Question about rx

[nikifori]

Is there any particular reason rx library is used?
Why do we need asychronous code in this repo?
What if we did not use rx at all, and the code was written straightforward?

Thanks.

[lq0104]

I have the same question. How should I do it if I don't use rx?

[juanmc2005]

Hello! There are several reasons why rx is a good choice in streaming applications.
For diart, it allows easily handling asynchronous and non-blocking processing (useful when you need to process large or many streams).

Additional benefits are also thread management and scheduling (instead of having to manually create and synchronize threads), error handling and ease of stream composition and transformation.

A big reason is also that it allows you to deal with backpressure (audio being sent faster than the model can process it). Although diart doesn't apply any specific strategy for this, it's nice to have the possibility to leverage it as part of the library from day one if you choose to use it with rx.

That said, there is no need to use rx in diart if you don't want to. All the steps of streaming diarization are implemented as separate blocks so you can easily compose them the way you prefer. If you check diart.blocks.diarization.SpeakerDiarization, you'll notice that not one single line of rx is to be seen.

This is by design. You can use diart at 3 different levels:

1. (lowest) you take each diarization block (segmentation, embedding, clustering, aggregation, etc.) and you chain them together manually
2. (high) you simply pass audio chunks to SpeakerDiarization, which handles point (1) for you
3. (highest) you use the opinionated implementation of StreamingInference and AudioSource, which handles point (2) for you

Lastly, note that in order for diart to provide point (3), we needed to choose and implement a way to handle audio streams (hence "opinionated"). We chose rx for this, but you can start from (2) and re-write your own (3) with co-routines, or threads, or a single thread.

If you find a better approach than rx, I'd gladly consider replacing or at least supporting the new one.