# Distilling a speech and music encoder with task arithmetic

arXiv: https://arxiv.org/abs/2505.13270 · HTML: https://mlsanity.com/paper/2505.13270/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 19 May 2025.

## Built on by (most cited first, 6 of 6)

- [USAD: Universal Speech and Audio Representation via Distillation](https://mlsanity.com/paper/2506.18843/) (2025)
- [Task Vector in TTS: Toward Emotionally Expressive Dialectal Speech Synthesis](https://mlsanity.com/paper/2512.18699/) (2025)
- [A correlation-permutation approach for speech-music encoders model merging](https://mlsanity.com/paper/2506.11403/) (2025)
- [Pseudo2Real: Task Arithmetic for Pseudo-Label Correction in Automatic Speech Recognition](https://mlsanity.com/paper/2510.08047/) (2025)
- [MMMOS: Multi-domain Multi-axis Audio Quality Assessment](https://mlsanity.com/paper/2507.04094/) (2025)
- [Ecologically-Constrained Task Arithmetic for Multi-Taxa Bioacoustic Classifiers Without Shared Data](https://mlsanity.com/paper/2605.03914/) (2026)

## Abstract (arXiv)

Despite the progress in self-supervised learning (SSL) for speech and music, existing models treat these domains separately, limiting their capacity for unified audio understanding. A unified model is desirable for applications that require general representations, e.g. audio large language models. Nonetheless, directly training a general model for speech and music is computationally expensive. Knowledge Distillation of teacher ensembles may be a natural solution, but we posit that decoupling the distillation of the speech and music SSL models allows for more flexibility. Thus, we propose to learn distilled task vectors and then linearly interpolate them to form a unified speech+music model. This strategy enables flexible domain emphasis through adjustable weights and is also simpler to train. Experiments on speech and music benchmarks demonstrate that our method yields superior overall performance compared to ensemble distillation.