Re:Show HN: The worst voicemail transcription ever
by 3Scribe on Monday August 05, 2019 @04:19PM
(#20618467)
That is true, made more so when doing phone calls (8kHz, Mono recordings) with a variety of accents. I've scored a live transcription system at 94% accurate over 250 samples of voicemail recordings from UK accents, with various automated services coming in at the following rates:
AWS 0.871603 Temi 0.871063 Google 0.867890 Speechmatics 0.858851 OtterAI 0.855769 Microsoft 0.854067 AssemblyAI 0.795696 IBM 0.777215 Remeeting 0.771764
I'm building a new transcription startup called 3Scribe (shameless self-promotion) that's scoring at just below 91%. Hoping to squeeze out another little bit of accuracy but at the moment struggling with React-Bootstrap, trying to build a beta MVP.
Think of how many accents there exist "in the wild". About the same as the number of people in the world.