Local Models

The local engine transcribes on your own machine: no API key, nothing uploaded, and it works offline once the model file is downloaded. There are two of them, and you pick one simply by selecting its model in the list — Parakeet, built for the CPU and the faster of the two, or Whisper, which covers far more languages.

Managing models

Open Settings → Transcription and select the Local Model (On-Device) engine. The Available Models list holds both families, Parakeet first, with All / English / Multilingual filter tabs. Each row has a Download button, a radio button that makes the model active once downloaded, and a trash icon ("Delete model") to free the disk space. Selecting a row switches the engine with it — there is no separate engine setting to keep in sync. Model files come from Hugging Face. An interrupted download keeps what it already received and continues from there — on the automatic retries (up to 4 attempts) and the next time you press Download, even after restarting the app; a download that stops receiving data for 60 seconds is retried. Each file is checked against the SHA-256 checksum Hugging Face publishes before it is used. Cancelling a download discards it.

Parakeet

Parakeet V3 (Multilingual) — parakeet-tdt-0.6b-v3-int8, 670 MB. A NeMo transducer run through sherpa-onnx, in 25 European languages with automatic detection: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. Punctuation and capitalisation are part of its output.

Two things make it the fast option. It runs on the CPU — no Metal, no Vulkan, no graphics driver to depend on — and the model stays loaded between dictations instead of being read from disk for each one. Measured on an Apple Silicon Mac against large-v3-turbo-q8_0 running on Metal: a 2.9-second dictation took 71 ms instead of 1347 ms, and 2.6 minutes of audio took 7.0 s instead of 11.7 s. The model is loaded once, which takes about a second, when you select it, when it finishes downloading, or when the app starts.

What it does not do: it provides no word-level timestamps, so when a meeting or a media file is transcribed with Parakeet and speaker identification is on, Speakly says that mixed passages may stay unidentified and lets you correct them by hand. Whisper's extra languages are not there either — outside the 25 above, use Whisper.

Whisper

Whisper runs through whisper.cpp and covers 99 languages. It is the right choice when your language is not one of Parakeet's, and the more accurate of the two on hard audio.

English-only

• Medium (English) — medium.en, 1464 MB

Multilingual (99 languages)

• Medium (Multilingual) — medium, 1464 MB
• Large V3 (Multilingual) — large-v3, 2953 MB
• Large V3 Turbo (Multilingual) — large-v3-turbo, 1617 MB

Quantized

• Large V3 Turbo Q5 (Quantized) — large-v3-turbo-q5_0, 574 MB
• Large V3 Turbo Q8 (Quantized) — large-v3-turbo-q8_0, 874 MB — the model a fresh install starts with on a machine with a GPU

There is no large-v3-turbo-en: the large and turbo models exist in multilingual form only.

GPU acceleration

GPU acceleration applies to Whisper only, and it means Metal, on Apple Silicon, and nowhere else. Speakly enables it when the platform is macOS on arm64. On Intel Macs, Windows and Linux, Whisper runs on the CPU — there is no Vulkan and no CUDA build. Parakeet never needs it.

• Apple Silicon (M1 and newer): the panel reads "Apple Metal detected! Local transcription will be 3x faster on your device." next to an Enable checkbox, on by default
• Everything else: the panel points you at Parakeet V3 as the fastest option on that machine, and no toggle is shown

Style and Dictionary

Neither local engine accepts the Style or Dictionary hints: whisper.cpp's prompt option is not wired in the bundled addon, and Parakeet is a transducer with no prompt input at all. Both features apply to the cloud engines, and Settings dims them while a local engine is selected. Post-processing rules still run on the text afterwards, whichever engine produced it.

Licence

Parakeet TDT 0.6B V3 is published by NVIDIA under CC-BY-4.0, which allows commercial use and requires attribution. Speakly shows that attribution next to the model in Settings and links to the model card.