【论文翻译】SCNET: SPARSE COMPRESSION NETWORK FOR MUSIC SOURCE SEPARATION

SCNET: SPARSE COMPRESSION NETWORK FOR MUSIC SOURCE SEPARATION

SCNET：用于音乐源分离的稀疏压缩网络

ABSTRACT

Deep learning-based methods have made significant achievements in music source separation. However, obtaining good results while maintaining a low model complexity remains challenging in super wide-band music source separation. Previous works either overlook the differences in subbands or inadequately address the problem of information loss when generating subband features. In this paper, we propose SCNet, a novel frequency-domain network to explicitly split the spectrogram of the mixture into several subbands and introduce a sparsity-based encoder to model different frequency bands.We use a higher compression ratio on subbands with less information to improve the information density and focus on modeling subbands with more information. In this way, the separation performance can be significantly improved using lower computational consumption. Experiment results show that the proposed model achieves a signal to distortion ratio (SDR) of 9.0 dB on the MUSDB18-HQ dataset without using extra data, which outperforms state-of-the-art methods. Specifically, SCNet’s CPU inference time is only 48% of HT Demucs, one of the previous state-of-the-art models.
Index Terms— Music separation, frequency domain, subband, sparse compression

摘要

基于深度学习的方法在音乐源分离领域取得了显著成果。然而，在超宽带音乐源分离中，如何在保持较低模型复杂度的同时获得良好的分离效果仍然是一个挑战。以往的研究要么忽略了子带之间的差异，要么在生成子带特征时未能充分解决信息丢失的问题。本文提出了一种新型频域网络SCNet，用于将混合信号的频谱图显式地分割成多个子带，并引入基于稀疏性的编码器来对不同的频段进行建模。我们对信息量较少的子带采用更高的压缩比来提高信息密度，并专注于对信息量较多的子带进行建模。这样，在降低计算消耗的同时，可以显著提高分离性能。实验结果表明，所提出的模型在MUSDB18-HQ数据集上无需额外数据即可达到9.0 dB的信号失真比（SDR），优于现有方法。具体来说，SCNet 的 CPU 推理时间仅为 HT Demucs（此前最先进的模型之一）的 48%。

关键词——音乐分离、频域、子带、稀疏压缩

1. INTRODUCTION

Music source separation (MSS) aims to separate the pure source signal and accompaniment from the mixed signal, such as separating human voice and accompaniment from singing. Many prevailing MSS methodologies predominantly explore the utilization of the frequency-domain [1, 2, 3, 4] or a fusion of time-domain and frequency-domain [5, 6, 7]. While these approaches have yielded impressive results, they often overlook the disparities between musical signals across various frequency bands, which subsequently becomes a performance bottleneck. BSRNN [8] explicitly splits the spectrogram of the mixture into subbands and performs interleaved band-level and sequence-level modeling to get the state-of-the-art separation performance. However, the method of splitting subbands in BSRNN is very complicated, and it is necessary to try different schemes to get the best results. Even more, the band-splitting scheme needs to be further modified according to the sound of specific instruments, which further increases the manual design operation.

1. 引言

音乐源分离 (MSS) 旨在从混合信号中分离出纯净的源信号和伴奏，例如从歌唱中分离出人声和伴奏。许多主流的 MSS 方法主要探索频域 [1, 2, 3, 4] 或时域与频域融合 [5, 6, 7] 的方法。虽然这些方法取得了令人瞩目的成果，但它们往往忽略了不同频段音乐信号之间的差异，从而成为性能瓶颈。BSRNN [8] 将混合信号的频谱图显式地分割成子带，并进行交错的频带级和序列级建模，从而获得了目前最先进的分离性能。然而，BSRNN 中分割子带的方法非常复杂，需要尝试不同的方案才能获得最佳结果。此外，还需要根据特定乐器的声音进一步修改频段分割方案，这进一步增加了手动设计操作。

