Several threshold variables

[Jeong-Haneul]

Several threshold variables are currently combined using an "or" operation, meaning that as long as one condition is satisfied, results are produced. This leads to outputs even in cases where the environment is noisy and the probability is low, or when it's quiet but the probability is high—both of which result in incorrect outcomes.

There seems to be an issue where, due to some incorrectly trained data, the system returns incorrect data with high probability, even when the audio contains low-volume noise.

I believe that switching to an "and" operation for the conditions could resolve this issue. However, would this change introduce more unknown problems? Alternatively, it might be beneficial to add an option that allows the choice between "or" and "and" operations, ensuring compatibility with existing code.

now

result = whisper_model.transcribe("temp.wav", language="ko", initial_prompt=initial_prompt,
temperature=[0, 0.5, 1.0], best_of=2, beam_size=2, no_speech_threshold = 0.05,
logprob_threshold = -0.35, condition_on_previous_text = False)

result
text ㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷㄷ
no 0.10252052545547485 <- ignored
prob -0.062140274047851565 <- It appears that high accuracy appears in silence due to the influence of bad data.
compression_ratio 20.8125 <- ignored