“The speaker estimates that current voice-to-voice models have a hallucination rate approximately 8 times higher than text-based models, making them unsuitable for many enterprise applications.”