Online Speaker Diarization Using Optimized SE-ResNet Architecture

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F46747885%3A24220%2F23%3A00011570" target="_blank" >RIV/46747885:24220/23:00011570 - isvavai.cz</a>

Výsledek na webu

<a href="https://dl.acm.org/doi/10.1007/978-3-031-40498-6_16" target="_blank" >https://dl.acm.org/doi/10.1007/978-3-031-40498-6_16</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-3-031-40498-6_16" target="_blank" >10.1007/978-3-031-40498-6_16</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Online Speaker Diarization Using Optimized SE-ResNet Architecture

Popis výsledku v původním jazyce

A new approach to speaker diarization (SD) suitable for real-time processing of streamed data is presented in this work. It utilizes a modified residual network with squeeze-and-excitation blocks (SE-ResNet-34) for extraction of speaker embeddings. These speaker embeddings are calculated in an optimized way by using cached buffers and are subsequently used for voice activity detection (VAD) as well as for block-online k-means clustering with a look-ahead mechanism. All these processing steps are first evaluated separately on a development set compiled from recordings of Czech broadcast programs. The whole scheme is then compared to an offline reference approach on various speech databases that are publicly available and include data in various languages. On this data, our method yields results similar to the reference system while operating on a CPU with a low real-time factor (RTF) below 0.1 and a latency of around 5.5 s.

Název v anglickém jazyce

Online Speaker Diarization Using Optimized SE-ResNet Architecture

Popis výsledku anglicky

A new approach to speaker diarization (SD) suitable for real-time processing of streamed data is presented in this work. It utilizes a modified residual network with squeeze-and-excitation blocks (SE-ResNet-34) for extraction of speaker embeddings. These speaker embeddings are calculated in an optimized way by using cached buffers and are subsequently used for voice activity detection (VAD) as well as for block-online k-means clustering with a look-ahead mechanism. All these processing steps are first evaluated separately on a development set compiled from recordings of Czech broadcast programs. The whole scheme is then compared to an offline reference approach on various speech databases that are publicly available and include data in various languages. On this data, our method yields results similar to the reference system while operating on a CPU with a low real-time factor (RTF) below 0.1 and a latency of around 5.5 s.

Druh

D - Stať ve sborníku

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/TO01000027" target="_blank" >TO01000027: NORDTRANS - Technologie pro automatický přepis řeči ve vybraných severských jazycích</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)<br>S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2023

Kód důvěrnosti údajů

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

Název statě ve sborníku

Lecture Notes in Computer Science

ISBN

978-303140497-9

ISSN

03029743

e-ISSN

—

Počet stran výsledku

12

Strana od-do

176-187

Název nakladatele

Springer

Místo vydání

Německo

Místo konání akce

Plzeň, ČR

Datum konání akce

1. 1. 2023

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—