Does anybody know what is approach implemented in Google Speech API and Siri for Turkish voice to text speech recognition?

I'm interested in details of these two services, not the general or theoretical research. The reason is that this is an agglutinative, and I doubt that the same techniques that were successfully implemented for English would work here well.


If you actually try to speak a very long Turkish word, for example Afyonkarahisarlılaştırabildiklerimizdenmişsinizcesine, is it recognised?

Today's speech recognition is end-to-end, that is, the input in training is just the audio and the transcript. There is ideally nothing language-specific in that part. It can be done, for example, with seq2seq architecture.

It can be augmented with unsupervised models - an acoustic model trained on raw audio with no transcript and/or language model trained on raw text with no audio - and ideally those also do not have any language-specific components, but they may have language-specific parameter values.

The language with this structure that drives much of the research is Japanese. Agglutination and generally morphology, including compound nouns that are written together in languages like English, can be addressed with character-level models, as opposed to strictly word-level models.

So the language model parameters for the minimum and maximum character n-grams to learn can be different for different languages. And because language models are used in many if not most tasks - and because agglutination is an issue in most English-specific approaches to most tasks - that is not specific to speech recognition.

Looking through https://research.google.com/pubs/HasimSak.html and https://research.google.com/search.html?q=kaisuke gives you a concrete idea of the experiments. However there are engineering reasons not to maintain separate components or even approaches for each language when you are supporting more than a hundred languages, and many of the additions forced by agglutinative languages like Japanese, Turkish, Hungarian and Finnish are also useful innovations if generalised to all other languages.

And it does not work perfectly yet.

| improve this answer | |
  • 1
    I sent a few sample 50 second clips to Google Speech API trial. I used movies in YouTube, and put the mic in front of a speaker to record. The results were incomprehensible garbage. I'm starting to suspect that they did not develop a custom solution for Turkish, and like you suggested used common engines. – Aksakal almost surely binary Mar 24 '18 at 16:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.