Solving Copyright Infringement on Short Video Platforms: Novel Datasets and an Audio Restoration Deep Learning Pipeline

1 Introduction

Short video platforms like YouTube Shorts, TikTok, and Instagram Reels have transformed content consumption, driving global growth and engagement. In 2023, the market for these platforms was valued at $26.69 billion, with projections reaching $289.52 billion by 2032 [Market Research Future, 2023]. These platforms captivate users, with 66% expressing a preference for short-form videos and spending over an hour daily engaging with them [DemandSage, 2023].

Despite their popularity, user-generated content introduces significant copyright challenges. As presented in Figure 1, users frequently overlay arbitrary background music (BGM) on the generic short videos, disrupting the original audio structure and complicating copyright compliance [Charfeddine et al., 2022; Hasanovic et al., 2022; YouTube Help, 2023]. This practice masks critical audio elements, such as dialogue or original soundtracks (OST), and undermines the accuracy of audio-based near-duplicate video retrieval systems [Avgoustinakis et al., 2021], which are essential for identifying infringements. Addressing these issues demands advanced methods capable of removing arbitrary BGM effectively while preserving the integrity of the original audio.

∗Corresponding authors

Music Source Separation (MSS) [Cano et al., 2018], such as Spleeter [Hennequin et al., 2020] Demucs [D´efossez et al., 2019], presented impressive performance in isolating overlapping audio sources. In our pipeline, MSS is employed for initial audio separation. To tackle the complex task of OSTBGM separation, we further incorporate the state-of-theart BS-RoFormer model [Lu et al., 2024], which leverages a Band-Split module to process audio across multiple frequency bands and Rotary Position Embedding (RoPE) to capture temporal and frequency dependencies. This makes BSRoFormer particularly well-suited for highly blended OSTBGM scenarios. However, while MSS excels at separating mixed sources, it cannot distinguish the OST from BGM, necessitating additional methods for accurate audio restoration.

To address this limitation, we integrate Cross-Modal Video-Music Retrieval (CMVMR) technologies [Jin et al., 2021; Pr´etet et al., 2021; Won et al., 2021; Yi et al., 2021] into our pipeline to align restored OSTs with their corresponding video content. Specifically, we fine-tune the UT-CMVMR

model [Mao et al., 2024], which uses a Video Conformer and Music Conformer enhanced with Optical Flow and Rhythm Quantization to extract spatial-temporal and rhythmic features. Using cross-attention mechanisms and optimizing with Triplet Loss, the model enhances the similarity between video and music embeddings without relying on metadata or tags, making it ideal for short-video environments.

To overcome the scarcity of datasets tailored for arbitrary BGM removal in short video platforms, we introduce two novel datasets: the Oh-Audio Separation Dataset 20K (OASD-20K) and Oh-Short Video Audio Restoration Dataset 160 (OSVAR-160). OASD-20K, designed for mixed-music separation, comprises 20,000 four-second audio clips, which are formed by combining OST tracks from the Korean TV series with BGM tracks from the YouTube Audio Library. This dataset reflects real-world OST-BGM blending complexities and provides high-quality training data for MSS research. OSVAR-160 consists of 160 one-minute video clips sourced from high-quality Korean TV series paired with distinct BGM tracks. It serves as a benchmark for evaluating pipeline performance in removing arbitrary BGM and restoring OST integrity.

This study introduces a novel and optimized pipeline for short video platforms like YouTube Shorts by integrating advanced audio separation and cross-modal retrieval techniques. Our approach effectively removes arbitrary BGM and restores OST integrity, addressing critical copyright challenges. The pipeline is rigorously trained and evaluated using our domain-specific datasets, OASD-20K and OSVAR-160, setting a new benchmark for arbitrary BGM removal in short video restoration tasks. The contributions are as follows:

• A novel pipeline: We introduce an innovative approach that integrates advanced MSS and CMVMR techniques to effectively remove arbitrary BGM, addressing the unique challenges faced by short video platforms.
• Two comprehensive datasets: We present OASD20K, a dataset designed for training and evaluating mixed-music separation, and OSVAR-160, a benchmark dataset for assessing arbitrary BGM removal from video content. These datasets bridge critical gaps in audio separation research for short video applications.
• Real-world applicability: Through rigorous experimentation, we demonstrate the pipeline's ability to preserve audio integrity, enhance content quality, and address copyright compliance on platforms such as YouTube Shorts and TikTok.

