[Bug]: Parakeet v3 times out during load and startup

[farooqu]

System Information

Pindrop: 1.22.5 (26)
macOS: Version 26.5.1 (Build 25F80)
Chip: Apple Silicon
Model: parakeet-tdt-0.6b-v3

What happened?

Parakeet v3 consistently failed to load during startup. Pindrop recognized the model as installed, but loading remained stuck until the 120-second timeout expired. Recordings afterward failed with modelNotLoaded.

I cleared Pindrop’s Application Support directory and downloaded the model again from scratch, but the same timeout still occurred.

Expected Behavior

After Pindrop downloads and recognizes a complete Parakeet model, startup should load that existing model without downloading a second copy into another Application Support directory.

Steps to Reproduce

1. Select and download parakeet-tdt-0.6b-v3.
2. Confirm that Pindrop reports the model as downloaded.
3. Quit and relaunch Pindrop.
4. Wait for the model to load during startup.
5. Model loading times out after 120 seconds.
6. Attempting to record afterward fails with modelNotLoaded.

Log File Attachment

pindrop-2026-07-29_14-28-01-1.log

Permissions

• Microphone access
• Accessibility access

[farooqu]

I investigated this further and confirmed that Pindrop was using two separate Parakeet cache locations.

The complete model downloaded by Pindrop was located at:

~/Library/Application Support/Pindrop/FluidInference/parakeet-coreml/parakeet-tdt-0.6b-v3

It was approximately 461 MB and contained all five required artifacts:

• Preprocessor.mlmodelc
• Encoder.mlmodelc
• Decoder.mlmodelc
• JointDecisionv3.mlmodelc
• parakeet_vocab.json

During model activation, however, FluidAudio used a different location:

~/Library/Application Support/FluidAudio/Models/parakeet-tdt-0.6b-v3

To confirm this, I quit Pindrop and moved the existing Parakeet directory out of FluidAudio/Models. I then launched Pindrop once while observing that directory.

At approximately two seconds, Pindrop created the parakeet-tdt-0.6b-v3 directory. After 30 seconds, it contained only about 24 MB:

• Decoder.mlmodelc was approximately 24 MB.
• Encoder.mlmodelc was effectively empty/incomplete.
• Preprocessor.mlmodelc, JointDecisionv3.mlmodelc, and parakeet_vocab.json were missing.

No other model directory was created.

This showed that although Pindrop had already downloaded and recognized a complete model in its own cache, model activation started downloading another copy into FluidAudio’s separate cache. That second download did not finish before the 120-second model-loading timeout.

As a local workaround, I copied the complete model from Pindrop’s cache into the runtime location FluidAudio had created:

~/Library/Application Support/FluidAudio/Models/parakeet-tdt-0.6b-v3

After copying the complete model there, the same Pindrop 1.22.5 application loaded Parakeet successfully in 12.03 seconds:

TranscriptionService.loadModel engine.loadModel task finished elapsed=12.03s

Model loaded successfully with Parakeet engine

No application code or settings were changed. The only change was placing the already-complete Pindrop model into the separate FluidAudio runtime cache.

[farooqu]

@watzon would you be open to a contribution for this bug?

[n2k3]

I experienced the same issue and confirmed the workaround works:

1. Quit Pindrop completely.
2. Renamed the empty runtime directory:
   ~/Library/Application Support/FluidAudio/Models/parakeet-tdt-0.6b-v3
3. Copied the complete model from:
   ~/Library/Application Support/Pindrop/FluidInference/parakeet-coreml/parakeet-tdt-0.6b-v3
   into the original suffix-less FluidAudio runtime path.
4. Relaunched Pindrop and activated Parakeet from the Models tab.

Afterward, Parakeet loaded successfully in 14.05 seconds and completed a test transcription successfully. No settings, preferences, recordings, or other caches were cleared.

[voodoomatz]

I can confirm the issue as well as the fix.
However, just a symlink instead of copying would be sufficient.