Granted Patent B2

US 12,738,262 · App. 18/253,673 · Granted Sep 15, 2026

Inventor: Tanzia Haque Tanzi (Huddinge, SE)

Assignee: ASSA ABLOY AB

G10L15/063G10L15/08G10L15/1815G10L2015/088G10L15/16

Abstract

Claims (50)

1 . A method for enabling training a machine-learning, ML, model for trigger-word detection, the method being performed in a training data provider, the method comprising:

receiving sound-based data, the sound-based data being based on sounds captured in a space to be monitored;

determining that the sound-based data corresponds to a trigger word, the determining comprising:

performing speech recognition of the sound-based data;

finding a section of sound-based data that, using the speech recognition, fails to be considered to be the trigger word, but is close to being considered to be the trigger word;

obtaining semantic vector data based on the found section of sound-based data; and

determining that the found section of sound-based data corresponds to the trigger word when a distance, in vector space, between the semantic vector data of the sound-based data and a vector corresponding to the trigger word, is less than a threshold distance;

labelling the sound-based data, including the found section, to correspond to the trigger word; and

providing the labelled sound-based data to train the ML model.

2 . The method according to claim 1 , wherein the sound-based data is in the form of mel-frequency cepstral coefficients, MFCCs.

3 . The method according to claim 1 , further comprising:

discarding sections of the sound-based data that fail to correspond to voice sounds.

4 . The method according to claim 1 , further comprising, after the providing the labelled sound-based data:

discarding all of the sound-based data.

5 . The method according to claim 1 , further comprising:

training a local ML model; and

transmitting at least part of the local ML model to a central location for aggregated learning of a central ML model.

6 . The method according to claim 5 , further comprising:

receiving an updated ML model being based on the central ML model.

7 . The method according to claim 5 , wherein the ML model is the local ML model.

8 . A training data provider for enabling training a machine-learning, ML, model for trigger-word detection, the training data provider comprising:

a processor; and

a memory storing instructions that, when executed by the processor, cause the training data provider to:

receive sound-based data, the sound-based data being based on sounds captured in a space to be monitored;

determine that the sound-based data corresponds to a trigger word, which comprises instructions that, when executed by the processor, cause the training data provider to:

perform speech recognition of the sound-based data;

find a section of sound-based data that, using the speech recognition, fails to be considered to be the trigger word, but is close to being considered to be the trigger word;

obtain semantic vector data based on the found section of sound-based data; and

determine that the found section of sound-based data corresponds to the trigger word when a distance, in vector space, between the semantic vector data of the sound-based data and a vector corresponding to the trigger word, is less than a threshold distance;

label the sound-based data, including the found section, to correspond to the trigger word; and

provide the labelled sound-based data to train the ML model.

9 . The training data provider according to claim 8 , wherein the sound-based data is in the form of mel-frequency cepstral coefficients, MFCCs.

10 . The training data provider according to claim 8 , further comprising instructions that, when executed by the processor, cause the training data provider to:

discard sections of the sound-based data that fail to correspond to voice sounds.

11 . The training data provider according to claim 8 , further comprising instructions that, when executed by the processor, cause the training data provider to, after the instructions to provide the labelled sound-based data, discard all of the sound-based data.

12 . The training data provider according to claim 8 , further comprising instructions that, when executed by the processor, cause the training data provider to:

train a local ML model; and

transmit at least part of the local ML model to a central location for aggregated learning of a central ML model.

13 . The training data provider according to claim 12 , further comprising instructions that, when executed by the processor, cause the training data provider to:

receive an updated ML model being based on the central ML model.

14 . The training data provider according to claim 12 , wherein the ML model is the local ML model.

15 . A non-transitory computer readable storage medium storing a computer program for enabling training a machine-learning, ML, model for trigger-word detection, the computer program comprising computer program code which, when executed on a training data provider, causes the training data provider to:

receive sound-based data, the sound-based data being based on sounds captured in a space to be monitored;

determine that the sound-based data corresponds to a trigger word, wherein the computer program code to determine that the sound-based data corresponds to the trigger word comprises instructions that, when executed by the processor, cause the training data provider to:

perform speech recognition of the sound-based data;

find a section of sound-based data that, using the speech recognition, fails to be considered to be the trigger word, but is close to being considered to be the trigger word;

obtain semantic vector data based on the found section of sound-based data; and

determine that the found section of sound-based data corresponds to the trigger word when a distance, in vector space, between the semantic vector data of the sound-based data and a vector corresponding to the trigger word, is less than a threshold distance;

label the sound-based data, including the found section, to correspond to the trigger word; and

provide the labelled sound-based data to train the ML model.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (33)