Moreover, music signals are typically recorded at a higher sample rate (e.g., 44.1k Hz) [9], which offers great details and enables neural network models to learn richer audio representations and achieve better separation performance. This, however, comes at the cost of linearly growing computational complexity, making them less deployable for resource-constrained applications. The simplest solution to this challenge is to downsample the audio to a lower sample rate. Nevertheless, this will drop the fine details captured from the high-sample rate sensor, and the omitted details could limit the model’s performance [7].Dropping details uniformly at all positions is clearly suboptimal, as not all frequency parts are equally informative. Within a spectrogram, the lower-frequency part containing more detailed features is more critical than the higherfrequency part. Inspired by [10], sparse, high-resolution features are far more informative than dense, low-resolution ones. A very natural idea is to skip computations for less-important parts and focus on more-important details.
此外，音乐信号通常以较高的采样率（例如 44.1k Hz）记录[9]，这提供了丰富的细节，并使神经网络模型能够学习更丰富的音频表示并实现更好的分离性能。然而，这是以线性增长的计算复杂性为代价的，使得它们不太适合资源受限的应用程序部署。应对这一挑战的最简单的解决方案是将音频下采样到较低的采样率。然而，这会丢失从高采样率传感器捕获的精细细节，并且省略的细节可能会限制模型的性能[7]。在所有位置均匀地丢失细节显然不是最佳的，因为并非所有频率部分都具有相同的信息。在频谱图中，包含更详细特征的低频部分比高频部分更重要。受此启发[10]，稀疏的高分辨率特征比密集的低分辨率特征提供更多信息。一个非常自然的想法是跳过不太重要部分的计算并专注于更重要的细节。

Therefore, in this paper, we propose an innovative frequencydomain network architecture called SCNet, which consists of an audio encoder, a separation network based on dual-path RNN [11], and an audio decoder.Inspired by [12], the proposed architecture of the encoder and decoder contains top-down, bottom-up, and skip connections, mirroring the brain’s hierarchical processing of sensory information [7, 13, 14].Considering the disparities among frequency subbands, we also employ different processing for them. Specifically, different from the intricate subband-splitting method used by BSRNN [8], we simply divide the spectrogram into three distinct frequency subbands: low, medium, and high, which dramatically reduces the complexity of splitting subbands.Moreover, unlike BSRNN, which has to set different parameters for different musical sources separately, we only need unified model parameters to separate multiple music sources.Furthermore, in order to make full use of high sampling rate music information to improve the separation performance and reduce the computational complexity of the model, we use different compression ratios and modeling operations for different frequency subbands. Recognizing that the low-frequency region is information-rich, while the mid- and high-frequency bands are comparatively less, we use a larger compression ratio for the midand high-frequency subbands to increase the information density.In addition, since the information in the low-frequency subband is more complex, we use relatively more parameters to focus on modeling it.In this way, the separation performance can be significantly improved using lower computational consumption. Experiment results show that the proposed model trained on the MUSDB18-HQ [9] dataset substantially outperforms the state-of-the-art model with 9.0 dB SDR.And the CPU inference time is only 48% of HT Demucs.

因此，本文提出了一种名为SCNet的创新频域网络架构，它由音频编码器、基于双路径RNN的分离网络[11]和音频解码器组成。受[12]的启发，所提出的编码器和解码器架构包含自顶向下、自底向上和跳跃连接，模拟了大脑对感觉信息的分层处理[7, 13, 14]。考虑到不同频带之间的差异，我们也对它们采用了不同的处理方法。具体而言，与 BSRNN [8] 使用的复杂子带分割方法不同，我们简单地将频谱图分割成三个不同的频率子带：低频、中频和高频，这显著降低了子带分割的复杂度。此外，与 BSRNN 需要为不同的音乐源分别设置不同的参数不同，我们只需要统一的模型参数即可分离多个音乐源。而且，为了充分利用高采样率音乐信息来提高分离性能并降低模型的计算复杂度，我们对不同的频率子带采用了不同的压缩比和建模操作。考虑到低频区域信息丰富，而中频和高频区域信息相对较少，我们对中频和高频子带使用了更大的压缩比来增加信息密度。此外，由于低频子带中的信息更为复杂，我们使用了相对更多的参数来对其进行建模。通过这种方式，可以在降低计算消耗的同时显著提高分离性能。实验结果表明，所提出的模型在 MUSDB18-HQ [9] 数据集上训练后，性能显著优于最先进的模型，SDR 为 9.0 dB。CPU 推理时间仅为 HT Demucs 的 48%。