By advancing techniques in OST restoration, mixed-music separation, and audio-based video retrieval, this research addresses pressing challenges in copyright compliance and content integrity. The proposed solutions are scalable, supporting sustainable growth and fostering ethical content practices in user-generated environments [Zhao and Guo, 2024].

2 RelatedWork

2.1 Copyright Challenges in Short Videos

The rapid growth of platforms like YouTube and TikTok presents concerns about copyright infringement [AL and Chopra, 2024; Blanding, 2024], spurring the development of advanced detection techniques to address unauthorized content modifications.

Many visual-based duplication detection methods utilize temporal consistency and feature-based analysis to identify duplicate videos [Henry et al., 2024; Han et al., 2024; Udaypur, 2024]. While these approaches are effective at identifying visual similarities, they often falter when faced with audio modifications-a prevalent tactic employed to bypass detection systems. To address this gap, audio-based retrieval methods leverage audio similarity learning techniques to retrieve near-duplicate videos [Avgoustinakis et al., 2021; Liu et al., 2010; Saracoglu et al., 2009].

Short video platforms introduce unique challenges, including inconsistent metadata and highly dynamic user-generated content. Addressing these complexities requires advanced audio processing methods, such as MSS, to isolate and remove arbitrary BGM, thereby restoring the original audio.

2.2 Music Source Separation Approaches

MSS aims to isolate individual components, such as vocals and background music, from mixed audio signals. Traditional signal processing-based methods rely on harmonic structures and time-frequency representations [Cohen and Berdugo, 2001; Virtanen, 2007], but they often struggle with complex audio mixtures.

Recent advancements in deep learning revolutionized MSS. Waveform-based approaches, such as Demucs [D´efossez et al., 2019], operate directly on raw audio signals, capturing fine-grained temporal details to achieve high-quality separation. Spectrogram-based methods [Huang et al., 2014; Jansson et al., 2017], on the other hand, analyze timefrequency representations and utilize CNNs and RNNs for precise audio component isolation [Cho et al., 2014; Sharif Razavian et al., 2014]. The introduction of band-split techniques and positional embeddings [Lu et al., 2024] further enhanced separation accuracy, particularly in scenarios involving overlapping audio sources.

These advancements address the challenges of arbitrary BGM separation in short-form video content, enabling accurate separation of OST from BGM.

2.3 Video-Music Retrieval Approaches

Other early methods, such as the Coherent Bag-of-AudioWords (CBoAW) [Liu et al., 2010], employed audio fingerprinting techniques to detect near-duplicate content and establish relationships between videos and background music. While effective in specific scenarios, these techniques were heavily dependent on handcrafted features, making them inflexible in handling diverse content genres or dynamic environments. As these limitations became evident, researchers began exploring more sophisticated approaches.

Video-Music Retrieval (VMR) methodologies introduced content-based approaches to align audio with video content, leveraging audiovisual features like rhythm and tempo to identify patterns. For instance, CBVMR [Hong et al., 2018] demonstrated robust performance in metadata-scarce environments by directly processing video frames and audio signals. However, these methods often incur substantial costs, limiting their scalability in real-world applications.

To address these challenges, cross-modal retrieval methods emerged as a powerful alternative. Unlike traditional approaches, which process audiovisual features independently and demand significant computational resources, cross-modal methods embed video and audio data into a shared latent space. This reduces the complexity of feature extraction and alignment while maintaining high performance [Gu et al., 2023; Li and Kumar, 2019]. For example, UT-CMVMR [Mao et al., 2024] employs Optical Flow to capture temporal video dynamics and Rhythm Quantization to analyze audio patterns, enabling efficient video-music alignment even in dynamic environments. These methods are well-suited for short video platforms where metadata is sparse.

