⚡ electrical engineering

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

The paper introduces SqueezeComposer, a framework that enables efficient and high-quality long-form music generation by leveraging a simple temporal speed-up trick to reduce computational requirements during the generation process before restoring the audio to its original speed.

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

Published 2026-03-24

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an architect trying to design a massive, 10-story skyscraper.

The Old Problem:
Usually, AI music generators try to build this skyscraper brick-by-brick, one second at a time. If the building is 10 minutes long, that's 600 seconds of bricks. The AI gets overwhelmed, runs out of memory, and the design starts falling apart because it can't remember what the foundation looked like 5 minutes ago. It's like trying to hold a 10-mile-long rope in your hands; eventually, you have to let go of one end to hold the other.

The New Solution: "SqueezeComposer"
The researchers behind this paper came up with a clever trick called Temporal Speed-up. Think of it as a "Fast-Forward" button for the AI's brain.

Here is how it works, using a simple analogy:

1. The "Time-Squeeze" (Compression)

Instead of asking the AI to build the skyscraper at normal speed, they tell it: "Hey, build this whole 10-minute song, but imagine time is moving 4 times faster."

The Analogy: Imagine you have a long, winding road. Instead of driving the whole road at 60 mph, you shrink the map so the road is now 4 times shorter. You can drive the whole route in a quarter of the time.
What happens: A 10-minute song becomes a "2-minute" song in the AI's eyes. The AI doesn't lose the melody or the structure; it just processes it much faster. This saves a huge amount of computer memory and energy.

2. The "Drafting" Phase (Generation)

Now, the AI builds the music in this "fast-forward" world. Because the song is shorter in this compressed world, the AI can easily remember the beginning while it's working on the end. It creates a coherent, long-form structure without getting confused or running out of memory.

The Analogy: The architect quickly sketches the entire 10-story building on a small piece of paper. Because the paper is small, they can see the whole design at once and make sure the roof matches the foundation.

3. The "Slow-Motion" Phase (Restoration)

Once the AI finishes the "fast" version, the magic happens. The system takes that 2-minute "fast" draft and stretches it back out to real-time speed.

The Analogy: You take that small sketch and blow it up to a giant billboard size. But here's the trick: the AI doesn't just stretch the image; it uses a second, smart step to fill in all the tiny details that were too small to see in the fast version. It adds the texture, the high notes, and the crisp sounds that make it sound like a real, high-quality song.

Why is this a big deal?

It's Simple: They didn't need to invent a new type of brain or a complex new language for the AI. They just changed the speed at which the AI thinks.
It's Flexible: This trick works with almost any existing music AI. It's like putting a turbocharger on a regular car; the engine stays the same, but it goes much further.
It Solves the "Long Song" Problem: Before this, AI could only make short clips (like 30 seconds). Now, it can compose full-length songs (10 minutes or more) that sound consistent from start to finish.

Real-World Examples

The paper tested this in two main ways:

Writing a whole song from scratch: The AI can now compose a full 10-minute symphony without losing its mind.
Writing music for a singer: If you give the AI a human singing a melody, it can now generate a full 5-minute instrumental backing track that fits perfectly, rather than just a 10-second loop.

In summary: SqueezeComposer is like teaching an AI to "think fast" to plan a long journey, and then "slow down" to enjoy the scenery. By squeezing time, they made it possible to create long, beautiful music that was previously too difficult for computers to handle.

1. Problem Statement

Generating coherent, high-quality long-form music (spanning several minutes to hours) remains a significant bottleneck in AI audio generation due to two primary challenges:

Computational & Memory Constraints: Current architectures (autoregressive models and diffusion models) struggle with the sheer sequence length required for long audio. For example, a 5-minute audio represented by a standard 10ms hop-size Mel-spectrogram results in a sequence length of 30,000 frames, making direct modeling computationally prohibitive.
Long-Range Dependency Modeling: Maintaining global musical coherence, thematic continuity, and structural consistency over extended durations is difficult. Existing methods often produce outputs that lack structural integrity or degrade in quality as the duration increases.

While recent works have attempted to compress audio representations (e.g., using latent spaces or discrete codecs), these methods still operate at the original playback speed, meaning the sequence length remains linearly proportional to the audio duration.

2. Methodology: SqueezeComposer

The authors propose SqueezeComposer, a model-agnostic framework based on a simple yet powerful hypothesis: AI models can understand and generate time-accelerated (speeded-up) audio.

