Question regarding model adaptation

alphacep / vosk-api

Offline speech recognition API for Android, iOS, Raspberry Pi and servers with Python, Java, C# and Node

Apache License 2.0

7.37k stars 1.04k forks source link

Question regarding model adaptation #1461

Open jumper047 opened 8 months ago

jumper047 commented 8 months ago

I'm trying to add improve recognition of the specific phrases (I'm using vosk as recognizer for "smart speaker"). I followed the guide - https://alphacephei.com/vosk/adaptation , generated new Gr.fst and it works, and resulting model recognizes only phrases I added. But what if I want to update phrases model can recognize, and not replace it? Is it possible to find phrases which were used to generate original Gr.fst?

nshmyrev commented 7 months ago

Default graph rebuilding process adds phrases, not replaces them. You probably modified something on the way.

jumper047 commented 7 months ago

Started rebuilding process from scratch and got same results. Also old and new Gr.fst drastically differs in size:

5969333 Gr.new.fst
24013795 Gr.fst

Commands I executed:

fstsymbols --save_osymbols=words.txt graph/Gr.fst > /dev/null
farcompilestrings --fst_type=compact --symbols=words.txt --keep_symbols text.txt | ngramcount | ngrammake | fstconvert --fst_type=ngram > Gr.new.fst

text.txt contains about 20 lines with some phrases