2. METHOD

2.1. Overall structure

Fig.1 illustrates the entire workflow of SCNet. The original mixture is initially subjected to a Short-Time Fourier Transform (STFT) calculation:
y = STFT(x) (1)
For a given waveform x, the STFT operation converts the waveform into a complex spectrogram y ∈ RF ×T ×(2·2), where F and T denote the frequency and time dimensions, respectively. The term (2 · 2) signifies the binaural channels and their respective real and imaginary components. The model takes y as its input and is composed of three main parts: an audio encoder, a separation network, and an audio decoder, which is similar to Conv-TasNet [15].The encoder aims to compress the spectrogram while enhancing its feature dimensions. The separation network then endeavors to learn sequence dependencies. Ultimately, the decoder reconstructs the spectrogram of each source. In the following, the separation step is elaborated.

2. 方法

2.1. 整体结构

图 1 展示了 SCNet 的完整工作流程。首先对原始混合信号进行短时傅里叶变换 (STFT) 计算：

y = STFT(x) (1)

对于给定的波形 x，STFT 操作将其转换为复数频谱图 y ∈ RF ×T ×(2·2)，其中 F 和 T 分别表示频率和时间维度。术语 (2 × 2) 表示双耳声道及其各自的实部和虚部。该模型以 y 作为输入，由三个主要部分组成：音频编码器、分离网络和音频解码器，类似于 Conv-TasNet [15]。编码器旨在压缩频谱图并增强其特征维度。分离网络则致力于学习序列依赖关系。最终，解码器重构每个声源的频谱图。以下将详细阐述分离步骤。

2.2. Encoder

The encoder is designed to reduce the resolution of ultra-wideband audio, laying the groundwork for subsequent processing steps. For an input y, the encoder processes y step by step through three sparse down-sampling blocks (SD blocks). Each traversal through a block modifies the frequency dimension’s length from Fin to Fin · R, where R represents the retention rate. Concurrently, the feature dimension undergoes adjustments, contingent upon the hyperparameters we establish.

2.2 编码器

编码器旨在降低超宽带音频的分辨率，为后续处理步骤奠定基础。对于输入 y，编码器通过三个稀疏下采样块（SD 块）逐步处理 y。每次遍历一个块都会将频率维度的长度从 Fin 修改为 Fin · R，其中 R 代表保留率。同时，特征维度也会根据我们设定的超参数进行调整。

SD blocks stand as the encoder’s centerpiece. Fig.2 shows its detailed architecture, incorporating a sparse down-sampling layer alongside a series of stacked convolution modules. Sparse downsampling layer (SD layer) engineered with three parallel convolution layers to compress the frequency axis, the strides of these convolutions are 1, 4, and 16 respectively. Such a configuration indicates a tripartite division of the entire frequency band, applying a higher compression ratio in the high frequency part.In order to reduce the model complexity, we use the same frequency band division ratio in the down-sampling layers of all SD blocks. Additionally, SD layer contributes to an elevation in the feature dimension, enabling the model to capture more intricate details. A GELU activation function [16] ensues post the down-sampling layer.

SD模块是编码器的核心。图2展示了其详细架构，包括一个稀疏下采样层和一系列堆叠的卷积模块。稀疏下采样层（SD层）由三个并行的卷积层组成，用于压缩频率轴，这些卷积层的步长分别为1、4和16。这种配置将整个频带划分为三个部分，在高频部分应用更高的压缩比。为了降低模型复杂度，我们在所有SD模块的下采样层中使用相同的频带划分比例。此外，SD层有助于提升特征维度，使模型能够捕捉到更复杂的细节。下采样层之后应用GELU激活函数[16]。

The design of the convolution module is inspired by the Conformer [17].As illustrated in Fig.3, we have opted for Groupnorm in lieu of Layernorm and Batchnorm as presented in the original design. In the low-frequency segment, we have incorporated additional convolutional blocks to enhance the granularity of our modeling.

