decoder was much slower than encoder about lyra

[aluebs]

Thank you for the feedback. With the current implementation the decoder is expected to be much slower than the encoder, since the encoder only has to extract the features, but the decoder runs the expensive generative model to recreate the time domain signal from the features.

[scguoi]

Is there has any plan to improve decoder performance?

[aluebs]

Yes, absolutely, we are currently working on reducing computational complexity of the decoder.

[aijianiula0601]

I find that the decoder consume 50ms with one frame(40ms) and opus decoder is 0.3ms, it that right?

[aluebs]

It is highly dependent on the architecture. Our matrix multiplication libraries are optimized for arm64 devices and for medium-end phones Lyra should be at least 4x faster than real-time. But that is still slower than Opus.

[aluebs]

The new Lyra 1.2.0 release is about 5x faster than the previous version. And moreover, now the encoder and decoder have similar complexities.