Predicting positive transfer for improved low-resource speech recognition using acoustic pseudo-tokens