US2018233140A1PendingUtilityA1

Determining speaker changes in audio input

64
Assignee: MICROSOFT TECHNOLOGY LICENSING LLCPriority: Feb 14, 2017Filed: Jul 11, 2017Published: Aug 16, 2018
Est. expiryFeb 14, 2037(~10.6 yrs left)· nominal 20-yr term from priority
Inventors:Kazuhito KoishidaUros Batricevic
G06F 18/29G06F 18/41G06N 3/044H04N 23/90H04N 23/611G06F 18/251G06F 18/28G07C 9/32G07C 9/28G06F 40/211G06F 40/35G06V 40/25G06V 40/103G06V 40/167G06V 40/10G06V 20/52G06V 10/94G06F 2221/2117G10L 15/00G10L 17/06G10L 17/22G06F 1/329G06F 21/00H04W 4/029G01S 13/867H04L 51/02G10L 2015/228G06F 21/32G06F 3/011G06F 1/3231G10L 17/08A61B 5/7475G06T 7/292A61B 5/1118G06T 2207/10024G06F 1/324G06T 2207/10016G06F 2203/0381H04L 67/12G06F 2221/2111G01S 11/14G06F 3/0304G06T 2207/30232H04N 21/44218A61B 5/1113G06T 2207/30201H04W 4/33G06F 1/3206G08B 29/186H04N 21/42203A61B 5/117G10L 17/04H04N 7/188G01S 5/16G10L 17/00A61B 5/05H04R 3/005G01S 13/38H04N 21/231G06F 3/017A61B 5/0205G10L 2015/0635G10L 15/22A61B 5/0507G06F 16/70G06F 3/04842G10L 15/28G08B 13/1427G06N 20/00G06T 2207/20101G10L 15/32G01S 5/18G06T 7/70G01S 13/888G10L 15/19G06N 5/025G06F 3/0482G06T 7/60H04R 1/406H04N 7/181G10L 15/063G10L 15/08G10L 15/02G06T 2207/30196G06N 5/047H04N 21/44222G06F 21/35G06T 7/248G10L 15/26G06F 3/0488G10L 2015/223G06T 7/74G06F 3/167G10L 15/18G10L 15/1815G01S 13/726G10L 2015/088G06T 2207/30204G10L 2015/225G10L 25/51H04L 63/102G10L 15/24G06N 3/0442G06N 3/09G06N 3/0499G10L 15/1822G01S 5/28H04L 67/535G06V 40/173G06V 2201/05G06V 40/23G06V 30/274G06V 40/172G06V 40/166G06V 20/653G06V 20/40H04N 23/11H04N 21/44224Y02D10/00
64
PatentIndex Score
0
Cited by
0
References
0
Claims

Abstract

Intelligent assistant systems, methods and computing devices are disclosed for identifying a speaker change. A method comprises receiving audio input comprising a speech fragment. A first voice model is trained with a first sub-fragment from the speech fragment. A second voice model is trained with a second sub-fragment from the speech fragment. The first sub-fragment is analyzed with the second voice model to yield a first confidence value. The second sub-fragment is analyzed with the first voice model to yield a second confidence value. Based at least on the first and second confidence values, the method determines if a speaker of the first sub-fragment is the speaker of the second sub-fragment.

Claims

exact text as granted — not AI-modified
1 . An intelligent digital assistant system, comprising:
 a logic processor; and   a storage device holding instructions executable by the logic processor to:
 receive audio input comprising a speech fragment; 
 train a first voice model with a first sub-fragment from the speech fragment; 
 train a second voice model with a second sub-fragment from the speech fragment; 
 analyze the first sub-fragment with the second voice model to yield a first confidence value; 
 analyze the second sub-fragment with the first voice model to yield a second confidence value; and 
 based at least on the first confidence value and the second confidence value, determine if a speaker of the first sub-fragment is the speaker of the second sub-fragment. 
   
     
     
         2 . The intelligent digital assistant system of claim of  1 , wherein the instructions are executable to, based at least on determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment, utilize at least the first sub-fragment and the second sub-fragment to determine a user intent of the speaker. 
     
     
         3 . The intelligent digital assistant system of claim of  1 , wherein the instructions are executable to, based at least on determining that the speaker of the first sub-fragment is not the speaker of the second sub-fragment, utilize at least the first sub-fragment and forego utilizing the second sub-fragment to determine a user intent of the speaker of the first sub-fragment. 
     
