Large model giving terrible transcripts.

[DhalgrenAurele]

trafficstars

I've been using a script in terminal to transcribe 1-3 minute .wav files, and it's been really annoying, but perfectly accurate. every transcript flawless.

Macwhisper, using the same "large" model, is getting maybe 1 word per sentence correct.

Using the same model on both, all ran locally, why does macwhisper get it so wrong? It also uses about double the resources.

I'm using an intel MacBook Pro for reference. anyone else running into this?

[mechiland]

Same here #2037

I feel that the ggml-large-v2.bin and ggml-large-v3.bin has been polluted. I've tested ggml-large-v1.bin works fine.

[panagiotidi]

Same, just tried ggml-large-v3 and the transcription is practically gibberish