Talrómur TTS models 22.10

  Authors

Örnólfsson, Gunnar Thor

  Item identifier

http://hdl.handle.net/20.500.12537/294

 Date issued

2022-04-29

 Type

toolService

 Language(s)

Icelandic

 Description

This release contains TTS models for all 8 voices from Talrómur [1], intended for use with the ESPNet toolkit [2]. Please refer to ESPNet's documentation to resolve any uncertainty. The models were produced using the recipes developed in and published as part of the ESPNet toolkit . The recipes are included in the 202207 version of the toolkit: [3]. Usage of the recipe is described in its README file: [4] . The models are trained using fully phonemized, stress-annotated inputs including syllable boundary markers ('.') and word boundary markers (','). These inputs can be obtained using the ice-g2p [5] Transcriber module with the described parameters dialed in. Both a Tacotron 2 model and a Fastspeech 2 model are included for each voice. The Tacotron model is slower but higher quality, whereas the Fastspeech model is faster and offers control over some prosodic features; speaking rate, pitch and energy. The models are also published on HuggingFace and instructions within the README in each zip file suppose that you're fetching them from there. When downloading from here, you can simply unzip the files and provide paths to the model's config.yaml and the model itself, in the respective parameters to espnet2.bin.tts_inference.Text2Speech. e.g. text2speech = Text2Speech( train_config="/path/to/config.yml", model_file="/path/to/model.pth",´ vocoder_config="/path/to/vocoder/config.yml", vocoder_file="/path/to/vocoder/model.pkl" ) Here you can see a demo of how ESPNet models can be used in inference: [6] // Þessi útgáfa inniheldur talgervingarlíkön fyrir allar 8 raddir Talróms [1], ætluð til notkunar innan ESPNet verkfærakistunnar [2]. Vinsamlegast skoðið skjölun ESPNet til að ráða fram úr óvissu með notkun. Líkönin voru þróuð með forskriftum sem voru gefnar út innan ESPNet. Forskriftir fyrir þessi líkön eru innifaldar í 202207 útgáfu verkfærakistunnar: [3]. Notkun forskriftarinnar er lýst í README skrá hennar: [4]. Inntak líkananna við þjálfun er hljóðritaður, áherslugreindur texti sem inniheldur tóka fyrir atkvæðaskil ('.') og orðaskil (','). Gögn á því formi má fá úr ice-g2p [5] Transcriber módúlnum með breytur stilltar eins og hér er lýst. Bæði Tacotron 2 og Fastspeech 2 eru gefin fyrir hverja rödd. Tacotron líkanið er hægvirkara en gefur meiri gæði, en Fastspeech líkanið er hraðvirkar og býður upp á stjórn yfir stökum prósódískum atriðum; leshraða, tónhæð og orku. Líkönin eru einnig útgefin á HuggingFace og leiðbeiningar í README skrám innan hverrar zip skrár hér reikna með því að líkönin séu sótt þaðan. Ef þau eru sótt héðan má einfaldlega afþjappa möppunum og gefa slóðir á config skrá og líkan fyrir það líkan sem nota á sem parametra á espnet2.bin.tts_inference.Text2Speech hlutnum. dæmi: text2speech = Text2Speech( train_config="/path/to/config.yml", model_file="/path/to/model.pth",´ vocoder_config="/path/to/vocoder/config.yml", vocoder_file="/path/to/vocoder/model.pkl" ) Dæmi um notkun á ESPNet líkani má sjá hér: [6] Links/Hlekkir [1] https://repository.clarin.is/repository/xmlui/handle/20.500.12537/104 [2] https://github.com/espnet/espnet [3] https://github.com/espnet/espnet/releases/tag/v.202207 [4] https://github.com/espnet/espnet/blob/96bd74641ceb463096067223d0734f70bddd8def/egs2/talromur/tts1/README.md [5] https://pypi.org/project/ice-g2p/ [6] https://colab.research.google.com/github/espnet/notebook/blob/master/espnet2_tts_realtime_demo.ipynb

 Publisher

Reykjavík University

 Acknowledgement

Ministry of Education, Science and Culture

Project code: T13 - Parametric synthesis for Icelandic

Project name: Language Technology for Icelandic 2019-2023

 Subject(s)

TTS Model

 Collection(s)

Clarin IS

Sýna fulla færslu atriðis