Dutch language support

#1

is anyone considering implementing support for the Dutch language? I’m involved in a project which will need Dutch in the end. I’m also talking to one of the creators of a large Dutch STT effort soon.
does anyone have a plan ready for introducing a new language? I may be able to help with Dutch and I want to know where this fits in.

3 Likes
#2

I do not know the nature of your project, but you might consider Rhasspy:

It is based on the good old jasper and then some! I am currently using it with Dutch language, a must for me.

#3

Any update on this request ? Dutch support is highly desired.

#4

I have switched to Rhasspy for this.

#5

I have also switched to Rhasspy. For other reasons (it’s completely open source!), but the support for Dutch is a nice bonus. Note that although STT works perfectly for Dutch, the default espeak TTS sounds quite robotic for our language. @Romkabouter has added Google WaveNet support to Rhasspy, which supports Dutch, but that’s not an offline solution and hence not one that I’d use. Rhasspy’s developer is investigating adding MaryTTS support for Dutch, though, so maybe the Dutch TTS options in Rhasspy will improve in the near future.

1 Like
#6

And for a more direct answer to your question: I recently interviewed Snips CTO Joseph Dureau for the Dutch magazine MagPi (I’m the editor-in-chief) and I asked him about Dutch language support. Because of your interest in this, I just published the interview on our website:

I’m quoting the relevant part:

Wanneer kunnen we in het Nederlands met Snips spreken?

Dat zal in 2019 niet meer gebeuren, vrees ik, maar zeker in 2020. We willen tegen volgend jaar 50% van de talen in de wereld ondersteunen, en zeker de belangrijkste Europese talen. Momenteel herkent onze assistent Frans, Engels, Duits, Spaans, Italiaans en Japans.

Een nieuwe taal ondersteunen is 3 tot 6 maanden werk. Een expert in die taal dient de grammatica te schrijven, een akoestisch model te trainen enzovoort. Het is best ingewikkeld.

2 Likes
#7

I bit of nuance is at its place, it is actually an almost offline solution.
A unique sentence, defined by language, wavenet voice and samplerate, is cached. The second time the sentence is called, the cache file is played.
This is probably in a lot of cases sufficient, except for a lot of random sentences.
So, depending on how much randomness you need, it is indeed an offline TTS