These advancements in video-music synchronization address the unique demands of short-form video platforms by bypassing metadata reliance and improving scalability for real-world applications.

3 Datasets

To comprehensively train, fine-tune, and evaluate ours, we established two novel datasets and adapted one off-the-shelf dataset, each tailored to a specific module of the pipeline. The OASD-20K dataset was developed to train and fine-tune mixed-music separation models, addressing key challenges in MSS research. For the video-music alignment module, we adapted the HIMV-200K dataset to create a CMVMR fine-tuning dataset, enabling precise training of cross-modal video-music retrieval models. Finally, the OSVAR-160 dataset was designed as a benchmark for evaluating the performance of the entire pipeline under real-world conditions, ensuring its applicability to short video scenarios.

3.1 Mixed-Music Separator Fine-tuning Dataset (OASD-20K)

OASD-20K is the first dataset specifically designed for mixed-music separation tailored to short video platforms. It addresses real-world scenarios where OST and BGM are blended, providing a robust benchmark for advancing MSS research. The dataset emphasizes ethical content usage and copyright compliance, aligning with the needs of modern content platforms. The construction process involves mixing OST and BGM tracks to generate diverse, realistic audio data representative of real-world use cases.

Audio Sources and Preparation

The OST tracks were randomly curated from a Korean TV series OST playlist 1 , while the BGM tracks were sourced from the YouTube Audio Library 2 -a diverse and copyrightcompliant repository commonly exploited by copyright infringers to obscure original OSTs. A total of 50 tracks were sampled from each source, with each track segmented into up to 10 four-second clips. This process yielded 500 unique OST clips and 500 unique BGM clips, creating a rich and varied dataset tailored for mixed-music separation tasks.

1 https://www.youtube.com/playlist?list=RDCLAK5uy miHnhrZJWX7HvNEX7CaSZzfk77iwWhBLI

2 https://www.youtube.com/audiolibrary

Mixing Process and Dataset Splits

The OST and BGM tracks were randomly mixed to generate 20,000 four-second audio clips, resulting in a dataset with approximately 22.2 hours of audio content. The dataset was split into training (16,000 clips; about 17.8 hours), validation (2,000 clips; about 2.2 hours), and test (2,000 clips; about 2.2 hours) sets, adhering to an 8:1:1 ratio. To ensure consistent volume levels across all clips, normalization was applied to -23 LUFS, following standard audio preprocessing practices[EBU-Recommendation, 2011].

3.2 Cross-modal Video-Music Retrieval Fine-tuning Dataset

The CMVMR fine-tuning dataset was derived from the HIMV-200K dataset [Hong et al., 2018], which contains 205,000 video-music pairs designed to ensure strong contextual alignment between video content and corresponding music. For this study, a subset of 1,000 video-music pairs was randomly selected to align with the pipeline's specific objectives. This refined subset enables the fine-tuning of UT-CMVMR, allowing for the classification of music that aligns seamlessly with video content. By eliminating tagbased dependencies, this approach reduces bias and enhances the model's adaptability to metadata-scarce environments.

Preprocessing

The preprocessing pipeline follows the approach utilized in UT-CMVMR [Mao et al., 2024], ensuring compatibility with its architecture and maintaining input consistency. Each of the 1,000 video-music pairs was segmented into 28second units, discarding shorter segments to preserve uniformity. These 28-second units were further divided into 4second sub-clips, increasing the dataset size to support robust model training and evaluation. Video frames were resized to 224×224 pixels, sampled at 1 frame per second (FPS), and concatenated into 896×224 matrices. Corresponding audio clips were transformed into 398×80 Mel-spectrograms, ensuring seamless integration with the model's input requirements. The dataset was divided into training (9,044 pairs), validation (1,130 pairs), and test (1,131 pairs) sets, maintaining an 8:1:1 split ratio.

