpainting which can generate audio with long-term consistency. To overcome the lack of MIDI/audio-paired datasets, we used not only an existing guitar dataset but also collected data from a high quality sample-based guitar synthesizer. Through quantitative and qualitative evaluations, we show that our proposed model has higher audio quality than the baseline model and generates more realistic timbre sounds than the previous leading work.

Soonbeom Choi, Juhan Nam

Recent studies in singing voice synthesis have achieved high-quality results leveraging advances in text-to-speech models based on deep neural networks. One of the main issues in training singing voice synthesis models is that they require melody and lyric labels to be temporally aligned with audio data. The temporal alignment is a time-exhausting manual work in preparing for the training data. To address the issue, we propose a melody-unsupervision model that requires only audio-and-lyrics pairs without temporal alignment in training time but generates singing voice audio given a melody and lyrics input in inference time. The proposed model is composed of a phoneme classifier and a singing voice generator jointly trained in an end-to-end manner. The model can be fine-tuned by adjusting the amount of supervision with temporally aligned melody labels. Through experiments in melody-unsupervision and semi-supervision settings, we compare the audio quality of synthesized singing voice. We also show that the proposed model is capable of being trained with speech audio and text labels but can generate singing voice in inference time.

SeungHeon Doh, Junwon Lee, Juhan Nam

We propose a machine-translation approach to automatically generate a playlist title from a set of music tracks. We take a sequence of track IDs as input and a sequence of words in a playlist title as output, adapting the sequence-to-sequence framework based on Recurrent Neural Network (RNN) and Transformer to the music data. Considering the orderless nature of music tracks in a playlist, we propose two techniques that remove the order of the input sequence. One is data augmentation by shuffling and the other is deleting the positional encoding. We also reorganize the existing music playlist datasets to generate phrase-level playlist titles. The result shows that the Transformer models generally outperform the RNN model. Also, removing the order of input sequence improves the performance further.

Hsiao-Tzu Hung, Joann Ching, Seungheon Doh et al.

While there are many music datasets with emotion labels in the literature, they cannot be used for research on symbolic-domain music analysis or generation, as there are usually audio files only. In this paper, we present the EMOPIA (pronounced `yee-mò-pi-uh') dataset, a shared multi-modal (audio and MIDI) database focusing on perceived emotion in pop piano music, to facilitate research on various tasks related to music emotion. The dataset contains 1,087 music clips from 387 songs and clip-level emotion labels annotated by four dedicated annotators. Since the clips are not restricted to one clip per song, they can also be used for song-level analysis. We present the methodology for building the dataset, covering the song list curation, clip selection, and emotion annotation processes. Moreover, we prototype use cases on clip-level music emotion classification and emotion-based symbolic music generation by training and evaluating corresponding models using the dataset. The result demonstrates the potential of EMOPIA for being used in future exploration on piano emotion-related MIR tasks.

Kyungyun Lee, Wonil Kim, Juhan Nam

Creating a good drum track to imitate a skilled performer in digital audio workstations (DAWs) can be a time-consuming process, especially for those unfamiliar with drums. In this work, we introduce PocketVAE, a groove generation system that applies grooves to users' rudimentary MIDI tracks, i.e, templates. Grooves can be either transferred from a reference track, generated randomly or with conditions, such as genres. Our system, consisting of different modules for each groove component, takes a two-step approach that is analogous to a music creation process. First, the note module updates the user template through addition and deletion of notes; Second, the velocity and microtiming modules add details to this generated note score. In order to model the drum notes, we apply a discrete latent representation method via Vector Quantized Variational Autoencoder (VQ-VAE), as drum notes have a discrete property, unlike velocity and microtiming values. We show that our two-step approach and the usage of a discrete encoding space improves the learning of the original data distribution. Additionally, we discuss the benefit of incorporating control elements - genre, velocity and microtiming patterns - into the model.

Taegyun Kwon, Dasaem Jeong, Juhan Nam

Recent advances in polyphonic piano transcription have been made primarily by a deliberate design of neural network architectures that detect different note states such as onset or sustain and model the temporal evolution of the states. The majority of them, however, use separate neural networks for each note state, thereby optimizing multiple loss functions, and also they handle the temporal evolution of note states by abstract connections between the state-wise neural networks or using a post-processing module. In this paper, we propose a unified neural network architecture where multiple note states are predicted as a softmax output with a single loss function and the temporal order is learned by an auto-regressive connection within the single neural network. This compact model allows to increase note states without architectural complexity. Using the MAESTRO dataset, we examine various combinations of multiple note states including on, onset, sustain, re-onset, offset, and off. We also show that the autoregressive module effectively learns inter-state dependency of notes. Finally, we show that our proposed model achieves performance comparable to state-of-the-arts with fewer parameters.