卷积模块的设计灵感来源于 Conformer [17]。如图 3 所示，我们选择了 Groupnorm 代替原始设计中使用的 Layernorm 和 Batchnorm。在低频部分，我们加入了额外的卷积块以增强建模的粒度。

On the whole, the encoder module prioritizes the preservation of low-frequency details.

总体而言，编码器模块优先考虑保留低频细节

2.3. Separation with dual-path module

The encoder primarily focuses on modeling inter subband dependencies due to the pronounced differences in information density across various subbands. After compressing sufficiently, the information density becomes more uniform, shifting our attention to global information modeling.To this end, we employ the dual-path RNN architecture from BSRNN. However, stacking multiple dual-path layers for the same sequence learning leads to limited performance due to diminishing marginal utility. TFCNet [18] addresses this by introducing a feature conversion module to project features into a new space. Inspired by this, we integrate the torch.rfft and torch.irfft functions [19] between adjacent dual-path layers.Specifically, the output from odd-numbered dual-path layers is denoted as Yi ∈ RB×Fr×T ×C(i = 1, 3, ...), with Fr signifying the compressed frequency dimension. The rfft function is applied:

Y˙i = RFFT(Yi[b, f, :, c]) ∈ RB×Fr×( T2 +1)×C
b = 1, · · ·, B f = 1, · · ·, F r c = 1, · · ·, C (2)

where Y˙i is a complex tensor. This tensor is split into its real and imaginary components, which are then concatenated in the feature
dimension, resulting in Xi+1 ∈ RB×Fr×( T2 +1)×2C for the next
layer. For the output of even-numbered layers, we adopt the inverse process of the aforementioned conversion.

2.3. 基于双路径模块的分离

编码器主要关注子带间依赖关系的建模，这是由于不同子带间的信息密度差异显著。经过充分压缩后，信息密度趋于均匀，此时我们需要将注意力转移到全局信息建模上。为此，我们采用了 BSRNN 中的双路径 RNN 架构。然而，对于同一序列的学习，堆叠多个双路径层会导致边际效用递减，从而限制性能。TFCNet [18] 通过引入特征转换模块将特征投影到新的空间来解决这个问题。受此启发，我们在相邻的双路径层之间集成了 torch.rfft 和 torch.irfft 函数 [19]。具体而言，奇数层双路径层的输出表示为 Yi ∈ RB×Fr×T×C (i = 1, 3, ...)，其中 Fr 表示压缩后的频率维度。应用 RFFT 函数：

Y˙i = RFFT(Yi[b, f, :, c]) ∈ RB×Fr×( T2 +1)×C

b = 1, · · ·, B f = 1, · · ·, F c = 1, · · ·, C ...对于偶数层的输出，我们采用上述转换的逆过程。

2.4. Decoder

The decoder employs skip connections to integrate hierarchical features from the encoder, gradually reconstructing the separated spectrogram via the sparse up-sampling layer (SU layer). Within these skip connections, we incorporate a fusion layer. Initially, the two inputs are summed together, and the result is duplicated across the feature dimension. Subsequently, a 2-D convolution with a kernel size of 3 and a stride of 1 is applied, followed by a gated linear unit (GLU) [20] layer:

GLU(a, b) = a × sigmoid(b) (3)
where input is split in half along feature dimension to form a and b. Given the prior replication steps, the GLU operation induces an effect akin to the swish function.

