Any luck with this?

[millimetre]

Hi James,

A buddy and I are embarking on working on something very similar to what it seems you worked on, getting Spleeter to run in real time on Raspberry Pi (ARM, low latency, real time, etc). Did you ever have any luck with this, or did you hit a wall? Any guidance you could provide would be much appreciated, and if it makes sense to create a consulting arrangement, we'd be interested in discussing that as well. Thanks in advance!

Paul

[james34602]

Hello Paul,

It depends on which Raspberry Pi you own, spoiler alert, you need a little stronger CPU to run SpleeterRT.
Last time I check, Raspberry Pi 4 are capable running SpleeterRT in reasonably fast level, the viable application is music playback and as a piece of ultility in waveform editing software like iZotope.

This repo uses double spectrogram buffer to enable us to run SpleeterRT(1 second latency) on relatively low power device.

What we haven't seen on my Github page is algorithmicly low latency SpleeterRT, the extremely low latency(46.4 ms) of it is currently non-open-source, due to business reason, it won't be open in short term, this one require a lot stronger CPU to run.

We are running a startup since then where we've made advancements and offer our technologies for third-party developers, such as realtime ML audio tech (audio reparation, dereverberation, separation and more).

You might like to contact api@splitter.ai to initiate talks and I'm sure we can solve your challenges.

Best regards,
James

[SuperKogito]

Hello James,
great project! and interesting results you mention here 👏
If you don't mind saying, when you mention a latency of 46.4 ms, does this include the same architecture of SpleeterRT neural net with the initial delay of Spleeter of ~1.5 seconds (with timeStep = 64, minimum height for NN input) or did you somehow manage to lower that initial delay also down.

thank you in advance for any response :)

[james34602]

Hello James, great project! and interesting results you mention here 👏 If you don't mind saying, when you mention a latency of 46.4 ms, does this include the same architecture of SpleeterRT neural net with the initial delay of Spleeter of ~1.5 seconds (with timeStep = 64, minimum height for NN input) or did you somehow manage to lower that initial delay also down.

thank you in advance for any response :)

Hi, I'm not active in responding to Github issues.

ok, here is the thing.
We offer a low latency SpleeterRT solution far lower than open source alternatives.

The latency figure is 64ms - 10 seconds, the runtime complexity is not ultra demanding, although more demanding than open source counterpart.

If you are interested, please contact api@splitter.ai