The core workflow involves a "Squeeze-and-Restore" paradigm:

A. Temporal Speed-up (Squeezing)

Concept: Instead of processing audio at normal speed, the input audio is accelerated by a factor $r$ (e.g., $4\times$ , $8\times$ , or $16\times$ ).
Effect: This reduces the temporal length of the audio representation by a factor of $r$ . For instance, an $8\times$ speed-up reduces a 10-minute song to a 1.25-minute sequence, drastically lowering memory usage and computational cost.
Feasibility: The authors demonstrate that AI models can effectively learn the musical structure (melody, harmony, rhythm) even when the audio is played at these accelerated rates.

B. Hierarchical Generation

The framework operates in two stages:

Abstract Generation (Accelerated Domain): A diffusion model (specifically a Diffusion Transformer, or DiT) generates music in the compressed, high-speed domain. This stage captures the global structure and abstract musical content.
Detail Refinement (Restored Domain): The generated accelerated representation is expanded back to the original temporal resolution. A second diffusion model (or refinement stage) adds high-frequency details and fine-grained temporal structure to restore the audio to its original speed.

C. Technical Implementation

Representation: The method operates on Mel-spectrograms.
Restoration Pipeline:
- Time-Scale Expansion: The accelerated Mel-spectrogram is upsampled to the original temporal length.
- Two-Stage Training:
  1. Prior Generation: A lightweight CNN predicts a coarse prior representation at the original resolution.
  2. Refinement: A DiT refines this coarse prior into a high-fidelity Mel-spectrogram using a diffusion process.
Vocoder Compatibility: A key finding is that Mel-spectrograms restored from accelerated audio remain compatible with off-the-shelf vocoders (e.g., BigVGAN) without requiring retraining, though fine-tuning the vocoder on restored data further improves quality.

3. Key Contributions

SqueezeComposer Framework: A general, model-agnostic approach that leverages temporal acceleration to reduce sequence length and computational cost while preserving musical coherence.
Hierarchical Generation Paradigm: A novel strategy where the accelerated domain serves as an abstract representation for global structure, and the restoration stage refines it into detailed, high-fidelity audio.
Empirical Validation: The authors validate the approach on two distinct tasks:
- Long-form Music Generation: Including continuation, completion, and generation from scratch.
- Whole-Song Singing Accompaniment: Generating instrumental tracks that align with vocal melodies for full songs.

4. Experimental Results

The authors evaluated SqueezeComposer on datasets including Pure Music, Lakh MIDI, and Song Data (400k songs).

Restoration Quality:
- Even with pretrained vocoders, $4\times$ acceleration maintained perceptually consistent audio with only slight metric degradation.
- Fine-tuning the vocoder on restored Mel-spectrograms effectively closed the quality gap, achieving performance comparable to the original BigVGAN.
- $8\times$ acceleration showed reasonable results, though with higher objective error, demonstrating the scalability of the approach.
Long-form Music Generation:
- Efficiency: SqueezeComposer achieved significantly lower Real-Time Factors (RTF) compared to baselines like PyramidCodec and MusicGen. For example, it generated 4-minute clips in ~0.08s RTF, whereas baselines were much slower or limited to short clips.
- Duration: The model successfully generated music up to 10 minutes (with $8\times$ speed-up), whereas baselines like MusicGen were limited to 15–60 seconds.
- Quality: The model achieved lower Fréchet Audio Distance (FAD) scores and competitive aesthetic scores (Content Enjoyment, Production Quality) compared to state-of-the-art models.
Singing Accompaniment:
- The model successfully generated whole-song accompaniments (240s) with consistent rhythmic and harmonic structure, a task where previous methods (e.g., SingSong, FastSAG) were limited to short 10-second clips.
- Ablation studies showed that focusing on instrument-only tracks yielded cleaner results, and different semantic encoders (MERT vs. muQ) performed comparably.

5. Significance

Scalability: SqueezeComposer offers a computationally efficient solution to the "long-context" problem in audio generation without requiring complex architectural changes or specialized codecs.
Simplicity: The "temporal speed-up" trick is a simple preprocessing step that can be applied to existing diffusion or autoregressive models, making it highly accessible.
Hierarchical Alignment: The approach naturally aligns with the principle of hierarchical generation (abstract-to-detailed), mimicking how humans might conceptualize music structure before filling in details.
Practical Application: It enables the generation of full-length songs and complex multi-track accompaniments, bridging the gap between short AI audio clips and professional-grade long-form music production.

In conclusion, SqueezeComposer demonstrates that by shifting the generation process to a time-accelerated domain, AI models can overcome memory and coherence barriers, enabling efficient and high-quality long-form music composition.