Final Report for Google Summer of Code 2019
This is a final report of the work which was done as part of Creation of an online Greek mail dictation system, using Sphinx and personalized acoustic/language model training hosted in https://github.com/eellak/gsoc2019-sphinx and https://snf-870149.vm.okeanos.grnet.gr.
The aim of the project is the implementation of a personalized Greek mail dictation system. The personalization is done both in the language model using the user's emails and in the acoustic model using previous recordings of the user. Also, the ASR output is passed through a post-processing system, where possible errors are corrected based on the adapted language model. By this way, we increase the accuracy of the default Greek model, which is low as a result of the limited amount of open source speech datasets.
Work and Repository
- Tool for extracting and cleaning sent emails of a Gmail user. Code Wiki
- Tool for creating adapted language models through email clustering. Code Wiki
- Tool for correcting ASR output. Code Wiki
- Various tools for preparing and evaluating a speech dataset. Code Wiki
- Simple tool for creating a speech dataset. Code
- API written in Flask. Code Wiki
- Online webpage using Angular 8. Code Wiki
The whole progress of the project was tracked on a daily basis in Projects section.
The project is hosted at https://snf-870149.vm.okeanos.grnet.gr
Note: Till now, we use self signed ssl certificates for both the webpage and the api. As a result, before using the webpage, the user should give permission in both of them by entering https://snf-870149.vm.okeanos.grnet.gr and https://snf-870149.vm.okeanos.grnet.gr:5000 and clicking Advanced and Proceed to url.
Some recommendations for future work can be found here.