the DeepFilterNet 2 paper has benchmarks which include the RNN model used in the article. TLDR is that DFN2 does a better job. I suspect they went with the RN

the DeepFilterNet 2 paper has benchmarks which include the RNN model used in the article. TLDR is that DFN2 does a better job. I suspect they went with the RNN model because the Pico 2 could run it with the 10ms max latency they wanted. DFN2 is fast, but I'm not sure it's lightweight enough for something like the Pico 2.