Posted: Thu May 17, 2007 4:16 pm Post subject: [asterisk-speech-rec] How to enable N-Best decoding? CRM:001
Quote:
Also, I've been finding that a lot of times the Speech engine
takes a long time (on the order of 1 or a several seconds) to
recognize speech.
There are a few things that affect the length of time before you'll get
results:
1. The end of speech delay is probably the most common. This is the
amount of time that we have to detect silence before the Speech Engine
begins doing a decode. If this is too low, it causes problems as callers
get off before they've really finished speaking (it's especially
important to set this high for utterances where you expect a lot of
pauses between words, such as when collecting large strings of digits).
You can adjust this by changing the vad_eos_delay in your lumenvox.conf
file; by default it's set to 2000 ms. For really short yes/no types of
prompts, you can set this value pretty low (800ms or less).
2. Extra background noise that the Engine is mistaking for speech could
also falsely cause it to not begin doing a decode, though this is a
pretty rare thing.
3. The Engine may also be taking a while to perform the actual decode.
Speech recognition is a pretty resource intensive process, and so
overworked hardware can sometimes cause this. The amount of time to
decode goes up as the length of the utterance and size of the grammars
go up as well (increasing the grammar size increases the Engine's search
space). If you use our Speech Tuner, you can see how long a decode
actually took by loading up some response files and browsing
interactions with the Call Browser, and clicking the "View Details"
button for a given decode. One of the details will be the decode time.
So I would recommend first seeing how long decodes are taking and then
try changing vad_eos_delay. Let us know how it goes.
Posted: Fri May 18, 2007 3:44 pm Post subject: [asterisk-speech-rec] How to enable N-Best decoding? CRM:001
Hi Stephen:
Thanks for the tip. Lowering the value of
vad_eos_delay did help. A couple of other questions:
1. Is there a way to change the value of vad_eos_delay
on-the-fly in my dialplan (instead of the
lumenvox.conf file)? As you have noted, in a given
speech application the value of vad_eos_delay may need
to change depending on the particular type of question
being asked.
2. So far I have not worked with grammars with more
than a hundred rules. How large a grammar can one have
before the speech engine computation times become
noticeable to the user (i.e in your experience, can
the engine handle a grammar file with hundreds or
thousands of grammar rules in reasonable time?)
Warm regards,
Adrian
--- Stephen Keller <StephenKeller@LumenVox.com> wrote:
Quote:
> Also, I've been finding that a lot of times the
Speech engine
> takes a long time (on the order of 1 or a several
seconds) to
> recognize speech.
There are a few things that affect the length of
time before you'll get
results:
1. The end of speech delay is probably the most
common. This is the
amount of time that we have to detect silence before
the Speech Engine
begins doing a decode. If this is too low, it causes
problems as callers
get off before they've really finished speaking
(it's especially
important to set this high for utterances where you
expect a lot of
pauses between words, such as when collecting large
strings of digits).
You can adjust this by changing the vad_eos_delay in
your lumenvox.conf
file; by default it's set to 2000 ms. For really
short yes/no types of
prompts, you can set this value pretty low (800ms or
less).
2. Extra background noise that the Engine is
mistaking for speech could
also falsely cause it to not begin doing a decode,
though this is a
pretty rare thing.
3. The Engine may also be taking a while to perform
the actual decode.
Speech recognition is a pretty resource intensive
process, and so
overworked hardware can sometimes cause this. The
amount of time to
decode goes up as the length of the utterance and
size of the grammars
go up as well (increasing the grammar size increases
the Engine's search
space). If you use our Speech Tuner, you can see how
long a decode
actually took by loading up some response files and
browsing
interactions with the Call Browser, and clicking the
"View Details"
button for a given decode. One of the details will
be the decode time.
So I would recommend first seeing how long decodes
are taking and then
try changing vad_eos_delay. Let us know how it goes.
Posted: Fri May 18, 2007 7:50 pm Post subject: [asterisk-speech-rec] How to enable N-Best decoding? CRM:001
Quote:
1. Is there a way to change the value of vad_eos_delay
on-the-fly in my dialplan (instead of the lumenvox.conf
file)?
Yes. Use Set(SPEECH_ENGINE(vad_eos_delay)=100) in your dial plan;
replace 100 with the value you want. Any value in lumenvox.conf can be
altered on-the-fly like this.
Quote:
2. So far I have not worked with grammars with more than a
hundred rules. How large a grammar can one have before the
speech engine computation times become noticeable to the user
(i.e in your experience, can the engine handle a grammar file
with hundreds or thousands of grammar rules in reasonable time?)
This really depends on your hardware. An old P3 with 512 MB of RAM is
going to bog down much sooner than a new dual-core system with a few
gigs of memory.
But generally we find that it becomes noticeable when you are
recognizing more than 2,000 individual words at a single time. Note that
this is different from the total size of the grammar. A grammar may have
10,000 unique words to be recognized but that entire space may not be
searched all at once depending on how the grammar is structured.
Also there is not a perfect correlation between grammar size and number
of words to be recognized; we have seen grammars that contain a few
thousand items to be recognized but are around a hundred thousand lines
long. These take longer for us to compile initially, but once we've
compiled them decoding utterances is still quite fast.
You can post new topics in this forum You can reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum