SqueezeComposer: Temporal Speed-up is A Simple Trick for Long-form Music Composing

Jianyi Chen; Rongxiu Zhong; Shilei Zhang; Kun Qian; Jinglei Liu; Yike Guo; Wei Xue

eess.ASMar 22, 2026

SqueezeComposer: Temporal Speed-up is A Simple Trick for Long-form Music Composing

Authors: Jianyi Chen, Rongxiu Zhong, Shilei Zhang, Kun Qian, Jinglei Liu, Yike Guo, Wei Xue

Organizations: The Hong Kong University of Science and Technology · JIUTIAN Research of China Mobile · Beijing Institute of Technology · China Mobile (Hong Kong) Innovation Research Institute

Abstract

Composing coherent long-form music remains a significant challenge due to the complexity of modeling long-range dependencies and the prohibitive memory and computational requirements associated with lengthy audio representations. In this work, we propose a simple yet powerful trick: we assume that AI models can understand and generate time-accelerated (speeded-up) audio at rates such as 2x, 4x, or even 8x. By first generating a high-speed version of the music, we greatly reduce the temporal length and resource requirements, making it feasible to handle long-form music that would otherwise exceed memory or computational limits. The generated audio is then restored to its original speed, recovering the full temporal structure. This temporal speed-up and slow-down strategy naturally follows the principle of hierarchical generation from abstract to detailed content, and can be conveniently applied to existing music generation models to enable long-form music generation. We instantiate this idea in SqueezeComposer, a framework that employs diffusion models for generation in the accelerated domain and refinement in the restored domain. We validate the effectiveness of this approach on two tasks: long-form music generation, which evaluates temporal-wise control (including continuation, completion, and generation from scratch), and whole-song singing accompaniment generation, which evaluates track-wise control. Experimental results demonstrate that our simple temporal speed-up trick enables efficient, scalable, and high-quality long-form music generation. Audio samples are available at https://SqueezeComposer.github.io/.

Figures & tables

Figure 1: The SqueezeComposer illustration: audio is compressed via time-scale modification, generated in a compact domain for various tasks (e.g., long-form composition, accompaniment), and then expanded and refined to the original resolution.

Figure 2: Training pipeline for temporal speeding-up and restoration. The input audio is speeded up and processed through a two-stage pipeline: CNN generates a prior condition, then DiT refines it to produce high-quality restored audio. The accelerated audio maintains the original sampling rate, ensuring vocoder compatibility. Training uses MSE loss for CNN prior generation and diffusion loss for DiT refinement.

Figure 3: Overview of the frameworks using SqueezeComposer for composing. (A) Long-form Music Composing: temporal speed-up enables efficient abstract-level generation using DiT with masking strategies (scratch, completion, continuation). (B) Whole-song Singing Accompaniment Generation: semantic-to-prior mapping followed by DiT refinement for harmonious accompaniment.

Figure 4: Audio duration distribution across four music datasets. All datasets show concentration in the 3-10 minute range, demonstrating that the majority of music files fall within the long-form category, with peaks around 5-7 minutes depending on the dataset type.

Model	Mel ${}_{\text{dis}}$ ( $\downarrow$ )	STFT ${}_{\text{dis}}$ ( $\downarrow$ )	Sampling-rate	TPS ( $\downarrow$ )	F a PS ( $\downarrow$ )
MuCodec Xu et al. (2024a)	3.1204	0.4065	48,000	25	25
Encodec Défossez et al. (2023)	3.3829	0.3266	24,000	600	75
DAC Kumar et al. (2023)	2.1274	0.3682	24,000	450	50
XCodec Ye et al. (2025)	2.4798	0.2263	16,000	400	50
SNAC Siuzdak et al. (2024)	2.6088	0.2518	32,000	156	-
Vocos Siuzdak (2024)	2.5959	0.3378	24,000	-	94.12

Table 1: Comparison of temporal speed-up restoration quality against existing audio intermediate representations. All audio vocoders are off-the-shelf vocoders without any retraining.

