Which local models stream
Streaming follows the model. There is no switch to turn it on: a model built for streaming always streams, and every other model transcribes after you stop. The catalog marks this per model, and the Streaming filter on the Models page lists them. The streaming models are Parakeet Unified EN 0.6B (the default), Nemotron Streaming 3.5, Nemotron Speech Streaming EN, Multitalker Parakeet Streaming EN, Voxtral Mini 4B Realtime, and Moonshine Streaming Tiny, Small, and Medium. Whisper, Canary, Cohere, SenseVoice, GigaAM, the other Parakeet and Moonshine models, and the rest do not stream. See Transcription models for the full catalog. Streaming models keep up to 1.65 seconds of every pause in the recording, where other models keep 0.45 seconds. The longer pause is how a streaming model tells where one sentence ends and the next begins. Stop after silence still counts from the same moment.Long recordings on other models
A model that does not stream can still work during the recording. With Split long recordings on (the default, in Settings → Advanced → Performance), a dictation longer than 40 seconds is cut at pauses into pieces of about 30 seconds, and each piece is transcribed while you keep talking. Most models lose accuracy or fail on long audio in one go, so this keeps them accurate. Whisper handles long audio itself and is never cut. If any piece fails, the whole recording is transcribed again after you stop.Cloud streaming
Three cloud providers stream: ElevenLabs (withscribe_v2_realtime), Deepgram, and Amazon Transcribe. Their text arrives while you are still talking, so releasing the keys only flushes the last few words.
Turn this on with the Transcribe as I speak switch in the cloud transcription settings. For ElevenLabs you need to select a realtime model; the batch model (scribe_v2) only transcribes after you stop. Amazon Transcribe always streams, because it has no batch endpoint.
If the live stream fails or produces nothing, SpeakoFlow falls back to transcribing the full recording in one batch request. You never lose your words. The batch and realtime models can produce slightly different wording for the same audio, and the substitution is logged.
See Cloud transcription for setup.
How it connects to the overlay
The overlay style in Settings → Dictation → Dictation overlay decides what you see, never how the audio is transcribed.- Live. A readable card with the running text.
- Minimal. The compact pill. A streaming model still streams behind it.
- None. No overlay. A streaming model still streams.