Adjust default parameters to make even small samples do something

[joanise]

Bug description

For regression testing, the 15 minute of data test case yielded 150 utterances from LJ, and that caused training to fail with this error:

2025-01-23 12:13:40.264 | ERROR    | everyvoice.utils:filter_dataset_based_on_target_text_representation_level:96 - Sorry you do not have enough characters data in your current validation filelist to run the model with a batch size of 16.

This appears to be due to having just 15 samples in the validation set.

We should adjust the default wizard and training defaults so that if the data has <160 utterances, things are setup so training can proceed anyway.

How to reproduce the bug

• Create a dataset with 150 samples
• run the wizard
• everyvoice preprocess config/everyvoice-text-to-spec.yaml
• everyvoice train text-to-spec config/everyvoice-text-to-spec.yaml

Or from the branch for #616 run go.sh and inspect the logs in regress-lj-150/

Error messages and logs

2025-01-23 12:13:40.264 | ERROR    | everyvoice.utils:filter_dataset_based_on_target_text_representation_level:96 - Sorry you do not have enough characters data in your current validation filelist to run the model with a batch size of 16.