Hacker News Reader

What about Kaldi? Do you not consider that a viable option?

I made an embarrassing amount of money back when I was a consultant, just helping people get Kaldi working in their production environments or small tweaks to the models. I really don't consider it viable unless you're in the field, and it's especially not viable in a production environment.

That’s a shame to hear, I’ve always has high hopes for it.

What makes it less suitable for production compared to TF?

Is it just the Kafkaesque build process (admittedly I last tried a year ago), or is model-fitting or prediction especially buggy?

Totally agree on that. For example, imho it is getting a bit ridiculous that Linux has a driver for just about any text-input device out there, except speech.

Speech input is not text input. I don't understand what you mean.

He means speaking commands (to a shell I presume) rather than type them.

I haven't used it, but I think I read that they have 6-10% error rate?