Microsoft Hails 'Historic Achievement' in Speech Recognition Technology

MacRumors

macrumors bot
Original poster
Apr 12, 2001
46,344
8,751



Researchers at Microsoft claim to have created a new speech recognition technology that transcribes conversational speech as well as a human does (via The Verge).

The system's word error rate is reportedly 5.9 percent, which is about equal to professional transcribers asked to work on the same recordings, according to Microsoft.

Microsoft researchers from the Speech & Dialog research group (Image: Allison Linn)

"We've reached human parity," said chief speech scientist Xuedong Huang in a statement, calling the milestone "an historic achievement".

To reach the milestone, the team used Microsoft's Computational Network Toolkit, a homegrown system for deep learning that the research team has made available on GitHub via an open source license. The system uses neural network technology that groups similar words together, which allows the models to generalize efficiently from word to word.

The neural networks draw on large amounts of data called training sets to teach the transcribing computers to recognize syntactical patterns in the sounds. Microsoft plans to use the technology in Cortana, its personal voice assistant in Windows and Xbox One, as well as in speech-to-text transcription software.

But the technology still has a long way to go before it can claim to master meaning (semantics) and contextual awareness - key characteristics of everyday language use that need to be grasped for Siri-like personal assistants to process requests and act upon them in a helpful way.

"We are moving away from a world where people must understand computers to a world in which computers must understand us," said Harry Shum, who heads the Microsoft AI Research group. However it will be a long time before computers can understand the real meaning of what's being said, he cautioned. "True artificial intelligence is still on the distant horizon."

Article Link: Microsoft Hails 'Historic Achievement' in Speech Recognition Technology
 

keysofanxiety

macrumors G3
Nov 23, 2011
9,534
24,709
Properly exciting times.

I remember when I was but a sprog, wide-eyed in wonder, sitting on my Dad's lap as we watched Next Gen. I don't think anybody back then would have imagined technology to be as advanced as it is now.
 
  • Like
Reactions: djlythium

Everlast66

macrumors newbie
Oct 27, 2014
10
30
Great, so now Windows 10 is even better at harvesting your personal conversations and relaying them back home to Microsoft.

When should we expect the update to be pushed, with these Win 10 machines not giving the option to opt out of updates?
 

keysofanxiety

macrumors G3
Nov 23, 2011
9,534
24,709
When should we expect the update to be pushed, with these Win 10 machines not giving the option to opt out of updates?
Have you enabled the Defer feature updates checkbox (Start> Settings> Update & Security > Advanced options)? It won't prevent security updates but does the trick for most things. :)
 
  • Like
Reactions: peterdevries

JeffyTheQuik

macrumors 68020
Aug 27, 2014
2,468
2,338
Charleston, SC and Everett, WA
GOOD!

Now, install them at every singe drive through in the USA.
[doublepost=1476878498][/doublepost]
Great, so now Windows 10 is even better at harvesting your personal conversations and relaying them back home to Microsoft.

When should we expect the update to be pushed, with these Win 10 machines not giving the option to opt out of updates?
I'm still trying to figure out why Siri has to have everything sent to Apple for decoding. MS Voice Command could do that 10 years ago with pretty good accuracy on a 256MB Windows Phone.
The methodology was:
Speak...
If it was on the phone, like contacts, it would just get the info for you, or call that person (if commanded to).
If it needed more help, it would go out to the web and get it for you.

Nothing sent to Microsoft, as it would work even if there wasn't Internet connectivity.,
 
  • Like
Reactions: drumcat and adib

Defthand

macrumors 65816
Sep 1, 2010
1,300
1,655
Meanwhile, Apple's voicemail transcriptions are embarrassing. I told them during the beta testing that they should withhold the feature until it can be improved. If people actually depended on it, it would be worse than the infamous Maps fiasco.
 

2457282

Suspended
Dec 6, 2012
3,327
3,006
Actually is "up to" 99%. It can be as low as "zero".

The quote I see under Accuracy says --
Accuracy
Control your computer by voice with speed and accuracy
Dragon speech recognition software is better than ever. Talk and your words appear on the screen. Say commands and your computer obeys. Dragon is 3x faster than typing and it's 99% accurate. Master Dragon right out of the box and start experiencing big productivity gains immediately.
 

CreatorCode

macrumors regular
Apr 15, 2015
139
118
US
The quote I see under Accuracy says --

[...]
DNS is very accurate if you speak clearly and directly, period. Contrary to what its name implies, comma, you cannot just speak naturally, period. You have to dictate specifically to it, period.

New paragraph.

The Microsoft experiment, comma, allegedly, comma, transcribes ordinary recorded speech and dialog without any additional effort on the part of the speaker, period.
 

Benjamin Frost

Suspended
May 9, 2015
2,405
4,924
London, England
DNS is very accurate if you speak clearly and directly, period. Contrary to what its name implies, comma, you cannot just speak naturally, period. You have to dictate specifically to it, period.

New paragraph.

The Microsoft experiment, comma, allegedly, comma, transcribes ordinary recorded speech and dialog without any additional effort on the part of the speaker, period.
It sounds impressive, but I very much doubt that it would be accurate at transcribing punctuation. Semi-colons, colons? These are tricky things for humans to get right, let alone a computer. Or: ... and dashes/hyphens and /. Also brackets. Exclamation marks? That last sentence would be extremely hard for a computer; if I didn't raise my voice at the end of the sentence to denote a question, how would it know to insert a question mark?
 

Rhonindk

macrumors 68040
Oct 3, 2014
3,876
7,515
watching the birth of the Dem WTH Party
I don't believe a word of this.

Human transcribers must vary widely in ability; secondly, there are many provisos given, which effectively means that it is still much lower accuracy than a human.
I work with a regional (USA) team - coast to coast, border to border, including territories. There are very frequent times, due in part to grammar, accent, and colloquialisms, that understanding is at a minimum.
Would love to know from where that percentage arose.

Then add in other languages and dialects.
 

merkinmuffley

macrumors 6502a
Dec 3, 2010
615
581
If this report is correct, this is a big step forward in voice recognition. I'm curious how much processing is behind it.