3 ms·

No. Translation is not easy. I speak Norwegian. There are about five million language users. The Norwegian median wage is perhaps the highest in the world, pos

by vages 6y ago

No. Translation is not easy. I speak Norwegian. There are about five million language users. The Norwegian median wage is perhaps the highest in the world, possibly the top fifth, so there is a lot of pull to have Norwegian users. But only a few voice assistants have Norwegian interfaces (Siri was the first, Google may have joined the crowd). I work in IT, and making a conversational interface is rarely on the table; never if you exclude chat bots. To get an idea of the work involved, I know the guy who translated Siri through friends. Allegedly, that was at least six months of a single person working. I don't know if he only did the translation or if he also did a lot of AI work, but that's a lot in either case. The result isn't too impressive either. Siri has problems understanding my girlfriend because of her Bergen dialect, which differs from my Oslo dialect mainly in how the R sound is pronounced. My two cents.

blueblisters 6y ago
A good speech recognition model takes more than 10k hours of labeled speech data to train. The more diversity the better. Most production speech models (Siri, Google) need some domain-specific training as well - for offline or low-resource inference. Additionally, language models that help improve speech recognition accuracy need a fairly decent amount of unsupervised text as well. If you are building a conversational speech model, also consider the data needed to label intents for a large number of phrases. For a language like Norwegian, you are unlikely to find so much labeled and unlabeled data without forking out a large amount of money to literally ask some Norwegians to help. I suspect the work is being done, but it's extremely time consuming and researchers are likely spending time building models that can learn high-quality high-level representations of language that can be transferred to different scripts/languages with a small amount of data.

by vages 6y ago

blueblisters 6y ago

A good speech recognition model takes more than 10k hours of labeled speech data to train. The more diversity the better. Most production speech models (Siri, Google) need some domain-specific training as well - for offline or low-resource inference. Additionally, language models that help improve speech recognition accuracy need a fairly decent amount of unsupervised text as well. If you are building a conversational speech model, also consider the data needed to label intents for a large number of phrases. For a language like Norwegian, you are unlikely to find so much labeled and unlabeled data without forking out a large amount of money to literally ask some Norwegians to help. I suspect the work is being done, but it's extremely time consuming and researchers are likely spending time building models that can learn high-quality high-level representations of language that can be transferred to different scripts/languages with a small amount of data.