Taejun Kim, Minsuk Choi, Evan Sacks et al.

A DJ mix is a sequence of music tracks concatenated seamlessly, typically rendered for audiences in a live setting by a DJ on stage. As a DJ mix is produced in a studio or the live version is recorded for music streaming services, computational methods to analyze DJ mixes, for example, extracting track information or understanding DJ techniques, have drawn research interests. Many of previous works are, however, limited to identifying individual tracks in a mix or segmenting it, and the sizes of the datasets are usually small. In this paper, we provide an in-depth analysis of DJ music by aligning a mix to its original music tracks. We set up the subsequence alignment such that the audio features are less sensitive to the tempo or key change of the original track in a mix. This approach provides temporally tight mix-to-track matching from which we can obtain cue-points, transition length, mix segmentation, and musical changes in DJ performance. Using 1,557 mixes from 1001Tracklists including 13,728 tracks and 20,765 transitions, we conduct the proposed analysis and show a wide range of statistics, which may elucidate the creative process of DJ music making.

Sangeun Kum, Jing-Hua Lin, Li Su et al.

The lack of labeled data is a major obstacle in many music information retrieval tasks such as melody extraction, where labeling is extremely laborious or costly. Semi-supervised learning (SSL) provides a solution to alleviate the issue by leveraging a large amount of unlabeled data. In this paper, we propose an SSL method using teacher-student models for vocal melody extraction. The teacher model is pre-trained with labeled data and guides the student model to make identical predictions given unlabeled input in a self-training setting. We examine three setups of teacher-student models with different data augmentation schemes and loss functions. Also, considering the scarcity of labeled data in the test phase, we artificially generate large-scale testing data with pitch labels from unlabeled data using an analysis-synthesis method. The results show that the SSL method significantly increases the performance against supervised learning only and the improvement depends on the teacher-student models, the size of unlabeled data, the number of self-training iterations, and other training details. We also find that it is essential to ensure that the unlabeled audio has vocal parts. Finally, we show that the proposed SSL method enables a baseline convolutional recurrent neural network model to achieve performance comparable to state-of-the-arts.

Jongpil Lee, Nicholas J. Bryan, Justin Salamon et al.

Deep representation learning offers a powerful paradigm for mapping input data onto an organized embedding space and is useful for many music information retrieval tasks. Two central methods for representation learning include deep metric learning and classification, both having the same goal of learning a representation that can generalize well across tasks. Along with generalization, the emerging concept of disentangled representations is also of great interest, where multiple semantic concepts (e.g., genre, mood, instrumentation) are learned jointly but remain separable in the learned representation space. In this paper we present a single representation learning framework that elucidates the relationship between metric learning, classification, and disentanglement in a holistic manner. For this, we (1) outline past work on the relationship between metric learning and classification, (2) extend this relationship to multi-label data by exploring three different learning approaches and their disentangled versions, and (3) evaluate all models on four tasks (training time, similarity retrieval, auto-tagging, and triplet prediction). We find that classification-based models are generally advantageous for training time, similarity retrieval, and auto-tagging, while deep metric learning exhibits better performance for triplet-prediction. Finally, we show that our proposed approach yields state-of-the-art results for music auto-tagging.

Jongpil Lee, Nicholas J. Bryan, Justin Salamon et al.

Music similarity search is useful for a variety of creative tasks such as replacing one music recording with another recording with a similar "feel", a common task in video editing. For this task, it is typically necessary to define a similarity metric to compare one recording to another. Music similarity, however, is hard to define and depends on multiple simultaneous notions of similarity (i.e. genre, mood, instrument, tempo). While prior work ignore this issue, we embrace this idea and introduce the concept of multidimensional similarity and unify both global and specialized similarity metrics into a single, semantically disentangled multidimensional similarity metric. To do so, we adapt a variant of deep metric learning called conditional similarity networks to the audio domain and extend it using track-based information to control the specificity of our model. We evaluate our method and show that our single, multidimensional model outperforms both specialized similarity spaces and alternative baselines. We also run a user-study and show that our approach is favored by human annotators as well.

