Small arrays cause unpredictable behavior in DIO

[KnurpsBram]

I'm interested in real-time voice conversion. I'm experimenting with showing small buffers of audio to WORLD, so as to perform conversion chunk-by-chunk. I would expect DIO to be able to recognize a frequency if at least two wavelengths are shown to the function. For 100Hz, this is 20ms. I see a weird behavior where DIO can't recognize 100Hz if the buffer is under 100ms. Strange! What could be the cause of this?

import numpy as np
import pyworld as pw

sr = 16000 # samplerate of 16kHz
f = 100 # Hz
dur = .5 # seconds
x = np.cos(2*np.pi*np.arange(sr*dur)*f/sr)
f0, t = pw.dio(x, sr)
print(f0)
f0, t = pw.dio(x[:int(.1*sr)], sr)
print(f0)
f0, t = pw.dio(x[:int(.12*sr)], sr)
print(f0)

[JeremyCCHsu]

Could you post this on https://github.com/mmorise/World ?
Unfortunately, I don't have an answer to this.
WORLD could have been designed to assume extra period of time for pitch evaluation.