Warping: Data-driven Mixture Preprocessing to Boost the Performance of Blind Speech Separation

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F46747885%3A24220%2F25%3A00013595" target="_blank" >RIV/46747885:24220/25:00013595 - isvavai.cz</a>

Výsledek na webu

<a href="https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11226076&utm_source=scopus&getft_integrator=scopus" target="_blank" >https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11226076&utm_source=scopus&getft_integrator=scopus</a>

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Warping: Data-driven Mixture Preprocessing to Boost the Performance of Blind Speech Separation

Popis výsledku v původním jazyce

Blind source separation (BSS) can be used to recover speech signals from mixtures recorded by microphones. However, their performances show significant limitations because of the deviations between the instantaneous mixing model and real audio mixtures transformed into the short-time Fourier domain (STFT). This paper presents a data-driven preprocessing technique called mixture warping, which aims to adjust the mixture to obey the instantaneous model as much as possible. As a proof of concept, we demonstrate its effect on a set of reverberant mixtures of two speakers. Warping implemented through a deep neural network is trained to estimate mixtures ideally modified towards the instantaneous model in the least squares sense. By applying it as a preprocessing stage, it boosts the BSS performance by up to 6.4 dB of signal-to-interference (SIR) and 4.7 dB of signal-to-distortion (SDR) on average without requiring any modification to the BSS methods.

Název v anglickém jazyce

Warping: Data-driven Mixture Preprocessing to Boost the Performance of Blind Speech Separation

Popis výsledku anglicky

Blind source separation (BSS) can be used to recover speech signals from mixtures recorded by microphones. However, their performances show significant limitations because of the deviations between the instantaneous mixing model and real audio mixtures transformed into the short-time Fourier domain (STFT). This paper presents a data-driven preprocessing technique called mixture warping, which aims to adjust the mixture to obey the instantaneous model as much as possible. As a proof of concept, we demonstrate its effect on a set of reverberant mixtures of two speakers. Warping implemented through a deep neural network is trained to estimate mixtures ideally modified towards the instantaneous model in the least squares sense. By applying it as a preprocessing stage, it boosts the BSS performance by up to 6.4 dB of signal-to-interference (SIR) and 4.7 dB of signal-to-distortion (SDR) on average without requiring any modification to the BSS methods.

Druh

O - Ostatní výsledky

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

<a href="/cs/project/GA25-18485S" target="_blank" >GA25-18485S: Hybridní extrakce signálů: Synergie fyzikálních, informačně-teoretických a datových znalostí</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2025

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů