Initial Training for Speech Recognition Software
|Submitter Email:||click here to access email|
My company is developing software for doctors to dictate reports. Our
software relies on a speech recognition engine that is trained to
recognize words. To improve on the current model, I am redesigning
the initial speech training component. As I would like to develop
effective material, I am looking for insight on the following questions:
What are the requirements for initial speech training text (to be read by
the user of speech recognition in order to initially train the speech
engine, and start working with a satisfactory level of recognition)?
Does it have to include all possible phonemes of a language?
Do they have to repeat certain number of times?
If the full phonemic inventory is not required, what would be necessary
for a language such as English?
What other requirements should I consider for such a text?
While I have attempted to do some research on this subject, I have had
trouble finding adequate guidelines for this, and speech corpora have
not really been searchable for texts like this. I have an exemplary text
of SpeechMagic software (provided by Nuance), but I would be grateful
for any additional examples people could provide. Any other resources
or guidelines for speech recognition development would also be greatly
Sums main page