Language Detection and Accuracy: V1 vs. Roadmap

[dmxplx]

Hi team,

I took good note that features such as voice assistants, voice commands, desktop control, translation, and automatic language detection are outside the scope of the initial project. With that in mind, I’m not quite sure whether my experience is a bug or something that should be discussed.

As a native speaker, I was impressed with the accuracy in French. However, when I switch to English, the accuracy is noticeably lower for me, despite having practiced the language daily for years. For example, when test-dictating www.ubuntu.com, I can get various incorrect outputs such as www.youbuntu.com, www.oubunt2.com, etc.

Does that mean that automatic language detection is not a "nice to have" but rather a must-have, for accuracy?

I’m curious to hear if other non-English native speakers have experienced similar issues, and to hear Jean-Baptiste’s POV on this who is also AFAIK fluent i both languages. If I’m only an edge case in this discussion, then no issues at all.

Thanks for making Ubuntu even more accessible!

[quine00]

Hi Dimitri,

Thanks for your feedback and testing, much appreciated!

I think you're exposing the lack of a vocabulary feature (see #1), rather than automatic language detection failing for you? I hope for instance that you get generally good results with accented English on non-technical dictation tasks.

Even with British English as my mother tongue, technical terms like those found on the Ubuntu homepage are often not transcribed correctly. There's methods to improve this that we haven't started implementing yet.

[dmxplx]

@quine00 thanks for the feedback.

[dmxplx]

@quine00 Just to clarify please, based on the diagram in Discourse, could we expect the local inference component to improve speech-to-text performance locally across multiple languages over time? https://ubuntucommunity.s3.us-east-2.amazonaws.com/original/3X/6/4/6479f07aceec10d6c9bc9642e9d11286f450707d.png Thanks