Shift between recordings among each pair can yield information about potential issues with the recordings. Besides, they might be needed for some analysis, e.g. when we need to correlate recordings or to combine extracted metrics to reduce noise.
In the simplest case, the shift evolves linearly with time, because of the audio drift of the recorders (the recorders probably have an intrinsic true sampling rate that differs from the theoretical frequency). The pace at which the shift changes over time is of the order of 1 second per day, which might not always be negligible. In this case, it is possible to achieve millisecond-level accuracy for shift estimation at any time of the recordings using a linear regression.
For some pairs, the shift increases linearly but also features jumps every few hours. A defect in some recorders might cause them to stop recording briefly sometimes. We can use the results to assess the frequency and extent of this defect.
Sometimes, the calculated shifts are random, which could mean that the true shift was larger than the 5 minute time-frame used by the algorithm to estimate them. In this case the algorithm returns random values. It is however demanding to compute cross correlations on much longer intervals.
lucasgautheron
commented
3 years ago
Shift between recordings among each pair can yield information about potential issues with the recordings. Besides, they might be needed for some analysis, e.g. when we need to correlate recordings or to combine extracted metrics to reduce noise.
In the simplest case, the shift evolves linearly with time, because of the audio drift of the recorders (the recorders probably have an intrinsic true sampling rate that differs from the theoretical frequency). The pace at which the shift changes over time is of the order of 1 second per day, which might not always be negligible. In this case, it is possible to achieve millisecond-level accuracy for shift estimation at any time of the recordings using a linear regression.
For some pairs, the shift increases linearly but also features jumps every few hours. A defect in some recorders might cause them to stop recording briefly sometimes. We can use the results to assess the frequency and extent of this defect.
Sometimes, the calculated shifts are random, which could mean that the true shift was larger than the 5 minute time-frame used by the algorithm to estimate them. In this case the algorithm returns random values. It is however demanding to compute cross correlations on much longer intervals.
This graph shows a few results :
shifts.pdf