Talrómur 3 (v1.0)

  Authors

Örnólfsson, Gunnar Thor ; Sigurgeirsson, Atli Þór ; Nikulásdóttir, Anna Björk and Schnell, Daniel

  Item identifier

http://hdl.handle.net/20.500.12537/387

 Project URL

https://github.com/icelandic-lt/emospeech-scripts

 Date issued

2025-08-25

 Type

audio, corpus

 Size

24 hours, 21518 utterances, 6.4 gb

 Language(s)

Icelandic

 Description

ENGLISH: Talrómur 3 is a public domain speech corpus for Text-To-Speech (TTS) research and development in the domain of emotional speech synthesis. The corpus consists of 21,528 audio clips of ten different speakers reading short sentences. The audio was recorded in 2024 by Reykjavík University and Grammatek as part of The Icelandic Language Technology Program. Each speaker read a script of 340 utterances in 6 different styles (neutral, happy, sad, angry, surprised and helpful). Additionally, each speaker read all letters in the icelandic alphabet and numbers from 0 to 20 both in isolation and as part of an enumeration. The audio is recorded at 48 kHz sample rate and 24 bit depth. Each audio file is stored in .flac format. In addition to the audio recordings, this corpus includes Voice Activity Detection (VAD) values for each utterance, obtained using Silero VAD, as well as Montreal Forced Aligner (MFA) phoneme-level alignments. These additional data can be used to trim the audio or during model training. ICELANDIC: Talrómur 3 er opið gagnasafn ætlað til rannsókna og þróunar á talgervlum fyrir íslensku á sviði tilfinningaríkrar talgervingar. Gagnasafnið inniheldur 21.528 stuttar upptökur frá tíu mismunandi röddum. Hákólinn í Reykjavík og Grammatek stóðu að söfnun gagnanna árið 2024 á vegum Máltækniáætlunar fyrir íslensku. Fyrir hverja rödd hafa 340 segðir verið lesnar inn í 6 mismunandi tilbrigðum (hlutlaus, gleði, depurð, reiði, undrun og hjálpsemi). Þar að auki voru lesnir inn allir stafir stafrófsins og tölur frá 0 upp í 20 á tvo mismunandi vegu: í einangrun og líkt og í upptalningu. Hljóðskrárnar voru teknar upp í 48 kHz og með 24 bita dýpt. Skrárnar eru geymdar á .flac sniði Auk hljóðskránna inniheldur þessi útgáfa raddvirknimerkingar (Voice Activity Detection values) fyrir hverja segð, fengnar með Silero VAD, auk Montreal Forced Aligner (MFA) samraðana á hljóðanaskala. Þessi viðbótargögn geta nýst til að snyrta hljóðskrárnar eða við þjálfun líkana.

 Publisher

Reykjavík University

 Acknowledgement

Ministry of Culture and Business Affairs

Project code: Recordings of emotional speech (T14)

Project name: Language Technology for Icelandic

 Subject(s)

TTS Emotional Speech Speech Alignments VAD

 Collection(s)

Clarin IS

 Other versions

List all versions  

Show full item record