poor results on narrowband communications speech

[audio441]

I've installed DeepFilterNet 0.5.6 as well as the linux command line version. I'm trying to process narrowband (300-3000 Hz) two-way radio communications. If I upload a clip to https://noisereducerai.com/process-audio/, it sounds great. The same clip processed locally fails to remove some noises and attenuates the some of the speech. A speaker can fluctuate in

The audio looks like this:

The output with atten_lim_db=30.0 looks like this:

The beginning and end of the clip are the same speaker in the same conditions.

I've tried adjusting lsnr_min in config.ini, does not change the audio. Same with min-db-thresh in the command line version.

This is just surprising considering how well noisereducer.ai works on the clip.
Any advice?

[danielr-ceva]

@audio441, you can try DPDFNet’s live demo to see whether it works well for your use case. DPDFNet is an enhanced version of the DeepFilterNet architecture, and the models are publicly available.

Make sure to select one of the 8 kHz models: DPDFNet2 or DPDFNet8.

Hope you find it useful!

[audio441]

@danielr-ceva This is great! I tried the demo on my content and it does a good job. I'm pre-processing audio for Whisper, will need to see if its WER improves. To the ear, this sounds good.