from minimize error to raise quality

In this post, I am going to share the finding (and audio samples) of applying perceptual quality as training target for causal model.

By employing Reinforcement Learning techniques, non-differentiable metric, such as speech perceptual quality score PESQ or STOI, can be used as model training target. However in the examples I found, they all use non-causal model, which is not suitable for realtime application (more background in “peek into the future”). As a keen embedded ML developer, I am curious about how much improvement it could bring to causal model.

Two models will be trained using the same dataset, identical model architecture (RNN based, causal, 4ms algorithmic delay) and size (300k parameters), the only difference is the target definition and corresponding training method.

Training targets:

1. Minimize error when compared with ideal clean speech. With synthesized dataset, clean speech can be conveniently used as ideal answer. A popular metric, SI-SDR is used.

2. Not so focus on errors, instead we want the output to be pleasant to our ears. Speech perceptual quality score is used as metric and the training target is to achieve as high score as possible. Let's call this Q target in the rest of article.

Two sets of result will be presented now. Test clips are unseen by the models during training.

Waveform comparison

First, both models have done a great job by the look of the waveform, STOI scores are close too. Holding my breath and change to spectrum view to inspect further.

Spectrum comparison

Q target output has done a visible better job in removing the background noise, extra 6dB at the peak in fact as the spectrograms show, not bad at all!

Some audio samples are available here for your listen.

My takeaways from this experiment:

• Train towards perceptual quality can definitely improve model’s understanding about speech, resulting in more precise noise removal. This part is very encouraging.

• No noticeable quality score improvement is kind of expected. Meaningful improvement likely require more of signal conditioning, filling in the corrupted waveform, therefore context is crucial and causal condition limits the amount of context can be used.

• The additional effort of applying Q target is not small, it might not be the best option for quick evaluation, but for production model development, it is worthwhile doing.