2.4 解码器

解码器采用跳跃连接来整合来自编码器的分层特征，并通过稀疏上采样层（SU层）逐步重建分离的频谱图。在这些跳跃连接中，我们加入了一个融合层。最初，两个输入信号相加，然后将结果沿特征维度进行复制。随后，应用一个卷积核大小为 3、步长为 1 的二维卷积，之后连接一个门控线性单元 (GLU) [20] 层：

GLU(a, b) = a × sigmoid(b) (3)

其中输入沿特征维度被分成两半，形成 a 和 b。鉴于先前的复制步骤，GLU 操作会产生类似于 swish 函数的效果。

2.5. Loss function

Previous work generally considers waveform similarity as the training target. But this is not so closely related to the spectrogram. So we use the root mean squared error (RMSE) loss of complex-valued spectrogram as the loss funtion:

(4)
whereandrepresent the real and imaginary parts of the source spectrogram respectively.

2.5 损失函数

以往的研究通常以波形相似度作为训练目标。但这与频谱图的相关性并不强。因此，我们采用复值频谱图的均方根误差（RMSE）损失作为损失函数：

(4)

其中和分别表示源频谱图的实部和虚部。

3. EXPERIMENTAL SETUP

3.1. Dataset

We validate the proposed algorithms on MUSDB18-HQ [9], the most popular dataset to date to benchmark binaural music separation algorithms. It contain 150 full lengths music tracks of different genres along with their isolated drums, bass, vocals and others stems. In addition to this, we also use MoisesDB [21] as a supplementary dataset. MoisesDB is a comprehensive multitrack dataset of 240 previously unreleased songs by 47 artists, covering 12 advanced genres. In the experiment, we will see how much performance improvement will be achieved by using MoisesDB as additional training data. Besides, the model’s adaptability will also be tested on MoisesDB.

3. 实验设置

3.1. 数据集

我们使用 MUSDB18-HQ [9] 数据集验证所提出的算法。MUSDB18-HQ 是目前最流行的双耳音乐分离算法基准测试数据集，包含 150 首不同风格的完整音乐曲目，以及分离出的鼓、贝斯、人声和其他音轨。此外，我们还使用 MoisesDB [21] 作为补充数据集。MoisesDB 是一个包含 240 首此前未发行歌曲的综合多轨数据集，这些歌曲来自 47 位艺术家，涵盖 12 种高级音乐风格。在实验中，我们将观察使用 MoisesDB 作为额外训练数据能带来多大的性能提升。此外，我们还将测试模型在 MoisesDB 上的适应性。

3.2. Model and training configurations

During training, we split the soundtrack into overlapped segments of duration 11 seconds. The time interval between two adjacent segments is 1 second. Additionally, we adopt the data augmentation methods, remix and scale, following the ones used in Demucs [13].Remix achieves data simulation by dynamically mixing tracks of different songs during training, while scale enhances model robustness by randomly adjusting the volume of tracks.

3.2 模型和训练配置

在训练过程中，我们将音轨分割成时长为 11 秒的重叠片段，相邻片段之间的时间间隔为 1 秒。此外，我们采用了数据增强方法，包括 remix 和 scale，这些方法与 Demucs [13] 中使用的方法类似。Remix 通过在训练过程中动态混合不同歌曲的音轨来实现数据模拟，而 scale 则通过随机调整音轨的音量来增强模型的鲁棒性。

The STFT window size is 92ms (4096-point FFT) with a hop size of 23ms. Thus, the spectrogram will have 2049 frequency bins. We do not use a window function in the STFT operation, as this has no effect on the results. Three sparse down-sampling layers (SD layers) sequentially increase features to 32, 64, and 128 dimensions. The hidden layer size of the convolutional module is one-fourth of the input size, with the three convolution layers in its middle having kernel sizes of 3, 3, and 1, respectively. We stack 6 dual-path RNN layers in the separation network. At odd layers, the number of hidden units is 128 in BiLSTM, which goes up to 256 in even-numbered layers.

