Streaming / real-time usage? about openl3 HOT 1 CLOSED

In principle this should be possible following the approach you described - load the model (to avoid having it re-loaded all the time) and pass it to get_audio_embeddings (new API about to be merged into master with separate methods for audio and image embeddings).

Regarding H/W: it depends on the desired hop size you plan to use. If you plan to use non-overlapping windows (i.e. just 1 prediction per second) it should be totally doable - I have a real-time demo running a similarly sized CNN (albeit in... wait for it... lasagne!) which runs happily on a macbook pro with CPU processing (no GPU required). With a GPU you'll be laughing. Just have to get the multiprocessing right to separate audio buffering, prediction, and UI updates into separate processes.

Closing this out as this isn't actually an issue, but feel free to re-open if you encounter any specific problems.

from openl3.