Verfahren und Vorrichtung zur Tiefenlernbasierten Sprachverbesserung mit Hoher Abtastrate

Details

Veröffentlichungs-Nr.

EP4756802

Anmeldetag

9. Dezember 2025

Veröffentlichung

10. Juni 2026

Rechtsraum

EP

IPC

G10L21/0232

Abstract

Methods, apparatus, programs, and storage media for enhancing speech signals based on a neural network are provided. The method includes receiving a speech signal and a corresponding reference speech signal. A first set of band features is extracted from the speech signal and a second set of band features is extracted from the reference speech signal, wherein each feature in the first set of band features corresponds to a frequency band of the speech signal and each feature in the second set of band features corresponds to a frequency band of the reference speech signal. The first set of band features is split into a low frequency band part and high frequency band part. Band features corresponding to the high frequency band part are mapped to a reduced number of high frequency band features. The low frequency band part and the reduced number of high frequency band features are combined into a third set of band features. A first mask is determined by inputting the third set of band features into a speech enhancement neural subnetwork. The first mask is split into a low frequency mask part and high frequency mask part. Entries corresponding to the high frequency mask part are mapped to an increased number of entries. The low frequency mask part and the increased number of entries are combined into a second mask. A fourth set of band features is determined based on the second mask and the speech signal. A loss function is evaluated based on the second set of band features and the fourth set of band features. Weights of the speech enhancement neural subnetwork are updated based on a value of the evaluated loss function.

Anmelder

Firma

Dolby Laboratories Licensing Corporation

Land

🇺🇸 USA