短时傅里叶变换 (STFT) 的窗口大小为 92 毫秒（4096 点 FFT），步长为 23 毫秒。因此，频谱图将包含 2049 个频率区间。我们在 STFT 操作中未使用窗函数，因为这不会影响结果。三个稀疏下采样层（SD 层）依次将特征维度提升至 32、64 和 128。卷积模块的隐藏层大小为输入维度的四分之一，中间三个卷积层的卷积核大小分别为 3、3 和 1。我们在分离网络中堆叠了 6 个双路径循环神经网络 (RNN) 层。在奇数层，双向长短期记忆网络 (BiLSTM) 的隐藏单元数为 128，而在偶数层，隐藏单元数增加到 256。

All our experiments are conducted on 8 Nvidia V100 GPUs. When training solely on the MUSDB18-HQ dataset, the model is trained for 130 epochs with the Adam [22] optimizer with an initial learning rate of 5e-4 and batch size of 4 for each GPU. Nevertheless, we adjust the learning rate to 3e-4 when introducing additional data to mitigate potential gradient explosion.

所有实验均在 8 个 Nvidia V100 GPU 上进行。当仅使用 MUSDB18-HQ 数据集进行训练时，模型使用 Adam [22] 优化器训练 130 个 epoch，初始学习率为 5e-4，每个 GPU 的批大小为 4。然而，当引入更多数据时，我们将学习率调整为 3e-4，以缓解潜在的梯度爆炸问题。

3.3. Evaluation metrics

In the experiments, we utilize Signal-to-Distortion Ratio (SDR) [23], as computed by museval [24] as the primary evaluation metric. Furthermore, Real-Time Factor (RTF) is also used to assess the model’s efficiency during testing. RTF is defined as the ratio of the time required to process a fixed input to the duration of that input.

3.3 评估指标

在实验中，我们采用信号失真比 (SDR) [23] 作为主要评估指标，该指标由 Museval [24] 计算得出。此外，我们还使用实时因子 (RTF) 来评估模型在测试期间的效率。RTF 定义为处理固定输入所需时间与该输入持续时间的比值。

4. RESULTS

4.1. Effect of sparsity rate

Table 1 shows the influence of different sparsity rates on performance. The term global compression ratio (GCR) denotes the reduction percentage of the frequency dimension following each SD block. Sparsity rate (SR) indicates the division strategy of the three frequency bands. As the results show, a reduced GCR will significantly improve performance, however, it concurrently increases computational demands. To manage computational expenses, we efrain from testing even lower GCR values. While SR exerts a minimal influence on efficiency, its correct selection is paramount. The results suggest that peak performance is reached when the share of low-frequency content approximates 17.5%. If it falls below this range, it might cause the loss of essential information. Conversely, exceeding this range can lead to a reduced range of the mid-frequency band, thereby impacting modeling effectiveness. Based on our experimental findings, we have selected a configuration with GCR set at 70% and a low-frequency proportion of 17.5% for our final model, as it delivers the optimal average SDR.

4. 结果

4.1. 稀疏率的影响

表1显示了不同稀疏率对性能的影响。全局压缩比（GCR）表示每个SD块之后频率维度的缩减百分比。稀疏率（SR）表示三个频段的划分策略。结果表明，降低GCR可以显著提高性能，但同时也会增加计算需求。为了控制计算成本，我们没有测试更低的GCR值。虽然SR对效率的影响很小，但正确选择SR至关重要。结果表明，当低频成分占比约为17.5%时，性能达到峰值。如果低于此范围，可能会导致关键信息的丢失。相反，如果高于此范围，则会导致中频频段范围缩小，从而影响建模效果。根据我们的实验结果，我们最终选择 GCR 设置为 70%，低频比例为 17.5% 的配置作为我们的最终模型，因为它提供了最佳的平均 SDR。

4.2. Results on MUSDB18-HQ

Table 2 compares the proposed SCNet with other previous works on the MUSDB18-HQ dataset. SCNet achieves 9.0 dB mean SDR on the test set without extra training data. Remarkably, our model marks a considerable advancement in the SDR metrics for drums and bass. Simultaneously, it keeps pace with leading-edge models in SDR metrics for vocals and other. We also utilize an additional dataset of 235 four-track songs provided by MoisesDB for training.By augmenting the training data, we observe improvements in the performance across all tracks. Furthermore, it’s worth noting that SCNet boasts a notably streamlined architecture with just 10.08M parameters, which is a mere quarter of the parameter count in HT Demucs. Attempting to stack more parameters in the large version, we doubled the channel dimension and conducted training using PyTorch’s mixed precision approach. This configuration resulted in a total of 41.2M parameters.

