Transcription Engines

Transcription Engines

Transcription engines are the speech-to-text plugins that turn your audio into text. TypeWhisper for Linux bundles both local engines that run on your machine and cloud engines that send audio to an external provider.

Local engines

Local engines run on your own hardware and do not require an API key.

Engine	Notes
WhisperCpp	Bundled whisper.cpp engine. Catalog includes the Large V3 models — large-v3 (f16, ~3.1 GB) and large-v3-q5_0 (~1.1 GB) — alongside the large-v3-turbo variants; the full models run the 32-layer decoder (more accurate, heavier) versus turbo's distilled 4-layer decoder, a good fit when you have GPU headroom. Has a configurable noSpeechThreshold for filtering silent segments to reduce hallucinated phrases. CUDA-capable.
SherpaOnnx	Local engine. Offers the Parakeet TDT 0.6B and Canary 180M Flash models, and is CUDA-capable.
Qwen3Stt	Local engine.
Voxtral	Local engine. Transcription only — it does not support translation.

CUDA acceleration applies to the bundled whisper.cpp and sherpa-onnx engines; other bundled local engines stay on CPU. See GPU & CUDA for setup and requirements.

Cloud engines

Cloud engines send audio to an external provider and require an API key.

Engine	Real-time websocket streaming
OpenAi	Yes (GPT Realtime)
OpenAiCompatible	No
Groq	No
Deepgram	Yes
AssemblyAi	Yes
ElevenLabs	Yes
Speechmatics	Yes
Soniox	Yes
Reson8	Yes
SmallestAi	Yes
Xai	Yes
Gladia	Yes
CloudflareAsr	No
GoogleCloudStt	No

Real-time streaming

The engines marked above with real-time websocket streaming stream audio to the provider over a websocket so partial text can appear in the overlay as you speak, instead of re-uploading the growing buffer on a poll. The full streaming set is:

AssemblyAi, Deepgram, ElevenLabs, Gladia, OpenAi (GPT Realtime), Reson8, SmallestAi, Soniox, Speechmatics, and Xai.

Streaming is enabled per session from the dictation settings and falls back to the batch polling preview when the engine or setting does not support it. See Dictation for how this surfaces in the recording overlay.

If a streaming session can't be finalized in time, the utterance fails over to batch transcription rather than being lost.

Per-engine limits

Engine	Limit
Gladia	Real-time streaming only. Batch transcription (file transcription, recorder) is unsupported and fails immediately with a clear error rather than failing obscurely later.
GoogleCloudStt	Audio longer than 60 seconds is rejected with a clear error — that is the synchronous API's limit. The request timeout is 120 s, which covers uploading the audio as well as recognizing it.
Voxtral	Transcription only; no translation.

Related pages

• Dictation — choosing and using a transcription engine while dictating.
• GPU & CUDA — CUDA acceleration for the bundled whisper.cpp and sherpa-onnx engines.
• Plugins — the plugin model these engines plug into.

---

Changelog

Date	Change
2026-06-17	Initial version.
2026-06-17	Added Transcription Engines screenshot.
2026-06-23	v0.10.0: noted the whisper.cpp Large V3 models and the sherpa-onnx Parakeet/Canary models; extended CUDA acceleration to sherpa-onnx.
2026-07-27	v0.13.0: added per-engine limits (Gladia streaming-only, Google Cloud STT 60 s cap, Voxtral no translation) and streaming finalize failover to batch.