this is a question that has interested me for some time now, mainly because I'm working on noise reduction for an existing speech recognition system myself.
Most papers on noise reduction techniques seem to focus on how to make speech more intelligible for humans, or how to improve vague terms like "speech quality".
I'm sure that, using criteria like these, you can identify filters that make noisy speech signals easier to listen to for humans. However, I am not sure that these criteria can simply be adapted when trying to evaluate speech signals that have been denoised to improve the accuracy of speech recognition system.
I don't really find papers that discuss this difference. Do speech intelligibility and speech quality correlate with the accuracy of speech recognition systems? Are there objective measures that can evaluate how "good" a denoised speech signal will be for a speech recognition system, for example if also given the original clean speech? Or is the only way to find out how good your noise reduction technique is, to train the speech recognition system on the denoised data and look at the accuracy?
I'd be happy if someone could point me into the right direction, or maybe give some papers that discuss this. Thanks in advance!