new realistic TTS service from Google

[mcthulhu]

Google is offering beta Cloud Text-to-Speech via the Google Cloud Platform at https://cloud.google.com/text-to-speech/, supporting natural-sounding WaveNet voices powered by DeepMind AI. WaveNet, instead of stringing together syllables, "analyzes the waveforms from a huge database of human speech and re-creates them at a rate of 24,000 samples per second," per https://www.theverge.com/2018/3/27/1716 ... nd-wavenet. There are 32 voices in 12 languages and variants, including English, French, Spanish, German, Dutch, Brazilian Portuguese, Turkish, Japanese, and Swedish. The supported voices are listed at https://cloud.google.com/text-to-speech/docs/voices. The generated speech can be saved to .mp3 files, etc.

The free tier of usage through the GCP API is up to 1 million characters per month, which sounds to me like more than enough for personal use. I think this is going to be fun to play with.

[rdearman]

Very timely post. I was using AWS Polly for something I needed. I'll see if this gives me more natural speech.

EDIT: WaveNet support seems to only implemented for English. It isn't available for French.

[mcthulhu]

That's too bad - I must have misread the list of supported languages. Thanks for the correction. The list does say that it includes both standard and WaveNet voices. I hope they expand the WaveNet support soon.

I've been pretty happy with AWS Polly so far, though.

[Kubelek]

https://cloud.google.com/text-to-speech/docs/voices

Wavenet is now available in many langauges, and it is very impressive imho. Quality varies but Polish is uncanny valley and much better than older commercial engines.
Pricing is reasonable, with 1 million characters for free monthly, and 16$/million above the limit. The standard TTS is also very good and much cheaper.

https://ankiweb.net/shared/info/814349176 a modified Awesome TTS addon already makes use of this functionality. Whoever made the addon estimates that the free limit is enough for 23 hours of audio every month (with some limits on the size of an individual query )

This could be great for languages with few audiobooks - EuPortuguese, I'm looking at you!

[tussentaal]

I find Nuance/Vocalizer voices more natural-sounding.