4.2. MUSDB18-HQ 数据集上的结果

表 2 将我们提出的 SCNet 与之前在 MUSDB18-HQ 数据集上的其他工作进行了比较。SCNet 在测试集上实现了 9.0 dB 的平均 SDR，且无需额外的训练数据。值得注意的是，我们的模型在鼓和贝斯的 SDR 指标方面取得了显著进步。同时，它在人声和其他声道的 SDR 指标方面也与领先模型保持同步。我们还使用了 MoisesDB 提供的包含 235 首四轨歌曲的额外数据集进行训练。通过扩充训练数据，我们观察到所有声道的性能均有所提升。此外，值得一提的是，SCNet 拥有非常精简的架构，仅有 1008 万个参数，仅为 HT Demucs 参数数量的四分之一。为了在大规模版本中增加参数，我们将通道维度增加了一倍，并使用 PyTorch 的混合精度方法进行训练。此配置共产生了 4120 万个参数。

4.3. Results of generalizability test on MoisesDB
We further assess the generalizability of our approach in Table 3. SCNet is trained on the MUSDB18-HQ and evaluated on the MoisesDB without fine-tuning. As a point of comparison, HT Demucs was trained with an additional 800 songs. The results demonstrate
that SCNet exhibits superior generalization even with limited data.

4.3. MoisesDB 上的泛化能力测试结果

我们在表 3 中进一步评估了我们方法的泛化能力。SCNet 在 MUSDB18-HQ 数据集上训练，并在 MoisesDB 数据集上进行评估，未进行微调。作为对比，HT Demucs 使用额外的 800 首歌曲进行训练。结果表明：

即使在数据有限的情况下，SCNet 也展现出优异的泛化能力。

4.4. Ablation experiments
To further validate the effectiveness of sparse compression, we conduct ablation experiments. Presented in Table 4, the control group utilizes the same frequency segmentation method but applies 4*down-sampling to each subband, retaining a GCR of 75%. The results indicate that sparse compression is crucial for achieving our desired outcomes.

4.4 消融实验

为了进一步验证稀疏压缩的有效性，我们进行了消融实验。如表4所示，对照组采用相同的频率分割方法，但对每个子带应用4倍下采样，保留75%的GCR。结果表明，稀疏压缩对于实现我们预期的结果至关重要。

5. CONCLUSION

In this paper, we propose an innovative frequency-domain network architecture called SCNet, which explicitly splits the spectrogram of the mixture into three subbands and introduces a sparsity-based encoder to model different subbands. By applying a higher compression ratio to the mid- and high-frequency subbands, we enhance the overall information density. This heightened density facilitates a more efficient utilization of the separation network. Experiment results show that, with lower computational consumption, SCNet can surpass the performance of existing state-of-the-art music source separation methods on the MUSDB18-HQ dataset, whether using extra data or not. In the future, we will study better sub-band splitting methods and further improve the separation module.

5. 结论

本文提出了一种名为 SCNet 的创新型频域网络架构。该架构将混合信号的频谱图显式地分割成三个子带，并引入基于稀疏性的编码器对不同的子带进行建模。通过对中频和高频子带应用更高的压缩比，我们提高了整体信息密度。这种更高的信息密度有助于更有效地利用分离网络。实验结果表明，无论是否使用额外数据，SCNet 在 MUSDB18-HQ 数据集上都能以更低的计算消耗超越现有最先进的音乐源分离方法。未来，我们将研究更优的子带分割方法，并进一步改进分离模块。