Model	Mel ${}_{\text{dis}}$ ( $\downarrow$ )	STFT ${}_{\text{dis}}$ ( $\downarrow$ )	Waveform ${}_{\text{dis}}$ ( $\downarrow$ )	F a PS ( $\downarrow$ )
BigVGAN Lee et al. (2023)	1.2115	0.5005	0.1575	93.75
BigVGAN-Squeeze-4( $\star$ )	1.8367	0.5414	0.1798	23.44
BigVGAN-Squeeze-8( $\star$ )	2.2512	0.6328	0.1822	11.72

Table 2: Comparison of temporal speed-up restoration quality with vocoder fine-tuning.

\star

indicates that the vocoder is fine-tuned using Mel spectrograms restored from temporally squeezed inputs.

Figure 5: Visualization of squeezing and ( $\times$ 4) restoration results on two example samples. From top to bottom: the temporally squeezed Mel spectrogram, the original ground-truth Mel spectrogram, the restored Mel spectrogram using a pretrained vocoder, and the corresponding restoration error; the restored Mel spectrogram using a vocoder fine-tuned on restored Mel spectrogram, and its corresponding error.

Model	CE ( $\uparrow$ )	CU ( $\uparrow$ )	PC ( $\uparrow$ )	PQ ( $\uparrow$ )	FAD(vgg) ( $\downarrow$ )	Duration ( $\uparrow$ )	RTF ( $\downarrow$ )
PyramidCodec Chen et al. (2024a) (continuation)	6.6442	7.9427	4.2218	8.1208	1.5201	180s	10.490
SqueezeComposer_x4 (continuation)	6.8499	7.8684	4.9940	7.8934	1.2321	240s	0.078
SqueezeComposer_x4 (completion)	6.7321	7.8267	4.8574	7.9340	1.3452	240s	0.081
PyramidCodec Chen et al. (2024a) (scratch)	6.6816	7.8873	4.4297	8.0238	-	180s	11.192
SqueezeComposer_x4 (scratch)	6.6919	7.4632	4.7583	7.9916	-	240s	0.079

Table 3: Evaluation Results for Continuation, Completion and from Scratch on rendered Lakh MIDI Dataset.

Model	CE ( $\uparrow$ )	CU ( $\uparrow$ )	PC ( $\uparrow$ )	PQ ( $\uparrow$ )	Duration ( $\uparrow$ )	RTF ( $\downarrow$ )
MusicGen Copet et al. (2023)	4.4182	6.5688	4.3799	6.7619	15s	6.518
AudioLDM Liu et al. (2024)	6.4087	6.8974	4.1393	7.2572	60s	0.005
MusicLDM Chen et al. (2024c)	4.8979	7.0832	4.8742	6.9630	15s	0.011
SqueezeComposer_x4	6.4319	6.8824	5.5222	6.8856	240s	0.074
SqueezeComposer_x8	6.0171	6.7128	5.0624	6.9417	600s	0.037

Table 4: Evaluation Results for Pure Instrumental Music Generation from Scratch.

Figure 6: Music continuation (left) and music completion (right) results using our temporal acceleration framework. All spectrograms are computed from audio that has been accelerated by a factor of 4 , and the generation is performed entirely in this accelerated domain, enabling long-horizon modeling and efficient inference. Top: Input spectrograms with missing regions. Middle: Our generated results, operating on the accelerated representation. Bottom: Ground-truth spectrograms for reference.

Dataset	Model	Config	CE ( $\uparrow$ )	CU ( $\uparrow$ )	PC ( $\uparrow$ )	PQ ( $\uparrow$ )	FAD ( $\downarrow$ )	Duration ( $\uparrow$ )
MUSDB18	FastSAG Chen et al. (2024b)	mixture	6.5173	7.0275	6.2978	6.9921	1.2109	10s
	SingSong Donahue et al. (2023)	mixture	6.1253	6.5430	5.9878	6.8324	0.9084	10s
	SqueezeComposer	mixture, x4, MERT	5.6126	6.7193	6.8603	7.3931	4.1104	240s
In-domain	RandSong	mixture	5.0145	5.6639	3.0143	6.4618	21.6837	240s
	RandSong	instrument	7.3543	7.8033	5.5412	7.8258	0.4331	240s
	FastSAG Chen et al. (2024b)	mixture	6.2566	6.6753	6.7159	7.1718	1.7423	10s