3.3 Pipeline Performance Evaluation Dataset

OSVAR-160 is a benchmark dataset to evaluate the pipeline's effectiveness in removing arbitrary BGM from short video content. Unlike OASD-20K and HIMV-200K, which focus on fine-tuning specific pipeline modules, it is designed to assess the pipeline's end-to-end performance in real-world scenarios. Arbitrary BGM tracks are overlaid on video content to simulate realistic short video environments, providing a challenging testbed for evaluating the system's ability to restore original audio integrity.

Video and Audio Sources

The video data was curated from publicly available clips of Korean TV series uploaded by tvN 3 , a premier South Korean broadcasting channel renowned for its high-quality productions, compelling storytelling, and superior production value.

3 https://www.youtube.com/@tvNDRAMA official

A total of 160 one-minute clips (approximately 2.67 hours) were selected from their YouTube uploads. For the background music (BGM), distinct tracks were sourced from the YouTube Audio Library, ensuring no overlap with OASD20K to maintain data independence. To ensure uninterrupted coverage, only tracks longer than one minute were chosen, guaranteeing that the BGM fully spans the video duration. This approach ensures consistent and uninterrupted audio content, facilitating robust evaluation of BGM removal tasks.

Dataset Composition

The video-audio pairs, with overlaid BGM, were segmented into 4-second clips, producing a total of 1,121 video-audio pairs. Audio normalization was applied following the same procedure as OASD-20K (-23 LUFS) to maintain consistent volume levels across the dataset. This segmentation process created a diverse set of short clips that include both OST and arbitrary BGM, simulating the complex challenges of short video platforms. The OSVAR-160 benchmark dataset is designed to highlight these challenges, providing a comprehensive testbed to evaluate the effectiveness of the proposed pipeline in arbitrary BGM removal and OST restoration.

4 Method

Our proposed method is designed to address the challenges of removing arbitrarily inserted BGM in short video content. The pipeline comprises four interconnected modules: the Background Music Separator, Speech Extractor, MixedMusic Separator, and Video-Music Matching Classifier. An overview of the pipeline is depicted in Figure 2.

4.1 Initial Audio Separation: BackgroundMusic and Speech Extraction

The first stage of the pipeline focuses on separating mixed audio (arbitrary BGM and original OST) and speech components from the video's audio track, providing clean inputs for downstream tasks. This stage consists of two key modules:

(1) Background Music Separator: The Background Music Separator utilizes the U-Net-based Demucs model [D´efossez et al., 2019] to decompose audio into its constituent tracks, such as vocals, bass, and drums. Demucs operates in the time domain, leveraging recurrent layers to capture long-term dependencies [Ronneberger et al., 2015], which ensures high-quality separation. The separated BGM components are then consolidated into a single "Mixed-Music" track using Pydub, creating a unified input for subsequent modules.

(2) Speech Extractor: The Speech Extractor applies the Pyannote Audio VAD [Bredin et al., 2020] to the vocal output generated by Demucs, accurately detecting and isolating speech. By leveraging recurrent neural networks (RNNs) and temporal convolutional networks (TCNs) [Kim and Reiter, 2017], the module processes overlapping audio frames with contextual embeddings, ensuring robust extraction of dialogue even in noisy or complex audio environments. This precise speech isolation is crucial for maintaining the integrity of spoken content during BGM removal.

This two-step process ensures precise separation of BGM and dialogue. While Demucs isolates the overall vocal track, the VAD module further refines the output by removing residual noise and non-speech elements, yielding clear and isolated dialogue. The refined speech tracks are then synchronized with the muted video, resulting in a "Dialogueonly Video" that provides a clean reference for subsequent audio restoration tasks. This initial stage produces clean, well-separated audio inputs, forming a solid foundation for the downstream tasks of mixed-music separation and videomusic matching.

