Hacker News Reader

Comment by krisoft | Hacker News Reader

> So essentially, that model has been trained to fool human perception.

Is there any other sensible goal function for a text to speech model?

> Similarly, ChatGPT is not trained to output sensible and meaningful statements but rather statements that appear to be to a human reader.

Almost certainly not true. If they could they would make it output sensible and meaningful statements all the time.

> Like feeding a simulator with actual data from real-world traffic scenarios and running Tesla's, or any other company's, autopilot in it?

Do you seriously think that self driving car companies are not doing this already?

View on HN

> Is there any other sensible goal function for a text to speech model?

Yes: to transfer information aurally. And the models are there, and have been for quite some time. They can simply stop with the trying-to-fool part.

The "fooling here" is "people wanting TTS to sound human". And obvious step for that is making the AI infer human voice traits from existing humans

It doesn't have to sound human. Fluent, ok, but human?

But no matter what, it's unnecessary to make a tool that can copy anyone's voice. That just lowers the threshold for abuse, while not adding almost nothing of value.

Listening to something that sounds non-human for a long period of time is fairly unpleasant; imagine trying to listen to an audiobook or podcast, or dialogue in an animated movie, when the voices are all obviously non-human/robotic. So TTS wouldn't be usable for a lot of cases where we might want it.

And with the way the models work, once you have a model that can sound human, it is unfortunately very easy for it to sound like any individual human as well.

It doesn't have to, and you can use ones that don't, they get the info across mostly. But it is jarring if it doesn't sound human, it's a speech impediment, or as a more generous take, an accent.

The correct inflections, pauses, annunciations, etc. are all important to humans, especially so for audio books and similar things that need to immerse.

Otherwise you have to strain to listen, similar to listening to someone with a heavy accent.

> especially so for audio books

Perhaps real human readers can help?

> The correct inflections, pauses

Because a model that can imitate a voice is still not capable of that. There's no need to have model that can do that. A robotic accent is best. Or perhaps you like to see your politicians make all kind of bizarre statements on youtube.

> Perhaps real human readers can help?

Certainly, and they do, but computers can annotate much faster, without attrition, and cost much less.

Generally you'll have human recordings for what you can, and TTS for anything missing.

There are a lot of books, live streams, podcasts, articles, etc. in the world.

> There's no need to have model that can do that.

I wouldn't say there's no need, or otherwise we wouldn't talk that way. It's a human element and the generated speech is for human ears.

> A robotic accent is best.

Best is subjective because that's a human preference. That being said, I'd say your preference is a far outlier.

Most people's "best" would be what they are used to hearing; the speech of native speakers.

Reply on news.ycombinator.com