Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FxSearcher: gradient-free text-driven audio transformation

Hojoon Ki, Jongsuk Kim, Minchan Kwon, Junmo Kim

TL;DR

FxSearcher tackles text-driven audio transformation with a gradient-free approach that searches over an audio FX chain using Bayesian optimization guided by a CLAP-based objective. The key idea is to maximize $S_{final} = S_{target} - S_{guide}$, where $S_{target}$ and $S_{guide}$ measure semantic alignment to the target description and to undesirable artifacts, respectively. By applying a six-FX chain and leveraging a guiding prompt, the method achieves high human and AI-aligned quality while supporting any differentiable or non-differentiable FX plugin, enabling broad sonic diversity. The results indicate strong perceptual relevance and robust performance across speech and instrumental domains, suggesting practical impact for accessible, high-quality, text-driven audio editing.

Abstract

Achieving diverse and high-quality audio transformations from text prompts remains challenging, as existing methods are fundamentally constrained by their reliance on a limited set of differentiable audio effects. This paper proposes FxSearcher, a novel gradient-free framework that discovers the optimal configuration of audio effects (FX) to transform a source signal according to a text prompt. Our method employs Bayesian Optimization and CLAP-based score function to perform this search efficiently. Furthermore, a guiding prompt is introduced to prevent undesirable artifacts and enhance human preference. To objectively evaluate our method, we propose an AI-based evaluation framework. The results demonstrate that the highest scores achieved by our method on these metrics align closely with human preferences. Demos are available at https://hojoonki.github.io/FxSearcher/

FxSearcher: gradient-free text-driven audio transformation

TL;DR

FxSearcher tackles text-driven audio transformation with a gradient-free approach that searches over an audio FX chain using Bayesian optimization guided by a CLAP-based objective. The key idea is to maximize , where and measure semantic alignment to the target description and to undesirable artifacts, respectively. By applying a six-FX chain and leveraging a guiding prompt, the method achieves high human and AI-aligned quality while supporting any differentiable or non-differentiable FX plugin, enabling broad sonic diversity. The results indicate strong perceptual relevance and robust performance across speech and instrumental domains, suggesting practical impact for accessible, high-quality, text-driven audio editing.

Abstract

Achieving diverse and high-quality audio transformations from text prompts remains challenging, as existing methods are fundamentally constrained by their reliance on a limited set of differentiable audio effects. This paper proposes FxSearcher, a novel gradient-free framework that discovers the optimal configuration of audio effects (FX) to transform a source signal according to a text prompt. Our method employs Bayesian Optimization and CLAP-based score function to perform this search efficiently. Furthermore, a guiding prompt is introduced to prevent undesirable artifacts and enhance human preference. To objectively evaluate our method, we propose an AI-based evaluation framework. The results demonstrate that the highest scores achieved by our method on these metrics align closely with human preferences. Demos are available at https://hojoonki.github.io/FxSearcher/

Paper Structure

This paper contains 13 sections, 5 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. PROPOSED METHOD
  3. Overview
  4. Overview
  5. Problem Definition
  6. Score Function
  7. Optimization
  8. EXPERIMENTS
  9. Experimental Details
  10. Evaluation Metrics
  11. Main Result
  12. Ablation Study
  13. CONCLUSIONS

Figures (4)

  • Figure 1: Comparison of operational mechanisms between other FX parameter generation models and our method, FxSearcher.
  • Figure 2: Overview of the proposed pipeline. The framework takes a source audio and a target prompt as input, generating a transformed audio and its corresponding optimal parameter set via an iterative optimization process.
  • Figure 3: Illustration of the two AI-based evaluation methods. Left: Qwen providing a 1-5 absolute rating for a single audio based on a prompt. Right: Gemini performing a pairwise A/B comparison to choose the better of two audio.
  • Figure 4: Effect of the guiding prompt on loudness distribution. The plot (left) shows the density of loudness values (in LUFS), while the table (right) summarizes their mean and standard deviation.