Hacker News Reader

Comment by jgrahamc | Hacker News Reader

I made one big mistake with that introduction when it was recorded on to the video tape for playback at the conference: I didn't realize how much people were going to laugh at the end and went straight into my presentation. I got to hear the laughter because the conference was live streamed using RealPlayer.

View on HN

I would like to see more modern text-to-speech that sounds good, but also inhuman. On one hand very easy to understand but on the other obviously constructed and not something a living person could generate. Like the scifi where robots always look like robots and it's illegal/immoral/taboo to have a robot that's indistinguishable from a person.

Maybe just an ordinary human sounding TTS that gets put through a mild vocoder of some kind.

It sounds like a guy recorded his own voice and ran it through an LPC.

That's a real problem with the vast majority of current TTS. Terrible in things like consistent intonation, proper pronunciation, believable pauses, while sounding human all the same and the result is super uncanny valley.

The gaming and movie industry understands this very well, they use human voice actors that can nail all of that and then make it sound more metallic or compressed or whatnot. Otherwise it does not fit.

And on the high end, making everyone sound like a professional public speaker. The machine sees mistakes as errors when in fact every non-hollywood speech contains multiple mistakes.

This is why it pisses me off when a techy person makes a YouTube video and just uses TTS instead of recording a voiceover. I know some people don’t have a good recording situation, but I get the sense that a lot of people just do it because they either think people can’t tell or they think it’s a clever hack or “the way of the future” or something.

It isn’t. Instead I find myself watching videos and getting a weird creepy feeling when I suddenly hear the voiceover mispronounce a word or put an emphasis in the wrong place. Part of it is the uncanny valley for sure, but the more pernicious thing is this: once I realize that the voice is AI-generated, I start to worry that the script might be too. Now I’m trying to figure out “is this guy just an amateur writer taking a while to get to his point, or is this an LLM-authored script that is never going to go beyond surface-level statements about the topic.”

I think this undersells the difficulty of recording a good voice-over, both technically and performatively.

I don’t think this is about a “good recording situation”. It’s likely people who think they suck at speaking/narrating or think they have a horrible accent or want to remain anonymous or just find the process annoying, and find it less embarrassing/more privacy-preserving/less of a hassle to use an AI voice.

Those are the most likely.

Another factor, less common, is when you want or have to speak a non-native language you're not used to pronounce, in which case you're usually afraid of not being understood.

PS: I think all the Text To Speech systems sounds horrible, the last generations are even irritating, as the user of the parent commented.

I was there and remember it well. The whole thing was completely unexpected and people lost it. Well done. :)

Reply on news.ycombinator.com