4.2 Mixed-Music Separator

The BS-RoFormer model [Lu et al., 2024] was adapted to separate OST and BGM in short video platforms, utilizing the OASD-20K dataset to align with real-world OST- BGM blending scenarios. Unlike previous models focused on multi-component separation (e.g., vocals, bass), BS-RoFormer's Band-Split module processes audio signals across parallel frequency bands [Luo and Yu, 2023], ensuring high-resolution spectral details even in short audio clips. Additionally, the Rotary Position Embedding (RoPE) [Su et al., 2024] enhances the model's ability to capture both temporal and frequency dependencies, improving its capacity to handle overlapping audio components. Fine-tuning of the model optimized its mask estimation using Time-domain MAE and Multi-resolution Complex Spectrogram MAE loss functions [Gus´o et al., 2022], which help preserve both temporal and spectral integrity. These refinements enable BSRoFormer to effectively isolate OST from BGM while maintaining high audio quality, making it particularly suited to the complex challenges posed by short video platforms.

4.3 Video-Music Matching

The final module of the pipeline focuses on recommending music that aligns with video content. While the original UTCMVMR model [Mao et al., 2024] employed both CrossEntropy and Triplet Loss for training, we adapted the model for short video platforms by fine-tuning it using only Triplet Loss. This modification reflects the exclusion of tag data and optimizes the similarity between video and music embeddings-treating videos as anchors, paired OSTs as positives, and non-paired BGM as negatives to enhance recommendation accuracy.

The model extracts video and audio embeddings through independent Conformer-based modules that capture both temporal and spatial dynamics. Pre-computed Optical Flow and Rhythm Quantization features, which represent motion and rhythmic characteristics, are embedded and concatenated with modality-specific embeddings. These are then normalized through fully connected layers. A Cross-Attention mechanism integrates these embeddings, learning alignment weights between modalities to capture nuanced relationships, such as rhythmic synchronization and mood alignment.

During training, Triplet Loss selects positive and negative samples based on pre-computed distances. A margin-based objective ensures that paired embeddings are closer in the latent space, while non-paired embeddings are pushed farther apart, optimizing the alignment between video and music. For this module, we use the two outputs from the mixedmusic separators-arbitrary BGM and OST-as inputs. The UT-CMVMR model is then trained to classify which of these two audio components is better suited for the video content.

Figure 3 illustrates the architecture and modifications of this module. By leveraging its architectural strengths and domain-specific fine-tuning, the model effectively addresses the dynamic nature of short video platforms, ensuring accurate music recommendations.

4.4 Integration of Modules

The four modules are systematically integrated into a cohesive pipeline, with the output of each module serving as an essential input for the next. This modular design ensures smooth data flow and effective collaboration between components. The Background Music Separator and Speech Extractor modules work in tandem to produce clean and wellseparated audio inputs, isolating vocal elements and generating 'Mixed-Music' tracks. These outputs form the foundation for the Mixed-Music Separator, which further separates the OST and BGM components, preparing the audio data for the subsequent Video-Music Matching process.

The final module, Video-Music Matching, uses the refined outputs from previous stages to recommend music by aligning its rhythmic characteristics with the visual context of the video. This module is fine-tuned specifically for short video platforms, ensuring the recommendations are contextually appropriate. By integrating these modules, the pipeline can address the challenges of short video platforms, providing a robust solution for arbitrary BGM removal and OST restoration. The complete pipeline structure is illustrated in

Figure 2.

5 Experimental Setup

The training process for the two learned modules, MixedMusic Separator and Video-Music Matching, took approximately 30 hours and 2 hours, respectively, utilizing a GPU cluster with NVIDIA Quadro RTX 6000 GPUs (24GB memory per GPU). The training was conducted in a Python 3.11 environment, with PyTorch 2.5.1+cu121.

For inference, processing one video through the entire pipeline takes approximately 9.17 seconds. This evaluation was performed using our curated OSVAR-160 dataset to ensure that the pipeline's performance is both realistic and comprehensive. A detailed breakdown of inference times for each module is provided in Table 1.

