Parakeet

About

Parakeet uses NVIDIA’s Parakeet TDT model for extremely fast on-device speech recognition on Apple Silicon. The compact 0.6B parameter model delivers strong accuracy across 25 European languages with word-level timestamps, all without requiring a cloud connection.

Features

• NVIDIA Parakeet TDT - extremely fast on Apple Silicon
• 25 European languages supported
• Live streaming preview and word-level timestamps
• No API key or cloud account needed
• Compact model size (~600 MB)
• Optional local vocabulary boosting from your TypeWhisper dictionary
• Import of your own Parakeet TDT models converted to Core ML, for example a fine-tune for one language

Transcription Models

Model	ID	Size	Languages
Parakeet Ultra (default)	parakeet-ultra	~630 MB	25 European languages
Parakeet TDT v3 0.6B	parakeet-tdt-0.6b-v3	~600 MB	25 European languages
Fluid Speech Mini	fluid-speech-mini	~250 MB	English
Parakeet TDT v2 0.6B	parakeet-tdt-0.6b-v2	~600 MB	English

Fluid Speech Mini needs version 1.5.0 of this add-on and TypeWhisper 1.8. Every model needs 8 GB of memory or more.

Configuration

No API key required. Parakeet runs entirely on-device.

If vocabulary boosting is enabled, Parakeet uses your custom TypeWhisper dictionary terms locally during transcription. Dictionary corrections are still applied automatically after transcription, so the dictionary helps both during and after the transcription step.

Setup

1. In TypeWhisper, open Settings > Discover plugins and install Parakeet. The add-on then appears in the Settings sidebar, and its page opens.
2. Choose a model, for example Ultra for multilingual dictation or Fluid Speech Mini for a smaller English download, then download and load it in the plugin settings.
3. Select Parakeet as your transcription engine in Settings or in a profile.

Import your own model

Besides the built-in models, Parakeet loads Parakeet TDT models that you bring yourself, for example a fine-tune for one language converted to Core ML. Importing needs version 1.7.0 of this add-on and TypeWhisper 1.8.2 or newer.

1. In the Parakeet settings, select Import Model….
2. Choose Hugging Face and paste the repository URL, or choose Local Folder and select the folder with the model.
3. Select Import. TypeWhisper copies the model, compiles it, loads it, and transcribes a short test before the import completes.

The folder or repository needs these parts. Each model can be a compiled .mlmodelc or an .mlpackage, which TypeWhisper compiles during the import:

Part	Accepted names
Preprocessor	Preprocessor, parakeet_preprocessor
Encoder	Encoder, parakeet_encoder
Decoder	Decoder, parakeet_decoder
Joint decision	JointDecisionv3, JointDecision, parakeet_joint_decision_single_step
Vocabulary	parakeet_vocab.json, vocab.json, or vocab.txt

This covers the layout of FluidInference’s Core ML repositories, such as FluidInference/parakeet-tdt-0.6b-v3-coreml, and the folder that FluidInference’s mobius converter writes. TypeWhisper copies only the files it needs.

mobius writes no vocabulary. For a model with the standard Parakeet vocabulary, TypeWhisper adds it during the import. A fine-tune that adds tokens needs its own vocabulary: copy vocab.txt or vocab.json from the original model into the folder before you import it. For example, the Greek fine-tune myned-ai/parakeet-tdt-0.6b-v3-greek adds 33 tokens, and its vocab.txt is in its Hugging Face repository. If the vocabulary is missing, the import stops with a message before it copies or downloads the model files, for a local folder and for a Hugging Face repository.

Supported are Parakeet TDT models with the vocabulary size of v2 (1,024 tokens) or of v3 and its extensions (8,192 tokens or more). Other speech recognition architectures cannot be imported into Parakeet.

Imported models appear below the built-in models, and you choose them like any other Parakeet model. The trash button deletes TypeWhisper’s copy of a model; your original files stay unchanged. TypeWhisper stores imported models in ~/Library/Application Support/TypeWhisper/PluginData/com.typewhisper.parakeet/custom-models.

Manual copy workaround

If Parakeet is already working on one Mac, you can also copy the local model files over as a workaround. This is separate from plugin installation.

1. Quit TypeWhisper on both Macs.
2. On the Mac where Parakeet already works, copy only the model folders you actually use:
   • ~/Library/Application Support/FluidAudio/Models/parakeet-tdt-0.6b-v3-coreml/
   • ~/Library/Application Support/FluidAudio/Models/parakeet-tdt-0.6b-v2-coreml/ if you still use the older v2 model
   • ~/Library/Application Support/FluidAudio/Models/parakeet-ctc-110m-coreml/ if you use vocabulary boosting
3. On the target Mac, install the same or a newer TypeWhisper version, launch it once, then quit it again.
4. Copy those folders into the exact same paths on the target Mac under ~/Library/Application Support/FluidAudio/Models/.
5. Reopen TypeWhisper and check whether the model is detected and can be loaded.

Install from file is for plugin bundles. It does not import Parakeet model folders for you.

This is a workaround, not the polished primary flow. If the target Mac blocks access to ~/Library/Application Support or CoreML cache usage, the copied model can still fail to load.