Artificial intelligence models that generate audio are being trained on datasets plagued with bias, offensive language and potential copyright infringement, sparking concerns about their use.