indistinguishable is a high and subjective bar; for many people, the "portrait mode" depth of field effect on iphone/android photos is "indistinguishable" from real depth of field; for people more familiar with what how real lenses render depth of field, it looks goofy and fake. I suspect voice synth will fall into the latter category for most people for quite some time; humans are incredibly fine-tuned to perceiving human voices.

a blind test conducted by a panel of at least 4 people of my choosing, cannot be reliably distinguished from a human narrator.

Can you design a test that would be failed by every TTS currently available on the market? If yes, could you elaborate on what that test would look like?

@nsokolsky Not sure that every TTS on the market would fail such a test; OpenAIs TTS is really good. But I'm thinking something like: make the test subjects listen to four 30-second snippets where three are good human narrators, and one is the AI voice. The subjects are tasked with identifying which voice is artificial. If at least two subjects hesitate before making their choice and at least one subject fails, then the model passes the test.

For comparison, do you believe, say ElevenLabs or OpenAI's TTS to meet this bar?

@DavidFWatson Haven’t tried ElevenLabs but I think OpenAI probably meets the bar.