Dereverberation before recognition: when it helps and when it does damage

Reverberation smears the signal, so the intuitive move is to clean it before recognition. That intuition is wrong often enough to state plainly: the standard modern answer is to train on reverberant data rather than to remove the reverb, and when a front end is used, it has to be part of the training configuration rather than a filter bolted on at test time.

The decision comes down to which strategy you can execute consistently. Two of them work.

Strategy one: train through it

Convolve clean training audio with measured impulse responses, mix in noise at 10 to 20 dB, and let the acoustic model learn the channel. This is the default for far-field systems, and it consistently beats front-end enhancement, for a simple reason: the model is trained on the distribution it will meet, and no artifact is introduced that the model has never seen.

The quality of the simulated rooms sets the ceiling. A handful of impulse responses from one building produces a model that has learned that building. Dozens of rooms, several distances, and a mix of measured and synthetic responses is where the gain comes from.

Strategy two: a linear front end, trained in place

Weighted prediction error is the safe front end: a linear filter that predicts late reverberation from past frames and subtracts it. It has no learned components, so it introduces no musical noise and no spectral holes, and its behavior is identical on every file.

Practical settings: a frame of about 512 samples with a hop of 256, a filter of 10 taps with a delay of 3 frames, and 3 iterations. The delay matters. Predicting from the immediate past would cancel the speech along with the reverb, and the delay is what separates the early signal from the late tail. It runs on CPU at roughly real time per channel, which is cheap enough to apply across a whole corpus.

It removes the late tail, which is the part that most damages recognition, and leaves early reflections. A measured reverberation time will therefore not fall dramatically after processing, and that is expected rather than a sign of failure. Check the right thing: compare the energy in the tail region of a convolved test file before and after, not the total reverberation time. If the audio also carries noise, the front end does not address it, and the ordering is dereverberation first, noise handling second, because the noise estimate is more reliable once the tail is gone.

Where neural dereverberation goes wrong

Learned enhancement models remove reverb and noise together and score well on enhancement metrics. Those metrics do not predict word error rate, and the failure mode is consistent: over-suppression of low-energy phonetic content, especially fricatives and final consonants, plus spectral holes that look acceptable on a spectrogram but shift the feature statistics away from what the acoustic model was trained on.

If a learned front end is right for the application, train the recognizer with it in the loop and select it by word error rate, never by an enhancement score. Choosing a front end by how clean it sounds is how a system gets worse while every intermediate metric improves.

Diagnosing the outcome

Four checks, each aimed at a specific failure.

• Compare word error rate on clean, simulated reverberant, and real far-field sets. An improvement on simulated and a regression on real almost always means the impulse responses behind the simulation do not represent the deployment rooms.
• Measure the 4 to 8 kHz energy ratio before and after processing. A large drop is the over-suppression signature.
• Check the active-speech level. A fall of more than about 2 dB after processing means the filter is eating speech, not only reverb.
• Listen to twenty files, alternating original and processed. Artifacts that fail to show in aggregate metrics are usually obvious to the ear within a minute.

When to skip it

If the room is small and the reverberation time is under about 0.3 seconds, dereverberation buys little and adds a step that can only lose information. If the model cannot be retrained, do not insert a front end at inference: a mismatch between training and test conditions is a larger effect than the reverb it removes. The justification grows with the distance from the microphone and with the length of the tail, and it holds only when training and inference run the same chain.