microsoft / NeuralSpeech

MIT License
1.37k stars 183 forks source link

when will NaturalSpeech open source? #40

Closed lucasjinreal closed 2 years ago

lucasjinreal commented 2 years ago

Just wonder does this NaturalSpeech: End-to-End Text to Speech Synthesis with Human-Level Quality paper will open source or not

YichongLeng commented 2 years ago

Since NaturalSpeech is deploying on Microsoft Azure, it will not be open-sourced soon.

LifeIsStrange commented 2 years ago

@YichongLeng Had Google not open-sourced BERT or even tensorflow and kept it for its "cloud" You Microsoft, us end users and other companies would be fucking miserable right now. Sharing AI progress is a key behaviour that is necessary for the greater good, for achieving actual progress. If you alter the group incentives by not sharing your advances, you risk creating a domino effect of an AI winter where everyone keep its advances to himself and eventually kill the entire field. Look I understand you might want to keep this model as an Azure exclusivity, but please engage to make it an exclusive for 1-2 year and then to open source it. Not doing so would be hypocrisy, and a whole field destructive behaviour. I don't think you would like to be remembered in history books as the company that started said domino effect. The proposed 1-2 year exclusivity period seems fair and quite standard.

lucasjinreal commented 2 years ago

This is normal, in the technique beginning, those companies like google and Microsoft trends to open source their work as more as possible because these work doesn't achieve a commercial product-level, so that they can share progress with academic frontiers. But once the technique becomes product-ready, they will using them to make money. And I don't think they gonna open source it, not because of it's hard to do, but they don't want anyone else has this capability, because this is the last step to the production-ready level.

LifeIsStrange commented 2 years ago

If you think ai research is reaching production quality performance, you are being delusional. Look at the leaderboards on paperswithcode, most tasks SOTAs have ~90% accuracy, in NLP that often means 1 error every 10 words AKA one error per sentence. We are at the demo stage of ai, and to reach higher accuracies, we have diminishing returns since there are less and less low hanging fruits. Advances will come more and more from integrating the work of others and making synergies, as it become increasingly harder for a single player to make progress, and as said, we are generally not at the productization stage for serious work.

lucasjinreal commented 2 years ago

@LifeIsStrange Hey, I am on your side. I just tell the truth and reality. I also feel very disgusted on Microsoft didn't open source this work. I am 100% sure their work must build upon many other open source work as well.

dreamflasher commented 1 year ago

@YichongLeng When will NaturalSpeech be available on Azure?