Speech2IPA Noise Robustness

[SanderGi]

As pointed out in #11, transcription is very sensitive to background noise. We have run some preliminary experiments:

1. Using webrtcvad to filter out non-speech segments: doesn't handle overlapping speech/sounds
2. Training models on various kinds of augmented noisy speech: does not generalize well to unseen types of noise

We need a low latency approach to remove noise that generalizes well to different types of noise. Some ideas that different PRs can explore:

1. Enable noise suppression via Web API on supported devices/browsers
2. Evaluate various open-source noise suppression models, see which we can run on-device in the browser and which would need to be hosted on a server with a GPU
3. Look into more advanced noise suppression based on binaural audio and/or speaker specific embeddings
4. Look into more noise robust Speech2IPA architectures, training objectives/regularization, and data augmentation approaches