Microsoft

How can we improve Video Indexer?

Too Low Accuracy of Language Detection with auto-detect mode

Too low accuarcy that Video Indexer can't detect its language with auto-detect mode like this:
https://www.youtube.com/watch?v=kRQnArIIffw
It's Japanese but Video Indexer defines this as en-us.

2 votes
Sign in
(thinking…)
Password icon
Signed in as (Sign out)

We’ll send you updates on this idea

Ayako shared this idea  ·   ·  Flag idea as inappropriate…  ·  Admin →

1 comment

Sign in
(thinking…)
Password icon
Signed in as (Sign out)
Submitting...
  • Shay Ben-Elazar commented  ·   ·  Flag as inappropriate

    Hi Ayako,
    I am a scientist working on the automatic Language Identification feature. Thank you for sharing this feedback! We are continuously improving our models and view these comments as an opportunity to better calibrate our models.
    In this case, our production model is predicting Japanese correctly with confidence 0.594 - slightly below the 0.6 confidence threshold where we report a result. Our fallback is to return En-US as the detected language in this case. To retrieve the (verbose) prediction regardless of the threshold you can use the ArtifactUrl API (https://api-portal.videoindexer.ai/docs/services/operations/operations/get-video-artifact-download-url)
    Additionally - our latest model (on route to production) is predicting Japanese with a confidence score > 0.7 on this sample, which will be reported in the video breakdown.

    Please let us know if there is anything else we can do to help.
    Best,
    Shay

Feedback and Knowledge Base