5.1 EvaluationMetrics

This study evaluates the performance of our pipeline using metrics tailored to each module and the overall task.

Mixed-Music Separator

The Mixed-Music Separator module is evaluated using two key metrics: Signal-to-Distortion Ratio (SDR) and ScaleInvariant SDR (SI-SDR), which measure the similarity between the reference signal and the separated signal in dB. SDR quantifies the overall quality of the separation by comparing the energy of the reference signal St to that of the error signal Se = St - Sest:

Higher SDR values indicate better separation quality and signal quality. SI-SDR refines SDR by eliminating the influence of scale differences in signal amplitude, focusing solely on structural similarity. SI-SDR is computed as:

where α = ( S⊤ t Sest ∥St∥2 ). The normalization ensures robustness in scenarios with dynamic volume variations, such as those commonly encountered on short video platforms.

The two metrics complement each other; SDR evaluates overall quality, while SI-SDR isolates structural differences, providing a robust and comprehensive evaluation of the separator's performance in restoring high-quality audio under diverse conditions.

Video-Music Matching

The Video-Music Matching module is evaluated using accuracy, which measures the proportion of correctly identified matches between videos and their intended music (i.e., OST or mismatched BGM). This task is framed as a binary classification problem, where the model determines whether the video is paired with the appropriate music, either the OST or the arbitrary BGM. While some prior studies employed other metrics for retrieval performance, accuracy was selected for this study because it directly aligns to validate precise alignment between video content and music recommendations. This metric ensures that the model not only identifies the correct music but also does so without relying on ranking, which is crucial for real-time applications on short video platforms.

Pipeline Evaluation: Arbitrary BGM Removal

To evaluate the pipeline's overall performance, particularly for arbitrary BGM removal, the same metrics used for the Mixed-Music Separator are applied (SDR and SI-SDR). These metrics assess the system's ability to separate audio components while preserving their temporal and spectral integrity. The use of SDR and SI-SDR ensures that both separation quality and restoration fidelity are rigorously validated.

6 Results

6.1 Mixed-Music Separator

The evaluation of the Mixed-Music Separator module shows significant improvements in OST and BGM separation, owing to domain-specific fine-tuning on the OASD-20K dataset. The proposed approach achieves an average SDR of 11.45 dB and SI-SDR of 11.06 dB for BGM separation, and an SDR of 8.23 dB and SI-SDR of 7.53 dB for OST separation.

For comparison, the original BS-RoFormer model, trained on a different dataset with 6-second audio clips, achieves an average SDR of 9.76 dB across multi-component separation tasks (e.g., vocals, drums, bass). While these tasks and datasets differ, the improved performance highlights the effectiveness of domain-specific fine-tuning for OST-BGM separation in the context of short video platforms.

As shown in Table 2, the fine-tuned model demonstrates robust performance in addressing complex overlaps and transitions characteristic of short-form content.

The improvements in SDR and SI-SDR values indicate that the separated audio components closely resemble the original signals, particularly for BGM separation, where the model demonstrates a strong capability to handle frequent overlaps and rapid transitions typical of short-form content. While the OST separation scores are slightly lower-likely due to the inherent complexity and variability of OST signals-the overall results affirm that BS-RoFormer, with domain-specific fine-tuning, effectively addresses the unique challenges of audio separation in short video platforms. These findings underscore the model's adaptability and robustness in real-world audio environments.

6.2 Video-Music Matching

The evaluation of the Video-Music Matching module demonstrates the superior performance of the UT-CMVMR model compared to the CBVMR baseline. Specifically, UTCMVMR achieves an impressive accuracy of 96.8%, significantly outperforming CBVMR's 65.4%, as detailed in Table 3. This improvement underscores the model's capability to effectively align video and audio content, even in metadatascarce environments such as YouTube Shorts and TikTok.

