Repository navigation
Replies: 1 comment
|
Answering my own question, since I was able to test it. Short version: yes, it works on Measured on a GitHub-hosted Install and import. The boosting tree. A full boosted decode.
On six invented street names spoken by macOS One caveat on that +17.4%. My first attempt timed one unboosted pass and then one boosted So, to my original three questions:
Small docs nit: |
Uh oh!
There was an error while loading. Please reload this page.
I'm evaluating
parakeet-tdt-0.6b-v3with the phrase-boosting (GPU-PB) added in v2.5.0 .The constraint is that the production target is an Apple Silicon Mac, not NVIDIA hardware.
malsd_batchdecoding withboosting_tree.key_phrases_file— run on macOS with MPS, or CPU-only, at usable speed? Oris NVIDIA hardware effectively required in practice?
the batched decoding it plugs into. If that's right, is a CPU-only or MPS configuration
supported, even if slower?
exported model that keeps the biasing behaviour? (
mlx-audio's Parakeet implementationappears to have no biasing code, so a straight export seems to lose it.)
Context: I'm following #14772, where phrase boosting was confirmed working for
parakeet-tdt-0.6b-v3. This is purely about whether that path is reachable off NVIDIA hardware. My most important question is does Apple work ?All reactions