Hacker News Reader

Comment by yladiz | Hacker News Reader

This is pretty cool (although, I have no idea what other technologies exist for this kind of thing), but it's definitely not convincing enough to a human listener. This sounds like it might be convincing enough for some programs like "Hey, Siri" but it's not gonna convince your mom. You can listen to the samples on the page linked here and you can immediately tell that Obama and Trump don't sound quite human.

View on HN

Well, the question is, do they just need to throw more computational power / training at this algorithm or is that the peak of their implementation?

This is something Google has been working a lot on [1] and Baidu also recently posted about their results too [2]. We're definitely pretty close to passing the human detectable level.

[1] https://deepmind.com/blog/wavenet-generative-model-raw-audio...

[2] http://research.baidu.com/deep-voice-production-quality-text...

Google and Baidu have only demonstrated single speaker TTS. Lyrebird's the first to demonstrate being able to generate arbitrary voices. Since this came out of a research lab, I would guess that the quality would only improve if they are given more compute and data.

The Google one is able to generate arbitrary voices based on the recordings used for training. So much so that they made it generate piano music.

Maybe. However, most CGI (in big Hollywood productions with almost no budget limit) is still very detectable, so much so that the biggest productions try to do as much as possible with real props and real actors instead of CGI (cf. interviews about the last Fast and Furious movie).

The human mind seems to be better at this than most creators credit it for.

I believe what makes the voices robotic is due to the little amount of audio they need to generate a "usuable" voice from the system.

Speech models usually use triphones, which turns out to be a huge amount of audio. This is particularly impressive because of how little data they need.

Google used their own datasets, which are most likely massive.

It might become more convincing if audio engineers would edit the results to hide artifacts and make it sound more natural.

Devils advocate: the noise masks the distortion, which is the giveaway.

Text to speech is still pretty distinguishable as not-human, and that seems like an easier problem (only has to work for one specific voice, not an arbitrary voice). So just on the basis alone I wonder if this isn't still a ways out

Interesting thought: is it easier or more difficult to make a synthetic voice undistinguishable from a human one, compared to producing speech copying a real voice?

To me, at least, the voices I heard in Lyrebird's demo actually sounded more 'real' than Microsoft Sam for example.

Of course, the voices produced by Lyrebird sound more "real" than Microsoft Sam, since I would define realness as sounding humanlike. However, I would strongly prefer Microsoft Sam over something generated by this algorithm for general use because this algorithm produces voices that are still in an uncanny valley, because it is almost human but not human enough, whereas Microsoft Sam is obviously not human.

I would argue that it depends on the sample rate. Over the telephone there are several TTS voices that are very convincing because the audio quality is lower.

Yet.

Reply on news.ycombinator.com