Skip to content
WM KeyboardWM Keyboard
Accessibility

Offline voice (Parakeet)

Dictation in 25 European languages on your phone, with NVIDIA's Parakeet TDT model telling them apart.

Offline Parakeet is a dictation engine that runs entirely on your phone, with NVIDIA’s Parakeet TDT 0.6B v3, one of the most accurate open speech recognition models of its size. It transcribes 25 European languages and works out for itself which one you’re speaking, so a layout in any of them can dictate without choosing a model per language. Once the model is downloaded, no audio leaves the phone.

The languages: 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. While a layout in another language is in use, dictation uses the system recognizer instead, the same way offline Whistle hands over its other languages.

WM Keyboard  Voice typing

Add Parakeet to the Recognition engine list and move it to the top, then press its row to open its screen and download a model from the Parakeet model group there. Each is large: 719 MB, the 4-bit build made for phones, and suits a recent phone with plenty of memory. On mobile data the download asks first. It comes from Hugging Face at a fixed revision, and both of its files are checked against their published SHA-256 before they count.

There are two models, the same size and speed:

  • Parakeet Ultra, Moondream’s further-trained copy of v3, with the same 25 languages. On Moondream’s own tests it gets about a sixth fewer words wrong across the 25 languages (FLEURS: 9.55% word error rate against v3’s 11.62%), and does better in background noise. The keyboard downloads its own ExecuTorch export of it.
  • Parakeet TDT 0.6B v3, NVIDIA’s original.

With both downloaded, dictation uses Ultra, and v3 can be deleted to free its space. Engine for each language can still give a language v3.

On a phone whose GPU supports Vulkan 1.1, the list also has Parakeet Ultra (GPU) and Parakeet TDT 0.6B v3 (GPU): the same models exported to run on the GPU, 819 MB each, with the same 4-bit weights and the same output. Downloading one also fetches ExecuTorch’s GPU runtime, a download of its own of about 13 MB. They are opt-in because the GPU is often the slower choice: on a phone with a Helio G99 and a Mali-G57 GPU they took 2.2 to 2.7 times as long as the standard builds (14 to 17 seconds for a 10-second clip, against about 5). Newer GPUs may do better, but none has been measured here.

A standard build of the same size wins over a GPU one, so a GPU build dictates only when it is the one you keep, or when you pick it for a language under Engine for each language. Until the GPU runtime is loaded (it loads with the keyboard, so the first time it can take a restart of the keyboard), dictation uses the standard build of the same model if that is downloaded, and otherwise reports the model as not loading.

On a phone with a Snapdragon 8 Gen 3, 8 Elite or 8 Elite Gen 5, the list also has Parakeet TDT 0.6B v3 (NPU, 10 s) and Parakeet TDT 0.6B v3 (NPU, 30 s): v3 compiled ahead of time for that chip’s NPU by the sherpa-onnx project, which publishes one build per chip. Each is one archive of about 440 MB from sherpa-onnx’s GitHub releases, checked against its SHA-256 and unpacked on the phone. A phone with another chip, including older Snapdragons that sherpa-onnx also builds for, does not list them, since the keyboard carries Qualcomm’s NPU libraries for those three chips only. There is no NPU build of Ultra.

An NPU build is compiled for clips of one length, 10 or 30 seconds. A shorter phrase is padded with silence to that length, so on the NPU a two-second phrase costs as much as a full one; a longer phrase is cut at its quietest moments into pieces that fit, and the pieces are transcribed one after another. With both downloaded, each phrase goes to the shorter build while it fits and moves to the 30-second one as it grows; the 10-second build alone cuts more often. A cut can land inside a word on fast continuous speech, which is the reason to keep the 30-second build for long dictation.

A downloaded NPU build is used over the standard builds, Ultra included. The first time, dictation also fetches sherpa-onnx’s NPU library (a few MB) and the phone’s NPU runtime, Qualcomm’s QNN libraries, the same ones on-device AI uses on the NPU; the GitHub edition asks you to accept Qualcomm’s licence before fetching them. Until both are on the phone, dictation uses the standard v3 if that is downloaded, and otherwise reports the model as not loading. The NPU builds have not been measured on a phone yet: the phone this was built on has a MediaTek chip, so how much faster they are, or whether a given phone’s Qualcomm drivers accept them, is still open.

Parakeet doesn’t write its answer a word at a time: an encoder reads a whole clip, and a small transducer then walks through it, writing a word piece where it hears one and skipping ahead where the model’s own duration head says nothing more is coming. That is most of why it’s quick for its size. A phrase can run to 50 seconds. The first phrase after the keyboard starts takes a few seconds longer while the model’s weights are prepared.

The words show while you speak, as they do with Asr4All and Sherpa: under the microphone, and in the field as underlined text in the default voice typing mode. Parakeet isn’t a streaming model, so the keyboard gets them by transcribing the phrase as it stands, again and again while it grows. A new pass starts once there is more than half a second of new audio, and never sooner than three times as long as the last pass took, so a slower phone shows the words less often rather than falling behind. The words can still change as the rest of the sentence arrives. When the phrase ends, one last pass over all of it gives the text that goes in, the same text Parakeet gave before it showed words as you spoke. Each pass covers the whole phrase, so the phone works harder during a long one; a pause that ends the phrase (see Hearing the end of a phrase) keeps phrases short.

The text arrives punctuated and capitalised by the model itself, and then goes through the same spacing, spoken punctuation and undo as any other dictation (see Voice typing).

WM Keyboard  Voice typing  Stop after a pause

A phrase ends when you stop talking for as long as Stop after a pause says, heard by the Phrase end detector you picked (see Hearing the end of a phrase). Both are also in the End of a phrase group on Parakeet’s own page, under its models. Stop after a pause is off by default; while it is off, a phrase only ends when you press the microphone or it reaches 50 seconds. Set it to a second or so and each phrase goes in at the pause, and with Keep listening on the next one starts by itself.

Parakeet has no detector of its own, so Auto gives it Silero once ONNX Runtime and a Silero copy are on the phone, and loudness until then. Silero is the better choice in a noisy room, where loudness can keep a phrase open.

Moondream’s Parakeet Ultra carries a small voice activity head of its own, which their Photon runtime uses to cut long recordings at pauses. The keyboard’s ExecuTorch export of Ultra was made with the four parts the Parakeet runner uses and left that head out, so it is not used yet; the phrase end detector does that job. Using it would take a new export of Ultra with the head as a fifth part, and a new download.

Several languages at once. A dictation that listens for several languages (see Language and engine) is Parakeet’s when every one of them is among its 25; it tells them apart on its own.

No words to listen for. Parakeet has no way to take a list of names or terms, so Words to listen for does not show while it is the engine.

Every build. The runtime, ExecuTorch built for the keyboard’s models alone, is in every build; F-Droid compiles it from source. A build made without it does not offer the engine. The NPU builds need Qualcomm’s libraries, so the F-Droid build does not offer them.

The licence. Both models are under CC-BY-4.0: v3 is NVIDIA’s, and Ultra is Moondream’s, built on it. Both are credited in the app’s About screen.