Microsoft’s Speech Recognition Tech Achieves Human Parity--Sort Of

[Lucian Armasu]

Microsoft announced its latest major milestone in machine learning-based speech recognition, which now equals the word error rate (WER) of humans. However, the WER doesn't tell the whole story, and machines still have a way to go in understanding humans.

Microsoft’s Speech Recognition Tech Achieves Human Parity--Sort Of : Read more

 

[jaber2]

Just want to know how long until universal translator

 

[kittle]

I wonder how that thing will work when plugged into siri on a noisy subway?

 

[Icepilot]

"... we're living in a time when machines are beginning to truly understand humans and the world around us."
One word too far, truly.

 

[stuartturner34]

“We’ve reached human parity,” said Xuedong Huang, the company’s chief speech scientist. “This is an historic achievement.”

A typo in a quote of a scientist talking about word error ratings. So meta.

 

[alextheblue]

I wonder how that thing will work when plugged into siri on a noisy subway?

That depends on your audio hardware/software more than anything. For example on a PC, it would depend on the type and quality of the microphone / mic array, the sound card, audio drivers, recording software, etc. There's a couple of places where there's opportunities for noise cancellation, depending on the gear and ware used. The result gets handed to this translation software, garbage in garbage out - you have to feed it good audio for it to do it's job. The situation isn't all that different for a smartphone. Unfortunately the iPhone probably wouldn't do the best job compared to a smartphone with a HAAC twin membrane quad-mic array.

 

[jackt]

using super computers or normal pc ?

 

[Kafantaris]

Microsoft's AI driven voice recognition has left all rivals in the dust. Great work by Dr. Xuedong Huang's speech team.

 

[bit_user]

Good job digging into the error rates, Lucian.

using super computers or normal pc ?

This is the question I had. How much compute does it use? It's not a small detail whether this requires a long time on a big GPU, or whether it can run on a smartphone in realtime. If too much compute is required, then this won't be deployed in most real-world uses cases for years.

BTW, humans are still way more energy efficient.

 

[bit_user]

“We’ve reached human parity,” said Xuedong Huang, the company’s chief speech scientist. “This is an historic achievement.”

A typo in a quote of a scientist talking about word error ratings. So meta.

It would be, but where's the error?