UT-CMVMR's strong performance can be attributed to its architectural innovations, such as the integration of Optical Flow and Rhythm Quantization modules, which enhance its ability to capture temporal dynamics and rhythmic patterns. The use of advanced cross-modal attention mechanisms further facilitates effective interaction between video and audio embeddings, enabling precise alignment. Fine-tuning on the domain-specific HIMV-200K dataset equips the model to effectively address the unique challenges of short video platforms, ensuring it can recommend rhythmically and contextually appropriate audio.

The high accuracy achieved by UT-CMVMR underscores its effectiveness in handling user-generated content and its practical utility in real-world applications, validating the benefits of domain-specific fine-tuning and architectural enhancements for short video platforms.

6.3 Pipeline Performance

The evaluation of arbitrary BGM separation across three distinct scenarios demonstrates that ours consistently achieves superior performance metrics compared to alternative approaches. These scenarios were designed to benchmark the pipeline's ability to reconstruct audio by comparing the original reference signals with: (1) The final output of ours, which separates arbitrary BGM while restoring OST, (2) Mixedmusic, where arbitrary BGM is added to the original OST, and (3) Dialogue-only audio, where BGM is removed entirely, isolating speech. Each scenario was evaluated against the reference signals using SDR and SI-SDR, offering a direct measure of audio quality and structural similarity.

Our approach achieves an average SDR of 138.63 dB and SI-SDR of 16.53 dB, significantly outperforming the other evaluation scenarios. As shown in Table 4, when arbitrary BGM is added to the original OST, the resulting mixed audio demonstrates a high SDR of 129.51 dB but a considerably lower SI-SDR of 8.65 dB. This indicates that while the overall signal remains clear and loud, the structural quality relative to the original is compromised due to the interference process from the added BGM. The SI-SDR metric here is crucial, as it emphasizes the intrinsic similarity between the reference and estimated signals, unaffected by amplitude discrepancies.

When comparing arbitrary BGM separation to the dialogue-only audio scenario (which achieves an average SDR of 131.24 dB and SI-SDR of 14.00 dB), ours demonstrates slightly better performance. This showcases the model's ability not only to effectively remove the BGM but also to reconstruct the OST in a manner that closely matches the original audio, significantly preserving essential spectral and temporal details. These results validate our effectiveness in maintaining both high-quality and structural similarity across diverse audio scenarios. By comparing the restored outputs to the original reference audio, the evaluation ensures that SDR and SI-SDR metrics reliably reflect the model's capacity to accurately reconstruct audio components.

7 Conclusion

This study proposes a novel pipeline designed to address the challenge of removing arbitrarily inserted BGM in short video platforms. By integrating advanced models with domain-specific fine-tuning, the pipeline demonstrates robust performance in separating OSTs from BGM while effectively restoring the original audio. The Mixed-Music Separator, fine-tuned on the domain-specific OASD-20K dataset, efficiently handles complex audio blends, achieving high SDR and SI-SDR values. The pipeline's final module, VideoMusic Matching, ensures precise alignment between audio and video content, leveraging the comprehensive HIMV200K dataset for fine-tuning. Additionally, the OSVAR-160 dataset serves as a benchmark for evaluating the pipeline under realistic short video scenarios.

In addition, we developed two novel datasets: OASD-20K, comprising 20,000 arbitrary BGM and OST mixed audio clips, and OSVAR-160, consisting of 1,121 video and mixedaudio pairs. These domain-specific datasets serve both as a resource for mixed-music separation tasks and as a benchmark for evaluating audio restoration in short video content.

While our approach achieves strong results, there remain opportunities for improvement. Future work could focus on enhancing the preservation of non-musical audio effects and further improving the quality of separated audio. Expanding the dataset to include diverse genres and languages will also enhance the adaptability across various global platforms.

Acknowledgements

This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ICAN (ICT Challenge and Advanced Network of HRD) support program (RS-202400436934), and the ITRC (Information Technology Research Center) program (IITP-2025-RS-2024-00436936) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation). Moreover, this work was supported by the IITP grant funded by the MSIT (RS2024-00439139, Development of a Cyber Crisis Response and Resilience Test Evaluation Systems).