Methods of and electronic devices for identifying an end-of-utterance moment in a digital audio signal

An electronic device for identifying a user utterance from a digital audio signal, the digital audio signal containing an audio recording of the user utterance, the electronic device being configured to:

• acquire a set of features for a respective segment of the digital audio signal, each set of features comprising at least acoustic-type features derivable from a respective segment of the digital audio signal, the segments of the digital audio signal being associated with respective time intervals of a pre-determined time length;
• acquire an indication of an end-of-utterance moment in time in the digital audio signal corresponding to a given moment in time after which the user utterance has ended;
• determine an adjusted end-of-utterance moment in time by adding a pre-determined temporal offset to the end-of-utterance moment in time;
• determine labels for respective sets of features based on the adjusted end-of-utterance moment in time and the time intervals of the corresponding segments of the digital audio signal, a given label being indicative of whether the user utterance has ended during the corresponding segment of the digital audio signal associated with the respective set of features;
• use the sets of features and the respective labels for training a Neural Network (NN) to predict during which segment of the digital audio signal the user utterance has ended,
• for a given training iteration of the NN, the electronic device being configured to:
• generate a plurality of individual penalty values, a given individual penalty value being associated with a respective segment of the digital audio signal, the digital audio signal including segments that precede the end-of-utterance moment in time, the given individual penalty value being indicative of a prediction error made by the NN for the respective segment of the digital audio signal;
• generate a combined penalty value for the given training iteration of the NN, the combined penalty value being a combination of the plurality of individual penalty values;
• use the combined penalty value to train the NN,
• wherein the electronic device is further configured to, during an in-use phase of the NN:
• acquire, by the electronic device, at least some of an in-use digital audio signal;
• determine, by the electronic device, a first set of features for a first segment of the in-use digital audio signal, the first segment including a most recently acquired portion of the in-use digital audio signal, the in-use digital audio signal being recorded in real-time during an in-use utterance, the first set of features comprising at least acoustic-type features derivable from the first segment of the digital audio signal;
• use, by the electronic device, the NN to determine based on the first set of features a first value indicative of a probability of the in-use user utterance having ended during the first segment of the in-use digital audio signal;
• responsive to the first value being above a pre-determined threshold, determine, by the electronic device, that the in-use user utterance ended during the first segment of the in-use digital audio signal,
• wherein the electronic device is further configured to, responsive to the first value being below the pre-determined threshold:
• determine a second set of features for a second segment of the in-use digital audio signal, the second segment being sequential to the first segment in the in-use digital audio signal, the second set of features comprising at least acoustic-type features derivable from the second segment of the digital audio signal;
• use, by the electronic device, the NN to determine based on the first set of features and the second set of features a second value indicative of a probability of the in-use user utterance having ended during the second segment of the in-use digital audio signal;
• and
• responsive to the second value being above the pre-determined threshold, determine, by the electronic device, that the in-use user utterance ended during the second segment of the in-use digital audio signal.

A method of training a Neural Network (NN) for identifying a user utterance from a digital audio signal, the digital audio signal containing an audio recording of the user utterance, the method executable by an electronic device, the method comprising:

• acquiring, by the electronic device, a set of features for a respective segment of the digital audio signal,
• a given set of features comprising at least acoustic-type features derivable from a respective segment of the digital audio signal,
• the segments of the digital audio signal being associated with respective time intervals;
• determining, by the electronic device, an adjusted end-of-utterance moment in time by adding a pre-determined temporal offset to an end-of-utterance moment in time in the digital audio signal,
• the end-of-utterance moment in time corresponding to a given moment in time after which the user utterance has ended;
• determining, by the electronic device, labels for respective sets of features based on the adjusted end-of-utterance moment in time and the time intervals of the corresponding segments of the digital audio signal,
• a given label being indicative of whether the user utterance has ended during the corresponding segment of the digital audio signal associated with the respective set of features;
• training, by the electronic device, the NN using the sets of features and the respective labels to predict during which segment of the digital audio signal the user utterance has ended,
• the training including, for a given training iteration of the NN:
• generating a plurality of individual penalty values, a given individual penalty value being associated with a respective segment of the digital audio signal, the digital audio signal including segments that precede the end-of-utterance moment in time, the given individual penalty value being indicative of a prediction error made by the NN for the respective segment of the digital audio signal;
• generating a combined penalty value for the given training iteration of the NN, the combined penalty value being a combination of the plurality of individual penalty values;
• using the combined penalty value to train the NN during the given training iteration.