Parakeet is a speech-recognition model family whose TDT checkpoints combine a FastConformer encoder with a transducer decoder. Parakeet-TDT 0.6B v2 is English-only, while v3 expands support to 25 European languages. Both have 600 million parameters and provide punctuation, capitalization and timestamps. Local runtimes include FluidAudio on Apple platforms and sherpa-onnx for cross-platform deployment.
The v2 card describes a FastConformer encoder with a TDT decoder, trained with full attention for English transcription. It supports word-level, segment-level and character-level timestamps. The v3 card describes FastConformer-TDT with automatic language detection across 25 European languages, plus word-level and segment-level timestamps. Both cards describe automatic punctuation and capitalization. Their model licenses are CC-BY-4.0, separate from the licenses of local runtime libraries.
The v2 card reports average WER 6.05% across its Hugging Face Open-ASR-Leaderboard evaluation datasets and RTFx 3380 with a batch size of 128. The v3 card reports average WER 6.34% across nine benchmark datasets and RTFx 3,332.74. These are the cards' reported figures, not a controlled comparison with Whisper or measurements of interactive dictation latency on a consumer laptop.
FluidAudio is a Swift SDK using Core ML and the Apple Neural Engine; its catalog includes Parakeet TDT v2 and v3. sherpa-onnx uses ONNX Runtime and documents converted int8 packages for both checkpoints across platforms. The runtime licenses are Apache-2.0, while the model cards specify CC-BY-4.0. Keep the checkpoint version and conversion format explicit when choosing a package for your application.
WhisperJot's Jot Local uses Parakeet-TDT 0.6B v2, the English-only model, for fast short utterances. On macOS it runs through FluidAudio with Core ML and the Neural Engine; on Windows and Linux it uses sherpa-onnx with an int8 ONNX package. Downloads are approximately 500 MB on Mac and approximately 600 MB on Windows and Linux. Auto chooses between Jot Local and Whisper-based Jot Local Pro per utterance. Local inference works offline after download.
Explore WhisperJot →Parakeet-TDT 0.6B v2 is English-only. The v3 card expands support to 25 European languages and describes automatic language detection. WhisperJot uses v2 for Jot Local, which therefore remains English-only. Language coverage depends on the exact checkpoint version selected in your application.
Both v2 and v3 cards describe word-level and segment-level timestamps. The v2 transcribe API also supports character-level timestamps. Applications choose which of these timestamp formats to expose. Check your application's output options to see which formats you can actually use or export.
Yes. FluidAudio provides local Apple-platform inference, and sherpa-onnx provides an offline ONNX Runtime path with converted Parakeet packages. WhisperJot uses those runtimes for Jot Local on its respective platforms and works offline after the one-time model download. Complete that setup while connected.
Primary documentation checked 2026-09-05. Specifications and runtime requirements can change.
Compare Whisper and Parakeet by model size, language scope, architecture and license, with task-based guidance and WhisperJot's per-utterance Auto mode.
Understand Whisper large-v3-turbo: fewer decoder layers, reference speed, translation limits, compressed variants and its role in WhisperJot Jot Local Pro.
Plan local Whisper transcription with the official VRAM table, whisper.cpp and Python routes, or WhisperJot's desktop models and honest download sizes.
See also: Parakeet glossary definition · System requirements · Pricing
