Spaces:

aadnk

/

whisper-webui

like 332

Running

Tracking issue: #5

Thanks for making a PR - being able to set the temperature will likely be a useful feature in some cases.

Though while initially testing this PR, I found that the best temperature was still 0 (using this video), and higher temperatures just started to introduce more noise. And at 2, the transcript was complete gibberish. Though that's probably because the audio in that video is particularly clear and easy to understand. Perhaps you have a few examples where the temperature setting is useful?

There's also a few other settings that are available in the CLI that's not currently available in the WebUI, such as "best_of", "beam_size", "patience", etc. I was thinking I'd use TabbedInterface and create an advanced version of the WebUI with every setting available currently in the CLI. That way, people who need the extra flexibility of these more "low-level" options can access them using the "Advanced" tab, and the rest can continue to use the default simpler UI in most cases. I'd prefer not to completely overwhelm people with too many options. Still there's probably room for at least one more setting in the default UI, like "Temperature".

Unfortunately, there's a few issues that prevent me from merging this PR right away:

1. First, I see that you added a "port" argument to app-network.py. This is slightly beyond the scope of this PR, but ... if you're going to add it, could use use ArgumentParser instead, as in cli.py? So that it's syntax is consistent with the CLI?
2. I see you've also removed a number of requirements from requirements.txt. The problem is that I actually do use or need some of these dependencies - "torchaudio" is used internally in "silero-vad", and "ffmpeg-python" is used to download audio from URLs (such as YouTube). It is possible "transformers" is redundant though, as Whisper itself depends on it. Still, I feel this change should at the very least be in its PR, as this is not needed to implement the temperature setting in the WebUI. And then only actually remove "transformers".