In this tutorial, we will charge the custom speech to text service, cris.ai, with one of the hardest task available : understanding the rpaper emimem, in the song "rap god". The motivation (beyond "because I can") is to showcase the limits of the service, and therefore it application domain.
To create a custom Speech-to-text model, you need three things : an audio dataset, a text dataset, and a cris.ai account. You then can create an accoustic model linking the audio to the text within cris.ai.
It has a few format requirements : you need a .wav format, in mono, with a debit of 16KB/s. Each sentence requires to be in its own file. It's the most boring part in the job.
It is a .txt with the following architecture :
When you have both audio and text, you can create an acoustic model which will adapt to the context. Depending on your needs, you might need to get a subscription in the Azure portal -however everything I did for testing was free.
open a tunnel
ssh -D 127.0.0.1:8034 $pyp17vmname.westeurope.cloudapp.azure.comthen have your browser using the ssh tunnel as a proxy. Here is how this looks in Firefox (which as proxy settings independant of the OS):
