Otsu threshold sometimes fails

[ttencate]

Example:

./master.py --store_audio_files --debug_recording_ids xc:295996 --debug_utterances --debug_otsu_threshold

The histogram looks like a single Gaussian, not bimodal, because the recording is almost all utterance and no silence.

See if we can detect such cases and just trim from the beginning if they happen. Perhaps simply if >80% is classified as utterance?