Machine Learning / AI

How to assemble train/val/test split and benchmark? How to properly split the data?

Tell us about the pronunciation assessment task: how to approach improving a model that detects errors in phonemes and sounds?

We found LR on a small machine, the batch size increased by 1000 times on the cluster. Should we change LR?

What should be done if a test has already been used and showed a poor result?

When to use train, validation, test, and benchmark?

There are many optimizer hyperparameters, 10,000 experiments, little time but many resources. What approaches to find the optimal combination?

What does the classifier take as input besides the audio embedding?

What classification task are we solving when evaluating phoneme pronunciation?

How do we train a phoneme classification model? What's in the target?

What ideas can be tried for precise phoneme alignment?

How to practically feel that a test has failed or that a distribution shift has occurred?

When should tests be changed? How to distinguish when a test is outdated from when it has burned out due to repeated use?

If the batch size is larger, will the gradient become smaller or larger?

How to perform forced alignment — determine which phoneme sounds from which second to which?

Are Learning Rate and Batch Size independent parameters? Is there a relationship between them?

Grid 100x100 vs 10000 random points — why is random better than grid even in 2D?

How to approach training on your own domain? What to do with the general model?

How to annotate audio data?

How to train a model? Seq-to-seq is expensive, and there are many hyperparameters. How to do it?

How to approach aligning a single word by phonemes? How to use wav2vec2 to obtain accurate phoneme boundaries?