Dnn-Basierte Sprachverbesserung mit Ultraniedrigem Speicher und Geringer Komplexität für Dsps

Details

Veröffentlichungs-Nr.

EP4787372

Anmeldetag

30. Januar 2025

Veröffentlichung

5. August 2026

Rechtsraum

EP

IPC

G10L25/30, G10L25/18

Abstract

There is disclosed an apparatus (10) for processing an input audio signal (20), comprising: a feature extractor (100) for extracting a set of features (332) from values (102) of the input audio signal (20); a first neural network, NN, processor (340), to process the set of features (332) through a first NN, to obtain a version of the set of features (342) as a first tensor having a frequency dimension with spectral values; a feature segmenter (345), to segment the first tensor (342) into a plurality of feature subbands (346); a second NN processor (370), to process each of the plurality of feature subbands (346) through a common second NN, to obtain a set of result subbands (371); a result composer (375), to compose the result subbands (371) onto one single second tensor (352) having a frequency dimension.

Anmelder

Firma

Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V.

Land

🇩🇪 Deutschland

Fachgebiete