Dictation glossary

Transcription latency

Transcription latency is the time between speech and usable text — first partial, or the finalized insert after you stop. Real-time factor, or RTF, compares processing time with audio duration: divide how long the engine ran by how many seconds you spoke. An RTF below one means the system is faster than real time. Dictation users feel end-to-end latency; engineers quote RTF to describe the recognizer in isolation.

Get WhisperJot Try the browser demo
In more detail

What is Latency?

Several clocks hide inside 'how long did it take?' Time to first partial is when the HUD lights up. Time to final is when injection happens, after endpointing, decoding, and cleanup. Cloud engines add network round trips. On-device engines add model load on a cold start. Real-time factor ignores the network and asks only whether five seconds of audio cost more or less than five seconds of compute. An RTF of 0.2 means those five seconds were decoded in about one second of processing, all else equal.

Short dictation bursts care about a different budget than an hour-long file. A model with excellent RTF on a GPU can still feel slow if you pay a fixed startup cost every time you tap the hotkey. Batching 30-second chunks, as some Whisper-style pipelines do, makes long audio predictable and short audio less snappy. Hardware matters: a Neural Engine or a large CPU can change RTF more than a software toggle. Measure on the machine you own, with the engine you actually select.

For writers who speak

Why it matters for dictation

If the pause after you speak is long, you type. Latency is why dictation apps ship a fast engine for messages and a slower one for prose, and why cloud mode can feel quick on fiber and sticky on a weak hotspot. Understanding RTF keeps you from comparing a phone captioning demo with a local desktop model as if they shared a clock.

In this product

How WhisperJot handles it

WhisperJot ships Jot Local for fast short utterances and Jot Local Pro for long-form accuracy, both on-device by default, plus live partials in the HUD so you are not waiting blind. Jot Cloud is the opt-in engine when you want to skip local model load and Neural Engine use; it needs a network and is subject to fair use. Meetings share the loaded local engine and cannot run at the same time as dictation.

Related terms

Keep reading in the glossary.

Live partial transcription

Live partial transcription means showing in-progress text while audio is still being captured, instead of waiting until you stop talking to reveal a finished transcript.

Read Live partials →

Cloud transcription

Cloud transcription means speech-to-text that runs on a remote server: your app uploads audio, a model in a data center produces text, and the words come back over the network.

Read Cloud transcription →

Apple Neural Engine

Apple Neural Engine is a dedicated block of silicon on Apple chips that runs neural networks locally, alongside the CPU and GPU.

Read Neural Engine →

Parakeet

Parakeet is a family of open speech recognition models from NVIDIA designed for fast automatic speech recognition.

Read Parakeet →
Questions

Straight answers.

See also: System requirements, Pricing.

What is transcription latency?

Transcription latency is the wait between speaking and seeing text. You can measure time to the first partial word or time to the cleaned insert after you stop. Cloud recognition includes network delay; on-device recognition includes however long the local model needs. Dictation is sensitive to this pause because it sits in a typing loop. File transcription can tolerate more delay in exchange for accuracy.

What is real-time factor in speech-to-text?

Real-time factor is processing time divided by audio duration. If a 10-second clip takes 2 seconds to transcribe, RTF is 0.2, which is faster than real time. If it takes 20 seconds, RTF is 2.0, and the job lags the recording. RTF describes the recognizer, not the full product: paste, cleanup, and network can add time that never appears in an RTF number from a benchmark.

Why is on-device dictation sometimes slower on the first utterance?

Local models often compile or load into memory on first use, then stay warm. That cold start is extra latency, not a permanent RTF. Subsequent utterances reuse the loaded engine. Cloud engines skip the local load and pay a network cost instead. If every utterance is slow, you are looking at hardware, a huge model, or a stuck download, not only a one-time warmup.

One hotkey, any focused app.

Private local transcription by default, with an optional opt-in cloud engine.

Get WhisperJot — $99/year