Unstable voiced regions
Voiced regions without stable intervals or voiced regions that have no sufficiently large subsequences of equal F0 are treated in a separate postprocessing step. Basically, the same propagation procedure is applied but the propagation starting or anchor point is found using looser conditions and additional information.
First, we compute the mean F0 of the last second of speech considering only frames with a verified F0, i.e. frames in equal sections of stable intervals used as starting points for F0 propagation or propagated frames in unstable intervals. We then compute candidate F0 values for all unstable frames of the voiced regions in the range of the mean F0. The permissible F0 range is the same as described in section 9.4.1. The anchor point for propagation is found by inspecting the equal section list of the stable intervals in the voiced regions, or a small section around the highest energy frame if no stable interval in the voiced region exists. The propagation starts from such a section if the mean of the F0 estimates does not deviate too largely from the last second’s mean F0. If no such section takes place, we leave the F0 estimate unchanged. In this case, no propagation takes place.