High-performance speech recognition with no supervision at all
1–10 of 66 posts
Re: High-performance speech recognition with no supervision at all
#2Quick summary via a Tweet: https://twitter.com/nasim_rahaman/status/1395812572037795847
Re: High-performance speech recognition with no supervision at all
#3Hilarious how the Arabic „hello“ in that picture is completely mangled since they screwed up ltr rendering
(basically instead of hello it says „o l l e h“)
Re: High-performance speech recognition with no supervision at all
#4In the title image, the Arabic word "ahlan" (welcome) is spelled correctly but left to right instead of right to left (should look like this: أهلا)
Re: High-performance speech recognition with no supervision at all
#5Will Facebook license this tech out? I can't be the only one who's noticed Google's speech recognition has gotten significantly less accurate the last couple of years, probably the result of some cost-cutting strategy.
Re: High-performance speech recognition with no supervision at all
#6Facebook's focus on unsupervised machine learning is a huge plus for under-resourced languages. They had a similar article for unsupervised machine translation[1] before, and I can see how it'll open doors for many African languages.
[1] https://engineering.fb.com/2018/08/31/ai-research/unsupervis...
Re: High-performance speech recognition with no supervision at all
#7Will Facebook license this tech out? I can't be the only one who's noticed Google's speech recognition has gotten significantly less accurate the last couple of years, probably the result of some cost-cutting strategy.
They have an open source repo linked from the article here: https://github.com/pytorch/fairseq/tree/master/examples/wav2...
Re: High-performance speech recognition with no supervision at all
#8Will Facebook license this tech out? I can't be the only one who's noticed Google's speech recognition has gotten significantly less accurate the last couple of years, probably the result of some cost-cutting strategy.
I wonder if it's less accurate for specific people, but more accurate generally. In other words speech recognition was more accurate in the beginning on english-speaking men, and maybe now it's better for women, kids, people with accents, other languages, etc...
They train this on examples of speaking and maybe it's more broad now.
Re: High-performance speech recognition with no supervision at all
#9Facebook's focus on unsupervised machine learning is a huge plus for under-resourced languages. They had a similar article for unsupervised machine translation[1] before, and I can see how it'll open doors for many African languages. [1] https://engineering.fb.com/2018/08/31/ai-research/unsupervis...
Not only that but it's also great for the English models as well. Many models are trained using the libre speech dataset with 960h of labeled audio.
What if you trained with 960000 hours instead??
Re: High-performance speech recognition with no supervision at all
#10Hopefully they license it to Tesla so that my car stops misinterpreting 'call Jodie' as 'cool jodi'.