v2.0.0 with G2P tools and JSGF support

Adapt-LM v2.0.0

New Features

The Adapt-LM tools have been updated to support the popular grapheme-2-phoneme tools Sequitur G2P and Phonetisaurus G2P. Both tools come with G2P models for 'en' and 'de' that have been trained on the existing Zamia Kaldi dictionaries and can be used to automatically generate out-of-vocabulary words during adaptation.

In addition you can now train JSGF language models with fixed grammar. Choose between JSGF and N-gram models depending on your use-case or simply train both, deploy them on your SEPIA STT Server and hot-swap during runtime 🙂.

All install, build and adapt scripts have been extended and optimized to support all the new features.
Please note that the API of the '3-adapt.sh' script has been adjusted to give access to all the new options.

As usual all models trained with these scripts can be copied over to your SEPIA STT Server and used right away 🤖 💬

Models and build files

Please check the v1.0.0 release if you need to manually download models and binaries

Release of new v1.0.0 for SEPIA STT v2

This is the first release of the new version that supports SEPIA STT Server v2:
https://github.com/SEPIA-Framework/sepia-stt-server

It has been simplified and many dependencies have been removed or replaced by much smaller alternatives.
This version works on x86 64Bit, ARM 32Bit and ARM 64Bit platforms and supports Raspberry Pi 4 :-)

Pre-built Kaldi and KenLM files are copied from: https://github.com/synesthesiam/prebuilt-apps/releases
Many thanks to synesthesiam!

[EDIT 2023.03.20]

• Added wheels for Sequitur G2P
• Added G2P models (up to 4-gram) for Zamia 'en' and 'de' phoneme sets

[EDIT 2023.03.23]

• Added wheels for Phonetisaurus G2P
• Added Phonetisaurus models for Zamia 'en' and 'de' phoneme sets