I need a specialized solution optimized to detect numbers between 1 and 1000 to be used on a smartphone. Best solution would be to have this SDK working offline. Any idea ? I do not find any configuration with Google Speech or Amazon Transcribe to allow "number only"
It is not quite right to strictly expect numbers from people, they usually say many things like "i don't know" or "wait a bit" even if you ask them for numbers. You will harm the experience significantly.
You have to analyze the recognition result intelligently and even if non-number is recognized you have to act accordingly.
To improve accuracy for numbers specifically you can use word hint feature of Google Speech API. Just add digits and other required words as a hint and Google will recognize them much more accurately. Amazon also has this feature they call it "custom vocabulary".
If you want to use an offline API you can certainly try Kaldi. You can adapt Kaldi vocabulary with numbers to improve the accuracy, it will be much better than Google API.
来源:https://stackoverflow.com/questions/55525503/api-or-sdk-to-make-speech-recognition-only-for-numbers-between-1-and-10000