Latency
Also: key-up latency
Updated
Definition
In dictation, latency is the time between the end of your speech and finished text appearing at the cursor. It is the number that decides whether a tool feels instant or feels like waiting.
Measure it from the moment you release the key, because that is the moment you are waiting from. Everything that happens after key-up counts: finishing recognition of the last few hundred milliseconds of audio, the formatting pass, and the insertion into the app. A tool that streams recognition has only the tail to finish; a tool that batches has the whole utterance still to process, so its latency grows with every second you spoke.
Under about a second, most people stop noticing. Around two or three seconds, they start glancing at the screen and then at the tool. Beyond that, dictation stops being a way to write and becomes a way to make a recording you then wait for.
Where the time goes
Recognition tail: how much audio remains to be processed at key-up, which streaming reduces to almost nothing. Formatting: a language-model pass over the transcript, typically the largest single cost in a local tool, and the reason a fast tool keeps that model small and gives it a strict time budget. Insertion: typing the result into the focused app through the accessibility layer, usually tens of milliseconds. Network, for cloud tools: a round trip plus queueing on a server you do not control, which is why cloud latency varies from fast to unusable depending on the hour.
Latency versus quality is a real trade
A formatting pass costs time, and skipping it is the fastest possible tool. It is also a tool that hands you a transcript to fix by hand, which costs more time than it saved. The right design offers the choice: a verbatim or lightly cleaned mode that lands almost immediately, and a fully formatted mode that takes a little longer and needs no fixing. What a tool should never do is let a slow model hold the text hostage; if the pass runs late, the plain transcript should land anyway.
How to measure dictation latency on your Mac
You do not need special tools. A screen recording that also captures your microphone puts your voice and the screen on the same timeline, so you can read the delay straight from the playback.
Measure several dictations of different lengths. One reading tells you little, and the difference between a short and a long dictation tells you whether the wait grows with how long you speak.
- Open the app you normally dictate into and place the cursor in a text field.
- Press Shift-⌘-5 to open Screenshot, and choose to record the entire screen or a portion of it.
- Click Options and choose your microphone, so the recording includes your voice.
- Start the recording, then dictate one short sentence and one long one, leaving a few seconds between them.
- Click the stop button in the menu bar. The recording is saved as a .mov file whose name starts with Screen Recording.
- Play the file and note two moments for each dictation: when your voice stops, and when the finished text is on screen. The gap between them is the latency.
- Compare the short and the long dictation. If the long one waits noticeably longer, the tool is processing in batch.
Why a bigger model means a longer wait
Model size is the most direct trade between accuracy and speed. The Whisper project publishes relative speeds for its own models: against the large model as 1x, tiny runs at about 10x, base at about 7x, small at about 4x, medium at about 2x, and turbo at about 8x. Those figures are relative to each other, and real speed depends on your hardware.
Speed alone does not make a model live. Whisper reads audio through a sliding 30-second window, and its model card states that the models cannot be used for real-time transcription out of the box. A dictation tool built on such a model has to add its own streaming or chunking, and how well it does that decides the delay you feel.
How to time a speech model in Terminal
If you want to see the raw cost of recognition without any app around it, you can time the open-source Whisper command-line tool on a recording of your own voice. This measures the model on your Mac and nothing else: no formatting, no insertion, no network.
The number is not the latency of any dictation app, because apps process audio while you speak. It is useful for comparing model sizes on your own hardware.
- Install the audio dependency with: brew install ffmpeg
- Install Whisper with: pip install -U openai-whisper
- Record yourself speaking for about ten seconds and save the file, for example as clip.wav.
- Run: time whisper clip.wav --model tiny and note the real time that Terminal prints.
- Run: time whisper clip.wav --model turbo and note the time again.
- Run each command a second time and compare those timings, so that one-time setup is not counted.
Questions people ask.
What is a good latency for dictation?
Under a second from key-up to finished text feels instant. One to two seconds is acceptable. Beyond that, the wait becomes the experience. Streaming recognition and a small, time-boxed formatting model are how local tools get under a second.
Why does my dictation tool get slower the longer I talk?
Because it recognizes in batch: nothing is processed until you stop, so a longer recording means a longer wait. Streaming tools process as you speak and keep the delay constant.
Is cloud dictation slower than local?
Often, and less predictably. A round trip to a server adds time on every dictation and varies with your connection and their load. Local tools have a fixed cost that depends only on your Mac.
How do I test how fast my dictation app is?
Record your screen with the microphone on using Shift-⌘-5, dictate a sentence, and play the recording back. The time between your voice stopping and the finished text appearing is the latency.
Can Whisper transcribe in real time?
Not by itself. The Whisper model card says the models cannot be used for real-time transcription out of the box, and the standard code processes audio in 30-second windows. Apps that feel live with Whisper have added their own streaming layer.
Does a smaller speech model make dictation faster?
Yes, at a cost in accuracy. The Whisper project lists its tiny model as about 10 times faster than its large model, and its turbo model as about 8 times faster.
