A method and a device for identifying an utterance from a signal is disclosed. the method includes acquiring a set of features for a respective segment of the signal, and an indication of an end-of-utterance moment in time in the signal corresponding to a moment in time after which the utterance has ended. The method includes determining an adjusted end-of-utterance moment in time, and labels for respective sets of features based on the adjusted end-of-utterance moment in time and time intervals of the corresponding segments. A given label is indicative of whether the utterance has ended during the corresponding segment associated with the respective set of features. The method also includes using the sets of features and the respective labels for training a Neural Network to predict during which segment of the digital audio signal the utterance has ended.