fix: real time gui not work

[RoversCode]

This PR addresses two issues and adds a pitch-shifting feature:

1. Fixed the issue where infer_wav_res produces meaningless noise
   The real-time self.resampler conversion was corrupting the audio output. While the root cause is still under investigation, librosa resampling has been implemented as a working solution.
   2 . Fixed the silent output issue
   When block time is too large, the vad chunk size becomes excessive, causing infinite loops in 'if not self.vad_speech_detected' condition, resulting in silent output.
2. Added pitch-shifting functionality
   Implemented pitch up/down features for input audio.

[Plachtaa]

Hi there, thanks for your hard work!
For 1. , I question your opinion on "The real-time self.resampler conversion was corrupting the audio output", reason for producing meaning less noise is most likely to be the case in #125 , please let me know if this is not your case.
For 2. , This is a good bonus feature, but I cannot run your script as there are multiple model checkpoint loading issues. If you agree, I will make some changes on your script before merging your PR.

Thanks

[RoversCode]

Thank you for your feedback!

Regarding point 1, through debugging, I've located that the issue is indeed with self.resampler - it converts valid speech signals into silence. I believe issue #125 is likely caused by the same problem, as I observe intermittent audio anomalies in real-time speech output even without making any modifications.

Regarding point 2, feel free to modify the code as needed, since I only added this feature in a simple way without much consideration for code optimization and performance.

I mainly wanted to report these two bugs. Specifically about the self.resampler issue, I'm also puzzled about the root cause. I've reviewed the code and found it's just using the basic torchaudio.transforms.Resample implementation. While strange, I can confirm the problem is definitely with this component