We are working on Call Center conversations. preparing LM for different domains. we have 50 hours recordings in utterance based format, we trainned acoustic model for the same but we are not getting good results as expected like 65%. We are getting somewhere in between 35-40% .
Now we are preparing LM on transcibed data. These conversation might be not accurate in grammar. will this grammar problem affect the LM? like we are getting "I like like" type of phrases. Do we need to prepare LM for orginal transcribed data or we have to preprocess (grammatical corrections) the data before making LM?
For domain specific Speech to Text, do we have to use only domain specific LM? We want to experiment with Domain specific Language Model and General domain Acoustic Model (50 hrs, utterance based). Is it the right approach to deal with. Will it improve our accuracy?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
we have 50 hours recordings in utterance based format
50 hours is too small for modern measures, you need 300-500 hours for good callcenter accuracy.
Do we need to prepare LM for orginal transcribed
Speech is not grammatical, you have to prepare LM on real speech data, not on grammatical data. Subtitles could help to get such data for example.
For domain specific Speech to Text, do we have to use only domain specific LM? We want to experiment with Domain specific Language Model and General domain Acoustic Model (50 hrs, utterance based). Is it the right approach to deal with.
We are working on Call Center conversations. preparing LM for different domains. we have 50 hours recordings in utterance based format, we trainned acoustic model for the same but we are not getting good results as expected like 65%. We are getting somewhere in between 35-40% .
Now we are preparing LM on transcibed data. These conversation might be not accurate in grammar. will this grammar problem affect the LM? like we are getting "I like like" type of phrases. Do we need to prepare LM for orginal transcribed data or we have to preprocess (grammatical corrections) the data before making LM?
For domain specific Speech to Text, do we have to use only domain specific LM? We want to experiment with Domain specific Language Model and General domain Acoustic Model (50 hrs, utterance based). Is it the right approach to deal with. Will it improve our accuracy?
50 hours is too small for modern measures, you need 300-500 hours for good callcenter accuracy.
Speech is not grammatical, you have to prepare LM on real speech data, not on grammatical data. Subtitles could help to get such data for example.
If you have enough text for LM training you can use only domain data. Usually it is not the case and you need to mix with generic langauge model as described in http://cmusphinx.sourceforge.net/wiki/tutoriallmadvanced.
Thanks for the guidance. Will ask query again if we will stuck somewhere.