Unfortunately there is no speech-recognition or speech-to-text tag yet so I go with the voice-recognition.
My question concerns various datasets for automated speech recognition and how training and test split should be generated. Even more specific, I'd like to know how important it is to separate speakers from the training, test and dev split.
In very imbalanced datasets one might have very few speaker which contribute a lot to the corpus and thus the model may be biased towards those speakers and have generalized less. Is this concern legitimate?