Table 5: Evaluation Results for Singing Accompaniment Generation

Music Accompaniment Generation

See all 12 papers

Jul 16, 2026

WanSong v1.0 Technical Report

Music generation foundation models have recently attracted significant industry attention. However, achieving efficient generation and high-fidelity long-form audio while supporting controllability remains challenging. To address these needs, we present WanSong, a simple yet powerful approach for long-form, commercial-grade song generation. Unlike autoregressive (AR) and cascaded multi-stage pipelines (\eg, AR followed by diffusion), WanSong is a pure diffusion-based model that directly generates high-fidelity, multilingual songs up to 5 minutes and outputs dual stems (vocals and background music) in a single run. In addition, our diffusion framework enables faster inference through step-distillation, and offers an efficient pathway for fine-tuning and customization to support downstream editing tasks.

Binghui Chen, Pandeng Li, Yu Liu +1

Wan Team, Alibaba Group

Controllable Music Generation Audio Diffusion Models Music Generation

Ask PDF

Apr 7, 2026

Anchored Cyclic Generation: A Novel Paradigm for Long-Sequence Symbolic Music Generation

Generating long sequences with structural coherence remains a fundamental challenge for autoregressive models across sequential generation tasks. In symbolic music generation, this challenge is particularly pronounced, as existing methods are constrained by the inherent severe error accumulation problem of autoregressive models, leading to poor performance in music quality and structural integrity. In this paper, we propose the Anchored Cyclic Generation (ACG) paradigm, which relies on anchor features from already identified music to guide subsequent generation during the autoregressive process, effectively mitigating error accumulation in autoregressive methods. Based on the ACG paradigm, we further propose the Hierarchical Anchored Cyclic Generation (Hi-ACG) framework, which employs a systematic global-to-local generation strategy and is highly compatible with our specifically designed piano token, an efficient musical representation. The experimental results demonstrate that compared to traditional autoregressive models, the ACG paradigm achieves reduces cosine distance by an average of 34.7% between predicted feature vectors and ground-truth semantic vectors. In long-sequence symbolic music generation tasks, the Hi-ACG framework significantly outperforms existing mainstream methods in both subjective and objective evaluations. Furthermore, the framework exhibits excellent task generalization capabilities, achieving superior performance in related tasks such as music completion.

Boyu Cao, Lekai Qian, Dehan Li +3

South China University of Technology, School of Future Technology

Symbolic Music Generation Music Generation Autoregressive Generation

Ask PDF

Feb 28, 2026

Efficient Long-Sequence Diffusion Modeling for Symbolic Music Generation

Symbolic music generation is a challenging task in multimedia generation, involving long sequences with hierarchical temporal structures, long-range dependencies, and fine-grained local details. Though recent diffusion-based models produce high quality generations, they tend to suffer from high training and inference costs with long symbolic sequences due to iterative denoising and sequence-length-related costs. To deal with such problem, we put forth a diffusing strategy named SMDIM to combine efficient global structure construction and light local refinement. SMDIM uses structured state space models to capture long range musical context at near linear cost, and selectively refines local musical details via a hybrid refinement scheme. Experiments performed on a wide range of symbolic music datasets which encompass various Western classical music, popular music and traditional folk music show that the SMDIM model outperforms the other state-of-the-art approaches on both the generation quality and the computational efficiency, and it has robust generalization to underexplored musical styles. These results show that SMDIM offers a principled solution for long-sequence symbolic music generation, including associated attributes that accompany the sequences. We provide a project webpage with audio examples and supplementary materials at https://3328702107.github.io/smdim-music/.

Jinhan Xu, Xing Tang, Houpeng Yang +7

School of Computer Science and Artificial Intelligence, Wuhan University of Technology, Wuhan, China 430070 · School of Computer Science, Hubei University of Technology, Wuhan, China 430068 · School of Information Engineering, Hubei University of Economics, Wuhan, China 430205; Hubei Key Laboratory of Digital Finance Innovation, Wuhan, China 430205

Symbolic Music Generation Diffusion Models Diffusion Model Acceleration Sequence Modeling

Ask PDF

Abstract

Music Accompaniment Generation

Explore similar work

WanSong v1.0 Technical Report

Anchored Cyclic Generation: A Novel Paradigm for Long-Sequence Symbolic Music Generation

Efficient Long-Sequence Diffusion Modeling for Symbolic Music Generation