     
         4 . The intelligent digital assistant system of  claim 1 , wherein the instructions are executable to:
 generate the first voice model and second voice model from a universal background model; and   based at least on determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment, update the universal background model to an updated universal background model using the first sub-fragment and the second sub-fragment.   
     
     
         5 . The intelligent digital assistant system of claim of  4 , wherein the instructions are executable to generate a third voice model from the updated universal background model by training the updated universal background model with another sub-fragment of speech. 
     
     
         6 . The intelligent digital assistant system of  claim 1 , wherein the first sub-fragment and the second sub-fragment have unequal temporal lengths. 
     
     
         7 . The intelligent digital assistant system of  claim 1 , wherein the instructions are executable to, based at least on the first confidence value and the second confidence value exceeding a predetermined threshold, determine that the speaker of the first sub-fragment is the speaker of the second sub-fragment. 
     
     
         8 . The intelligent digital assistant system of  claim 1 , wherein the instructions are executable to, based at least on the first confidence value and the second confidence value being less than or equal to a predetermined threshold, determine that the speaker of the first sub-fragment is not the speaker of the second sub-fragment. 
     
     
         9 . The intelligent digital assistant system of  claim 1 , wherein determining if the speaker of the first sub-fragment is the speaker of the second sub-fragment comprises:
 computing an average of the first confidence value and the second confidence value; and   if the average exceeds a predetermined threshold, then determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment.   
     
     
         10 . At a computing device, a method for identifying a speaker change, the method comprising:
 receiving audio input comprising a speech fragment;   training a first voice model with a first sub-fragment from the speech fragment;   training a second voice model with a second sub-fragment from the speech fragment;   analyzing the first sub-fragment with the second voice model to yield a first confidence value;   analyzing the second sub-fragment with the first voice model to yield a second confidence value; and   based at least on the first confidence value and the second confidence value, determining if a speaker of the first sub-fragment is the speaker of the second sub-fragment.   
     
     
         11 . The method of  claim 10 , further comprising, based at least on determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment, utilizing at least the first sub-fragment and the second sub-fragment to determine a user intent of the speaker. 
     
     
         12 . The method of  claim 10 , further comprising, based at least on determining that the speaker of the first sub-fragment is not the speaker of the second sub-fragment, utilizing at least the first sub-fragment and foregoing utilizing the second sub-fragment to determine a user intent of the speaker of the first sub-fragment. 
     
     
         13 . The method of  claim 10 , further comprising:
 generating the first voice model and second voice model from a universal background model; and   based at least on determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment, updating the universal background model to an updated universal background model using the first sub-fragment and the second sub-fragment.   
     
     
         14 . The method of  claim 13 , further comprising generating a third voice model from the updated universal background model by training the updated universal background model with another sub-fragment of speech. 
     
     
         15 . The method of  claim 10 , wherein the first sub-fragment and the second sub-fragment have unequal temporal lengths. 
     
     
         16 . The method of  claim 10 , further comprising, based at least on the first confidence value and the second confidence value exceeding a predetermined threshold, determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment. 
     
     
         17 . The method of  claim 10 , further comprising, based at least on the first confidence value and the second confidence value being less than or equal to a predetermined threshold, determining that the speaker of the first sub-fragment is not the speaker of the second sub-fragment. 
     
     
         18 . The method of  claim 10 , wherein determining if the speaker of the first sub-fragment is the speaker of the second sub-fragment comprises:
 computing an average of the first confidence value and the second confidence value; and   if the average exceeds a predetermined threshold, then determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment.   
     
     
         19 . A computing device, comprising:
 at least one microphone;   a logic processor; and   a storage device holding instructions executable by the logic processor to:
 via the at least one microphone, receive audio input comprising a speech fragment; 
 generate a first sub-fragment and a second sub-fragment from the speech fragment; 
 train a first voice model with the first sub-fragment; 
 train a second voice model with the second sub-fragment; 
 analyze the first sub-fragment with the second voice model to yield a first confidence value; 
 analyze the second sub-fragment with the first voice model to yield a second confidence value; and 
 based at least on the first confidence value and the second confidence value, determine if a speaker of the first sub-fragment is the speaker of the second sub-fragment. 
   
     
     
         20 . The computing device of  claim 19 , wherein the instructions are executable to, based at least on determining that the speaker of the first sub-fragment is the speaker of the second sub-fragment, utilize at least the first sub-fragment and the second sub-fragment to determine a user intent of the speaker.

Join the waitlist — get patent alerts

Track US2018233140A1 — get alerts on status changes and closely related new filings.

We store only your email — no account needed. See our privacy policy.