A computer-implemented method for training a neural network to process as inputs a plurality of audio signals comprising a first audio signal and a second audio signal obtained from a microphone array arranged in a conference room and provide as output an indication of an estimated probability of voice activity for a plurality of directions relative to the microphone array, wherein the method comprises executing, by a computer, multiple training rounds, wherein each training round comprises:
- determining a training data set comprising a first training audio signal, a second training audio signal, and a target direction;
- applying the first training audio signal and the second training audio signal as respective inputs to the neural network;
- determining an error signal based on an output of the neural network and the target direction;
- and
- using a learning rule to adjust weights of the neural network based on the error signal,
- wherein determining a training data set is based on:
- a user speech signal selected from a plurality of speech signals each indicating speech from a first person;
- a noise sound signal selected from a plurality of noise sound signals each indicating an ambient noise sound;
- a room model selected from a plurality of room models each indicating a spatial extension and one or more acoustic properties of a conference room;
- a user position within the spatial extension indicated by the selected room model;
- and
- an array position defining a first microphone position and a second microphone position within the spatial extension indicated by the selected room model,
- wherein determining the first training audio signal and the second training audio signal comprises:
- determining a first impulse response of a first acoustic path from the user position to the first microphone position and a second impulse response of a second acoustic path from the user position to the second microphone position based on one or more acoustic properties indicated by the selected room model;
- estimating first sound arriving at the first microphone position by combining the result of applying the first impulse response to the speech indicated by the selected user speech signal and the ambient noise sound indicated by the selected noise sound signal;
- estimating second sound arriving at the second microphone position by combining the result of applying the second impulse response to the speech indicated by the selected user speech signal and the ambient noise sound indicated by the selected noise sound signal;
- and
- determining the first training audio signal based on the first sound and the second training audio signal based on the second sound,
- wherein determining the target direction comprises:
- determining a direction of arrival of user speech at the microphone array based on the user position and the array position;
- and
- selecting the target direction from the plurality of directions such that the target direction does not differ more from the determined direction of arrival than any other direction in the plurality of directions, and
- wherein determining the error signal comprises:
- determining a classification error for each of the plurality of directions based on the respective probability of voice activity indicated in the output of the neural network such that the classification error for the target direction increases when the indicated probability of voice activity decreases and the classification error for each other direction increases when the indicated probability of voice activity increases;
- and
- determining the error signal based on the classification errors for all of the plurality of directions.