Year: 2025
Type: conference-paper
Abstract: We present a Neural-Driven Multi-Band Processor (NDMP), an innovative audio processing framework that augments a static six-band Parametric Equalizer (PEQ) with per-band compression, optimized via neural inference for two tasks: Automatic Equalization (AutoEQ) and Production Style Transfer (NDMP-ST). NDMP is trained using a self-supervised strategy where reference audio is degraded with random NDMP parameters and gain, eliminating the need for paired input-target data. As a fully differentiable operator, NDMP supports end-to-end training with audio-domain losses. The model learns to predict tonal and dynamic parameters that match the degraded input to the reference. At inference, AutoEQ adjusts tonal balance for unseen inputs without a reference, while NDMP-ST imposes production-style characteristics by predicting tailored parameters. We evaluated the MUSDB18 dataset with objective metrics (e.g., SI-SDR, SNR, PESQ, STOI) and listening tests and NDMP consistently outperforms traditional PEQ. NDMP offers an efficient and robust neural enhancement framework for modern audio production, enriching static processing with learned dynamic control. © 2025 Parakrant Sarkar et al.
Source: CityU Scholars
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7133402504
FWCI:
Cites:
Cited by:
Related to: