Missing API - The model silently fails and produces audio with long silences, happens to be at 1000/1000

Once every 20 generations or so (~5%) I see that the model has hit 1000/1000 and I know that its a failed generation. Overall this is a great result for the model, especially since re-running the generation almost always succeeds.

The problem is that the python lib does not give any feedback that can help flag the issue and re-run generation. I have been doing that manually when I encounter an issue in the generated audio (long stretches of silence). Although I did notice that there is a 100% correlation with sampling stats printing 1000/1000, so I ended up doing intercepting tqdm bars on close and retrieving the information from there, to decide if the generation has succeeded.

I think we need a better API that can indicate whether the generation is a success. Maybe the returned wav file can be None if its 1000/1000. That would be good enough.