I have problems with some pronunciations for abbreviations.
For example for NSERC the pronunciation produced is N S AH R K but it should
be EH N S AH R K or something like that.
So, my question is how does sphinx handle abbreviations? How can I fix this
problem?
Any suggestion or solutions?
Bests,
Amin
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I've used lmtool (online one) for producing N S AH R K
Here is the problem:
There are abbreviations that "speaker" may use in his/her "speech", such as
NSERC. In order to be able to recognize those words (NSERC), they should be
paired with a pronunciation (EH N S EH R K) n the .dic file.
Is there any tool for producing the correct pronunciations for abbreviations?
Another abbreviation that one can think of is SFU. which should be EH S EH F
YU
Am I clear enough?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Online lmtool which uses pronounce is very limited. There must be a
preprocessing step before it or actually it's way better to use more advanced
tools. Consider
Karl IV
1000 miles
NBC
Those all are words which require tokenization. In practice this job is done
by TTS engine, for example FreeTTS or OpenMARY which has all required code to
do that properly. We are planning to integrate engine by the end of the summer
but this is just a plan. I agree this is a much needed functionality that we
lack now.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
They have a set of rules, like state machine, which they use to extract the
pronunciation of each word, but the set of rules seems to be not
complete/accurate enough.
I looked at Festival as well, same problem there.
The problem seems to be those being somehow out of date (may be not the
engine, but the rule-sets)
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Since the problem of abbreviation detection is sort of external to ASR, it
might help to ask somewhere in NLP community for the help. Maybe some more or
less accurate toolkit for this task is available.
I'll be also interested to find that out, if find have a link please share it
here.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hey all,
I have problems with some pronunciations for abbreviations.
For example for NSERC the pronunciation produced is N S AH R K but it should
be EH N S AH R K or something like that.
So, my question is how does sphinx handle abbreviations? How can I fix this
problem?
Any suggestion or solutions?
Bests,
Amin
Hello
Sorry, what exactly do you use to produce pronunciation for this word? A web
service?
Not sure what do you mean by "handle". You can add abbreviation to the
dictionary as any other word
Is there a problem to solve?
Maybe I wasn't clear enough.
I've used lmtool (online one) for producing N S AH R K
Here is the problem:
There are abbreviations that "speaker" may use in his/her "speech", such as
NSERC. In order to be able to recognize those words (NSERC), they should be
paired with a pronunciation (EH N S EH R K) n the .dic file.
Is there any tool for producing the correct pronunciations for abbreviations?
Another abbreviation that one can think of is SFU. which should be EH S EH F
YU
Am I clear enough?
Hello
Online lmtool which uses pronounce is very limited. There must be a
preprocessing step before it or actually it's way better to use more advanced
tools. Consider
Karl IV
1000 miles
NBC
Those all are words which require tokenization. In practice this job is done
by TTS engine, for example FreeTTS or OpenMARY which has all required code to
do that properly. We are planning to integrate engine by the end of the summer
but this is just a plan. I agree this is a much needed functionality that we
lack now.
Actually I have tried FreeTTS engine for this.
They have a set of rules, like state machine, which they use to extract the
pronunciation of each word, but the set of rules seems to be not
complete/accurate enough.
I looked at Festival as well, same problem there.
The problem seems to be those being somehow out of date (may be not the
engine, but the rule-sets)
That's true. The problem is complex and there is even non-trivial research in
this area, for example, for medical texts
http://www.lrec-conf.org/proceedings/lrec2010/pdf/737_Paper.pdf
Since the problem of abbreviation detection is sort of external to ASR, it
might help to ask somewhere in NLP community for the help. Maybe some more or
less accurate toolkit for this task is available.
I'll be also interested to find that out, if find have a link please share it
here.