Seungheon Doh, Jongpil Lee, Tae Hong Park et al.

Word embedding pioneered by Mikolov et al. is a staple technique for word representations in natural language processing (NLP) research which has also found popularity in music information retrieval tasks. Depending on the type of text data for word embedding, however, vocabulary size and the degree of musical pertinence can significantly vary. In this work, we (1) train the distributed representation of words using combinations of both general text data and music-specific data and (2) evaluate the system in terms of how they associate listening contexts with musical compositions.

Taejun Kim, Juhan Nam

End-to-end learning models using raw waveforms as input have shown superior performances in many audio recognition tasks. However, most model architectures are based on convolutional neural networks (CNN) which were mainly developed for visual recognition tasks. In this paper, we propose an extension of squeeze-and-excitation networks (SENets) which adds temporal feedback control from the top-layer features to channel-wise feature activations in lower layers using a recurrent module. This is analogous to the adaptive gain control mechanism of outer hair-cell in the human auditory system. We apply the proposed model to speech command recognition and show that it slightly outperforms the SENets and other CNN-based models. We also investigate the details of the performance improvement by conducting failure analysis and visualizing the channel-wise feature scaling induced by the temporal feedback.

Jeong Choi, Jongpil Lee, Jiyoung Park et al.

Audio-based music classification and tagging is typically based on categorical supervised learning with a fixed set of labels. This intrinsically cannot handle unseen labels such as newly added music genres or semantic words that users arbitrarily choose for music retrieval. Zero-shot learning can address this problem by leveraging an additional semantic space of labels where side information about the labels is used to unveil the relationship between each other. In this work, we investigate the zero-shot learning in the music domain and organize two different setups of side information. One is using human-labeled attribute information based on Free Music Archive and OpenMIC-2018 datasets. The other is using general word semantic information based on Million Song Dataset and Last.fm tag annotations. Considering a music track is usually multi-labeled in music classification and tagging datasets, we also propose a data split scheme and associated evaluation settings for the multi-label zero-shot learning. Finally, we report experimental results and discuss the effectiveness and new possibilities of zero-shot learning in the music domain.

Jongpil Lee, Jiyoung Park, Juhan Nam

Supervised music representation learning has been performed mainly using semantic labels such as music genres. However, annotating music with semantic labels requires time and cost. In this work, we investigate the use of factual metadata such as artist, album, and track information, which are naturally annotated to songs, for supervised music representation learning. The results show that each of the metadata has individual concept characteristics, and using them jointly improves overall performance.

Kyungyun Lee, Juhan Nam

Previous approaches in singer identification have used one of monophonic vocal tracks or mixed tracks containing multiple instruments, leaving a semantic gap between these two domains of audio. In this paper, we present a system to learn a joint embedding space of monophonic and mixed tracks for singing voice. We use a metric learning method, which ensures that tracks from both domains of the same singer are mapped closer to each other than those of different singers. We train the system on a large synthetic dataset generated by music mashup to reflect real-world music recordings. Our approach opens up new possibilities for cross-domain tasks, e.g., given a monophonic track of a singer as a query, retrieving mixed tracks sung by the same singer from the database. Also, it requires no additional vocal enhancement steps such as source separation. We show the effectiveness of our system for singer identification and query-by-singer in both the same-domain and cross-domain tasks.

Jeong Choi, Jongpil Lee, Jiyoung Park et al.

Music classification and tagging is conducted through categorical supervised learning with a fixed set of labels. In principle, this cannot make predictions on unseen labels. Zero-shot learning is an approach to solve the problem by using side information about the semantic labels. We recently investigated this concept of zero-shot learning in music classification and tagging task by projecting both audio and label space on a single semantic space. In this work, we extend the work to verify the generalization ability of zero-shot learning model by conducting knowledge transfer to different music corpora.

Jiyoung Park, Donghyun Kim, Jongpil Lee et al.

Artist recognition is a task of modeling the artist's musical style. This problem is challenging because there is no clear standard. We propose a hybrid method of the generative model i-vector and the discriminative model deep convolutional neural network. We show that this approach achieves state-of-the-art performance by complementing each other. In addition, we briefly explain the advantages and disadvantages of each approach.