FEAT: Parakeet compatability?

[DPBattaglia]

Can't get parakeet v2 or v3 llm to work.

the below error is what I get when I start the server with model path: "mlx-community/parakeet-tdt-0.6b-v3"
maybe you can add a model type: "parakeet" instead of "whisper" and remove reference to "sample_rate" being used in the backend?

Unsure if something like this could be included for use with the MLX version of parakeet noting it includes diarisation - (openai api compatible with parakeet) - https://github.com/jfgonsalves/parakeet-diarized

2025-10-16 15:14:16 | ERROR | app.handler.mlx_whisper:_process_request:215 | ✦ Error processing audio transcription request: ModelDimensions.init() got an unexpected keyword argument 'sample_rate'

2025-10-16 15:14:16 | ERROR | app.core.queue:_process_request:158 | ✦ Error processing request transcription-02ead7dc-9871-4922-a0af-e90aa8ccc4a5: ModelDimensions.init() got an unexpected keyword argument 'sample_rate'

2025-10-16 15:14:16 | ERROR | app.api.endpoints:create_audio_transcriptions:181 | ✦ Error processing transcription request: ModelDimensions.init() got an unexpected keyword argument 'sample_rate'

[DPBattaglia]

or better yet maybe one of these - not quite sure which is the better ones: -

1. https://github.com/senstella/parakeet-mlx
2. https://github.com/FluidInference/FluidAudio
3. https://github.com/Blaizzy/mlx-audio

Fluid audio looks pretty good.

[DPBattaglia]

I think looking into it FluidAudio would be a pretty cool way to go. However, I noticed this wrapper, if it could enable pretty quick integration, used to integrate parakeet-mlx in https://github.com/RamboRogers/mlx-gui/blob/05773388fc4c22d82045271437064ef7e52fc18c/src/mlx_gui/mlx_integration.py#L813

unsure if there is a way to tailor the audio/completions api to easily point to parakeet type rather then when using whisper.

Either way would really like to see this as parakeet in our experience is way better then the whisper alternative. We have been getting quite a few issues in the output repeating /hallucinating

[cubist38]

Hi @DPBattaglia , thank you for your recommendation! I’ll look into it when I have some free time — I expect to have it implemented in the next few days.