[BUG] Local Whisper model is repeatedly re-downloaded instead of being recognized on disk

[mane81]

Summary

AudioWhisper 2.1.0 repeatedly re-downloads the local Whisper model instead of recognizing the already downloaded model on disk.

Environment

• AudioWhisper version: 2.1.0
• macOS version: macOS 26.3.1
• Hardware: Apple Silicon / ARM64

What I expected

After downloading a local Whisper model once, AudioWhisper should reuse the local files for transcription.

What happened

AudioWhisper reports the model download as finished, but when transcription is triggered it behaves as if the model is missing and starts downloading again.

Reproduction

1. Install AudioWhisper
2. Select local transcription
3. Select Whisper model tiny
4. Let the model download complete
5. Trigger transcription
6. AudioWhisper starts downloading the model again

Local state I verified

The model files are already present on disk and appear complete:

• ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/openai_whisper-tiny
• ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/openai_whisper-base

Observed sizes:

• openai_whisper-tiny: about 73 MB
• openai_whisper-base: about 143 MB

The tiny model folder contains expected files such as:

• config.json
• generation_config.json
• AudioEncoder.mlmodelc/...
• MelSpectrogram.mlmodelc/...
• TextDecoder.mlmodelc/...

App settings observed locally

defaults read com.audiowhisper.app shows:

• transcriptionProvider = local
• selectedWhisperModel = tiny

Extra debugging already attempted

I checked both paths referenced by the app behavior:

• ~/Documents/huggingface/models/...
• ~/.cache/huggingface/hub

I also created ~/.cache/huggingface/hub as a symlink to the existing model directory so both expected cache locations resolve correctly, but AudioWhisper still re-downloads the model.

Why this looks like a bug

The model files already exist and look complete, but the app does not persist or re-recognize the downloaded state. This looks like a model detection / cache recognition bug rather than a failed download.

[m13v]

this is almost certainly a path mismatch in the model validation logic. WhisperKit caches models under a specific directory structure and the app is probably checking a different path than where the download actually lands.

we ran into the same thing building a WhisperKit-based transcription service. the issue was that WhisperKit's ModelFileManager resolves the cache path using the model variant name (like openai_whisper-tiny) but the app's "is model downloaded" check was looking at a slightly different directory hierarchy. so the files existed on disk but the validation function did not find them because it was checking a parent directory or using a different path resolution.

worth checking: does the app use WhisperKit's built-in modelFolder() resolution, or does it construct the path manually? if it is manually constructed, that is likely where the mismatch lives. also the symlink workaround you tried would not help if the validation is checking file existence relative to a different root than what the symlink resolves to.

[m13v]

the WhisperKit model loading and cache path resolution we use: https://github.com/m13v/fazm/blob/main/Desktop/Sources/TranscriptionService.swift