But with just its initial 320 hours of data, Te Hiku was able to build a speech-to-text engine with an initial word error rate of 14 per cent, according to Mahelona, a Native Hawaiian who’s been working at Te Hiku for seven years. For reference, Google’s ASR achieves a word error rate of 6.7 per cent with a 12,500-hour data set, according to one 2018 conference abstract. “The fact that they are getting word error rates that low for just over 300 hours, for a language that basically didn’t have speech recognition before, that"s very impressive,” Davis says.
|