Joint training of speech enhancement and ASR sounds obvious — just backprop the recognition loss through the denoiser — but in practice it's messy, because aggressive denoising often distorts features the acoustic model relies on. This NTT paper proposes a cleaner objective: instead of enhancing to "clean," enhance to a target signal-to-noise ratio that the downstream recognizer actually prefers, and let the model learn where on the noise-distortion tradeoff to sit.

The SNRi target is estimated per utterance and used as an explicit control signal to the enhancement network, so the front-end stops over-processing easy segments and stops under-processing hard ones. The walkthrough covers how the target is derived, how it plugs into standard enhancement architectures, and the CHiME-style benchmark gains it delivers over both cascaded and vanilla joint-training baselines. If you're shipping ASR into noisy real-world environments — cars, call centers, wearables — this is a subtle but powerful trick to add to your pipeline.