WhisperAX: force closes on file upload when Audio Encoder/Text Decoder set to GPU

[marknomarky]

Device: iPhone 13
OS: iOS 18.4.1
WhisperAX v. 0.7.0 (2)
SDK v. 1.2.0

Narrative: when attempting to set Audio Encoder and Text Decoder both to GPU and uploading a file to transcribe (file format: M4A, file size: 8.9 MB), the app loads the file (at least to the point where it shows the length of the audio in seconds) and then immediately crashes. Not sure if this is expected behavior at this point or if it is a concern.

NOTE: The attached file sounds as though it's a conversation between a doctor and a patient - this is intentional for our use case but only exists as example data and does not represent actual doctor/patient interaction or any PII.

I had to zip the file as Github will not allow M4A attachments

DocCord-example.zip

[ZachNagengast]

Do you recall which model you were using for this transcription? I tested on a few devices with base and wasn't able to replicate. In general though the GPU can sometimes cause issues on certain iOS versions (CoreML bug). Can you try using the Neural Engine as well to see if it resolves the issue? That would narrow it down a bit.

[marknomarky]

This would be on whisper-large-v3-v20240930 and switching back to Neural Engine does resolve the issue.

[ZachNagengast]

Understood, thanks for trying, will investigate further. In the meantime, you could try building the example app here to test with a debuggable environment.