Adapting generic speech recognition fashions to particular people is a difficult downside as a result of shortage of customized knowledge. Current works have proposed boosting the quantity of coaching knowledge utilizing customized text-to-speech synthesis. Right here, we ask two basic questions on this technique: when is artificial knowledge efficient for personalization, and why is it efficient in these circumstances? To handle the primary query, we adapt a state-of-the-art automated speech recognition (ASR) mannequin to focus on audio system from 4 benchmark datasets consultant of various speaker sorts. We present that ASR personalization with artificial knowledge is efficient in all circumstances, however significantly when (i) the goal speaker is underrepresented within the international knowledge, and (ii) the capability of the worldwide mannequin is restricted. To handle the second query of why customized artificial knowledge is efficient, we use controllable speech synthesis (CSS) to generate speech with assorted types and content material. Surprisingly, we discover that the textual content content material of the artificial knowledge, relatively than fashion, is essential for speaker adaptation. These outcomes lead us to suggest an information choice technique for ASR personalization based mostly on speech content material.
