[BUG] 'Translate to English' has no effect with FluidAudio / Parakeet

[mane81]

Summary

Translate to English appears to have no effect when using the Parakeet / FluidAudio engine.

Environment

• App: OpenSuperWhisper
• Engine: FluidAudio / Parakeet v3
• macOS: macOS 26.3.1
• Hardware: Apple Silicon Mac mini

What I expected

When Translate to English is enabled, non-English speech should be translated to English during transcription.

What happened

With Parakeet v3 selected, transcription still returns the original language even though Translate to English is enabled.

Current settings observed locally

• selectedEngine = fluidaudio
• fluidAudioModelVersion = v3
• translateToEnglish = 1
• whisperLanguage = de

Reproduction

1. Open Settings
2. Select engine Parakeet
3. Enable Translate to English
4. Speak German (or another supported non-English language)
5. The output remains in the original language instead of being translated to English

Extra note

From local code inspection, this may be implemented only for the Whisper engine path:

• WhisperEngine.swift sets params.translate = settings.translateToEnglish
• FluidAudioEngine.swift does not appear to use translateToEnglish

So this may be either:

• a missing implementation for FluidAudio/Parakeet, or
• a UI/settings bug where the toggle is shown for an engine that does not support it

[m13v]

Ran into this same gap when switching between ASR engines in our transcription pipeline. Parakeet/FluidAudio models are ASR-only, they don't support translation as a task type the way Whisper does. Whisper has that neat translate task baked into its architecture, but CTC-based models like Parakeet are strictly transcribe-in-source-language. What worked for us was adding a lightweight translation step after transcription. We pipe the ASR output through a small on-device model for the translation pass. Not as seamless as Whisper's built-in translate, but the transcription quality from Parakeet is noticeably better for non-English languages, so the two-step approach actually gives better results overall. The UI should probably gray out the translate toggle when a non-Whisper engine is selected to avoid confusion.

[m13v]

Here is how we set up the transcription pipeline with engine-specific feature detection. The service checks model capabilities before exposing translate/transcribe options: https://github.com/m13v/fazm/blob/main/Desktop/Sources/TranscriptionService.swift