What the hints are and when to use them
Sometimes automated transcription fails to recognize individual words because they are either not known to the recognizer or because other more common very similar sounding words exist. Here are some examples of such misrecognitions:
- "Then Daniel replied with discretion and discernment to Erioc, the captain of the king's bodyguard" - this should be Arioch which is an ancient Babylonian name
- "The big is kind of puffed up a little bit okay, " - this should be biga which is a type Italian dough
- "You talk about ball movement Max Kellly." - this should be Max Kellerman, a basketball player.
- "That has not always been a case with Kuwait." - this should be Kawhi Leonard, a basketball player.
As you can see these are mistakes in single words and short phrases in otherwise correctly recognized text. You can fix many of such transcription errors by providing correct hints to the recognizer.
Specifying hints
A hint is a word (e.g. Arioch) or a short phase (e.g. Max Kellerman). Phrases need to have underscores in place of spaces, so e.g Max_Kellerman. Phrases may be several words long. Capitalization provided in the hints will be preserved.
It is possible to provide a weight next to the hint, e.g. Max_Kellerman:10. The weight value ranges from 1 to 10 with the default being 5 if the value is not provided. The higher value will make the hint more likely to be used. Using high values of weight increases somewhat the risk of the hint being picked by mistake.
Up to 50 hints can be specified per transcription request.
Using hints from the API
Hints can be specified in transcribe API requests withing the settings.asr, e.g.:
...
,
"settings": {
"asr": {
"speechContext" : "normal",
"noInputTimeout": -1,
"completeTimeout": -1,
"sensitivity" : 0.5,
"hints" : ["Max_Kellerman:10, James_Harden"]
}
}
Using hints from the Web Console
In the Web Console UI you can enter the hints in the dialog form that is used to upload the file for transcription, e.g.:
New Feature - misspellings
Hints as described above will not work very well if the word is foreign word or a word with non obvious pronunciation. An example would be Kawhi Leonard. If the hint was Kawhi_Leonard it is highly likely that it would not work well because the first name is pronounced like "Kawai".
In order to facilitate such words we have just introduced alternative pronunciations (misspellings) within hints. A hint with misspelling(s) looks like this:
Kawhi_Leonard[Kawai_Leonard|Kawai_Lenard]:10
One important constraint with hint misspellings is that the misspelling should not be like another word or phrase that is expected to occur in the transcript. e.g. this would not be a good hint with misspelling: lakhs[lacks]:5
Currently, hint misspellings work in following output:
- results of sync transcription,
- response of polling async transcription with content set to transcript
Comments
0 comments
Please sign in to leave a comment.