nvidia

/

diar_streaming_sortformer_4spk-v2

like 130

Follow

NVIDIA 64.2k

Hi @FIT17 ,

Thank you for your kind words about the 4-speaker model and for the great feedback. We really appreciate hearing about your use cases.
To answer your question: yes, an 8-speaker version is on our roadmap.
We are currently working on it, and the release is targeted for the first half of 2026.
We value your input, and please stay tuned for future announcements.

Hi, @imedenikov
It's great to hear that the 8-speaker version is on the roadmap for the first half of 2026.
Could you let me know if there is a more exact release date set for it yet, Since we are already in the first half of 2026?

Hi @FIT17 , our current target is the end of June. Please note that, while we are pushing for this timeline, we cannot fully guarantee it at this moment.

Hi @imedennikov

Thanks again for the earlier update. Now that we're in early June, I wanted to
check in: is the 8-speaker model still on track for the end-of-June target?

I completely understand it couldn't be fully guaranteed back in March — just
hoping to get your latest read on whether a June release still looks realistic,
or if the timeline has shifted at all. Even a rough update would help us plan on
our side.

Really looking forward to it. Thanks!

Hi @FIT17 , unfortunately the timeline has shifted a bit. The realistic target for the release is around August-September.

Hi @FIT17 , unfortunately the timeline has shifted a bit. The realistic target for the release is around August-September.

Thank you for your contributions. Will the new model focus solely on adding more speakers, or will there also be improvements related to parallel speech or overall accuracy?
ty again looking forward the new release 🙃

Hi @altunenes , the new model should perform better overall.

Hello, would it be possible for you to also make a lightweight version of the 8 speaker model?

Hello, would it be possible for you to also make a lightweight version of the 8 speaker model?

Hello @GradientDescent2718 ! Thanks for reaching out.
We do not have plans to release a separate lightweight version at this time. However, our new model is heavily optimized—you will see faster performance and lower memory usage compared to v2.1.
If you require an even smaller footprint for your specific use case, we recommend exploring distillation on your end.

BTW, If anyone here is interested in early access to the new model, please react to this comment with a ➕!

➕➕➕ please :-)

BTW, If anyone here is interested in early access to the new model, please react to this comment with a ➕!

➕

Can't wait to see it!

➕➕➕➕➕➕

If you require an even smaller footprint for your specific use case, we recommend exploring distillation on your end.

Hello @imedennikov , would you happen to know the effects of reducing the number of transformer and conformer blocks in the model? I am wondering if that would be an effective way to create a smaller model for distillation.

Hello @imedennikov , would you happen to know the effects of reducing the number of transformer and conformer blocks in the model? I am wondering if that would be an effective way to create a smaller model for distillation.

Hi @GradientDescent2718 ,
Sorry for the delayed reply.
While we haven't run ablation studies on this specifically for Sortformer models, a recent paper (https://arxiv.org/abs/2506.18623) covers this exact question in depth.
This should serve as an excellent starting point for your own distillation efforts.
Best of luck with your optimizations!

BTW, If anyone here is interested in early access to the new model, please react to this comment with a ➕!

is there a specific form or process to have the early access to the model?

@GordonChang - Once ready, a note will be posted here.