Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Latent Space Interpolation for Music Transitions

Rydén, Olle LU (2026) In Master's Theses in Mathematical Sciences FMSM01 20252
Mathematical Statistics
Abstract
Music generation has progressed rapidly in recent years, but practical track-to-track transitions are still largely handled with conventional signal processing such as equal-power crossfades and manual beatmatching. This thesis evaluates whether latent-space interpolation in a pretrained neural audio codec can serve as a short, targeted generation operator for music transitions, where the goal is a coherent bridge rather than a simple overlap of two songs. Using EnCodec, bar-aligned segments are tempo-aligned (with optional pitch preservation), encoded to a latent timeline, and blended frame-wise using different interpolation schedules (linear, sigmoid/eased, and spherical linear interpolation). The resulting decoded transitions are... (More)
Music generation has progressed rapidly in recent years, but practical track-to-track transitions are still largely handled with conventional signal processing such as equal-power crossfades and manual beatmatching. This thesis evaluates whether latent-space interpolation in a pretrained neural audio codec can serve as a short, targeted generation operator for music transitions, where the goal is a coherent bridge rather than a simple overlap of two songs. Using EnCodec, bar-aligned segments are tempo-aligned (with optional pitch preservation), encoded to a latent timeline, and blended frame-wise using different interpolation schedules (linear, sigmoid/eased, and spherical linear interpolation). The resulting decoded transitions are compared against baseline equal-power waveform crossfades built on the same alignment pipeline.

Evaluation combines objective similarity metrics, latent-space diagnostics, and listening tests. Time–frequency coherence (with cross-correlation alignment) is used to compare latent transitions to the baseline, while Principal Component Analysis trajectories, correlation structure, and per-channel frequency contribution probes are used to characterize how EnCodec represents musical structure during interpolation. In an A/B listening test (AB) across multiple transitions (N = 14), latent interpolation (spherical interpolation; SLERP) was preferred at a similar rate as the baseline, with a substantial fraction of “no preference” responses, indicating comparable perceptual quality rather than a clear winner.

Diagnostics further suggest structured latent trajectories and systematic channel behavior, and indicate that latent magnitude correlates with decoded power with a small temporal offset. Overall, latent interpolation in a neural codec provides a viable alternative transition operator and motivates future work on learned refinement, automatic transition point selection, and conditional generative inpainting over codec latents. (Less)
Popular Abstract
Most music transitions are still made using tools such as beatmatching and crossfades. This work investigates whether a neurally trained audio codec can be used to create transitions directly in a compressed latent space, and whether the result can compete with traditional signal processing.
Please use this url to cite or link to this publication:
author
Rydén, Olle LU
supervisor
organization
alternative title
Interpolering i det Latenta Rummet för Musikövergångar
course
FMSM01 20252
year
type
H2 - Master's Degree (Two Years)
subject
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMS-3551-2026
ISSN
1404-6342
other publication id
2026:E18
language
English
id
9225409
date added to LUP
2026-04-20 11:41:24
date last changed
2026-06-04 15:43:18
@misc{9225409,
  abstract     = {{Music generation has progressed rapidly in recent years, but practical track-to-track transitions are still largely handled with conventional signal processing such as equal-power crossfades and manual beatmatching. This thesis evaluates whether latent-space interpolation in a pretrained neural audio codec can serve as a short, targeted generation operator for music transitions, where the goal is a coherent bridge rather than a simple overlap of two songs. Using EnCodec, bar-aligned segments are tempo-aligned (with optional pitch preservation), encoded to a latent timeline, and blended frame-wise using different interpolation schedules (linear, sigmoid/eased, and spherical linear interpolation). The resulting decoded transitions are compared against baseline equal-power waveform crossfades built on the same alignment pipeline.

Evaluation combines objective similarity metrics, latent-space diagnostics, and listening tests. Time–frequency coherence (with cross-correlation alignment) is used to compare latent transitions to the baseline, while Principal Component Analysis trajectories, correlation structure, and per-channel frequency contribution probes are used to characterize how EnCodec represents musical structure during interpolation. In an A/B listening test (AB) across multiple transitions (N = 14), latent interpolation (spherical interpolation; SLERP) was preferred at a similar rate as the baseline, with a substantial fraction of “no preference” responses, indicating comparable perceptual quality rather than a clear winner. 

Diagnostics further suggest structured latent trajectories and systematic channel behavior, and indicate that latent magnitude correlates with decoded power with a small temporal offset. Overall, latent interpolation in a neural codec provides a viable alternative transition operator and motivates future work on learned refinement, automatic transition point selection, and conditional generative inpainting over codec latents.}},
  author       = {{Rydén, Olle}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master's Theses in Mathematical Sciences}},
  title        = {{Latent Space Interpolation for Music Transitions}},
  year         = {{2026}},
}