Open lippytak opened 10 years ago
We can/should test www.wit.ai for these responses. We could train their algorithm with some sample audio files from the actual EBT numbers where we've had ambiguous/inaccurate transcription and trigger wit.ai (instead of Twilio) as the transcriber for our known-difficult responses. (HT @migurski for the note on Wit)
So far we've seen one inaccurate transcription of "$70" to "$17": https://api.twilio.com/2010-04-01/Accounts/AC3532b1373340d80781d8a0a88f62e484/Recordings/RE08253be0bbfa23bf6d1bc276313c3936.mp3
Are there any technical ways to address these?
See the research issue for details: https://github.com/codeforamerica/balance/issues/16