New IPA phonemizer engine - Goruut.

[neurlang]

Hello I've made a new IPA phonemizer engine, in golang. It is available online at hashtron.cloud or self-hosted under a MIT license. github

What works:

Supported Languages

• English
• Spanish
• German
• Arabic
• Czech
• Slovak

Looking for more languages to add!

Target languages

• IPA
• Espeak
• Antvaset.com

What doesn't work:

• Currently dates, times, or anything with numbers doesn't work (fix planned)
• Slashes, dashes etc or words separated by anything other than spaces.
• More punctuation needs to be trimmed.

Feedback welcomed.

[neurlang]

goruut phonemizer

Supported languages

English, Spanish, Portuguese, German, French, Italian, Arabic, Farsi, Luxembourgish, Dutch, Russian, Swedish, Czech, Slovak

Planned languages

Swahili, Japanese

Looking for more languages to add!

Try it online

https://www.hashtron.cloud/

[komainuone]

Hello,
how does the new IPA phonemizer engine work? Is this documented anywhere? Can we add new languages easily?
Thank you.

[neurlang]

First of all, thank you for the good questions.

The IPA phonemizer searches the word's longest prefix match in the language's grammar. Once a match is identified, the correct pronunciation is selected by a hashtron machine learning model. The process is then repeated for the remaining phones in the word. The performance is on the order of milliseconds for the whole sentence.

Adding a new language has been documented, the dict folder now does have a readme of how to add a new language to goruut. Sharing a dataset with the community is fully optional.

[martinarisk]

I would like to upvote this. Using goruut 0.4.0 and the quality is very good for both phonemizer and reverse phonemizer.

[eginhard]

You can use the code in idiap#88 to use goruut with Coqui.

[neurlang]

Goruut/Pygoruut v0.6.1

• 139 languages support
• fix compilation on 386
• retrained all models for forward / reverse phonemization (breaking change, new minor version)
• new transformer based models for homograph word inference (english)
• Handle numerics in Arabic and English
• working study initiation from null, {} map in language.json
• Add groups of same diacritics to hebrew2
• analysis2 normalization/diacritics sorting
• Fixed -overwrite in homograph scripts
• Split on hanzi/nonhanzi boundary
• don't crash if language dict not found in phonemization steps
• User interface
  • toucan TTS integration
  • word editor
• New Languages:
  • cantonese
  • minnan/taiwanese
  • minnan/hokkien