People are often surprised that a meeting recorder can produce a full transcript with the Wi-Fi switched off. The assumption is understandable: for most of the last decade, speech recognition was something you rented from a server. You uploaded audio, a data centre thought about it, and text came back.
That is no longer the only option. The speech models that matter now are small enough to sit in a file on your laptop and fast enough to be useful there. Here is what is actually happening when JesRecap turns a recording into text without touching the network.
What a local speech model is
A transcription model is a single file — a few hundred megabytes of learned numbers. It is not a program that phones home and it does not contain your data; it contains statistical patterns about how speech sounds and how language fits together, learned once during training by whoever built it.
Whisper-family models, the type JesRecap uses, work in two stages. An encoder converts short slices of audio into a compact numeric representation of what was heard. A decoder then writes out text one piece at a time, using both that representation and the words it has already produced. That second half is why these models handle real speech so well: because the decoder knows the sentence so far, it can pick "their quarterly review" over three phonetically similar alternatives.
Both stages are just arithmetic over the model file. There is no step in which the audio has to travel anywhere.
Why a laptop can do this now
Three things changed at roughly the same time.
Hardware. Modern laptop chips ship with a lot of memory bandwidth and vector units designed for exactly this kind of dense arithmetic. Apple Silicon in particular gives the whole machine access to one shared pool of fast memory, which suits a model that needs to stream a few hundred megabytes of weights repeatedly.
Quantisation. Models are trained at high numeric precision, but they do not need to run that way. Storing each weight in fewer bits — a process called quantisation — can shrink a model dramatically while affecting output far less than you would expect. It also makes inference faster, because the main bottleneck is usually moving weights through memory rather than the maths itself.
Efficient runtimes. A lot of engineering has gone into running these models well on consumer CPUs and GPUs rather than only on server accelerators. JesRecap inherits its inference layer from JesType, our offline dictation app, which has been doing on-device speech recognition for the same reasons.
One more practical detail: transcribing a recording is not the same problem as transcribing live. If you are happy to wait until the meeting ends, the model can work through the audio at whatever pace your machine allows, using the whole CPU because nothing else is competing. That is why JesRecap treats live transcript as optional — it is nice to watch, but deferring the work to the end keeps memory use lower and lets a larger model stay in play.
Model size: what you are actually trading
Bigger models are generally more accurate and always slower. That is the whole trade, but it is worth understanding where the extra accuracy shows up, because it is not evenly spread.
On clean audio — one speaker, decent microphone, plain vocabulary — small models do remarkably well, and a larger one mostly gives you back a few punctuation and capitalisation decisions. The gap widens sharply on hard audio: accents the model has heard less of, crosstalk, a bad connection, proper nouns, technical jargon, industry acronyms, numbers spoken quickly. That is exactly the material real meetings are made of, which is why the default worth trying is a mid-sized model rather than the smallest one.
JesRecap ships a short list rather than a wall of options. Roughly:
- Around 150 MB — a balanced English model. Fast, light, a sensible starting point on any machine.
- Around 500 MB — English-tuned or quantised multilingual options. Noticeably better on difficult audio and non-English speech; slower, and worth it when the recording matters.
- Smaller, speed-focused options — useful when you want a live transcript on a modest machine, or you mostly record clean one-on-one English.
We deliberately do not publish accuracy percentages or words-per-second figures. Any number we quoted would be a number measured on our audio and our hardware, and it would tell you very little about your recordings on your laptop. Your own first meeting is a better benchmark than our marketing.
Because both source tracks are kept on disk, changing your mind is cheap. If a transcript reads poorly, switch model and retranscribe the same meeting. No re-recording, no re-upload, no per-minute charge for the second attempt.
"Downloaded once, runs forever"
The first time you set JesRecap up, it fetches the model you chose from
models.jesrecap.com and saves it on your machine. That download
is a plain file transfer of a public model file — it carries no audio, no
transcript, and nothing about your meetings.
After that, the app has everything it needs. Transcription runs against the local file, so:
- It works on a plane, on hotel Wi-Fi, or with networking disabled entirely.
- There is no per-minute cost and no monthly quota, because there is no meter to read.
- The model does not silently change under you. The version you downloaded is the version you keep, so a transcript you produce next year is generated the same way as one you produce today.
- If our servers vanished, your setup would keep working. Nothing about transcription depends on us being reachable.
This is the part of local-first that is easy to undersell, because it looks like an absence of features. What it really is, is an absence of dependencies. The only network traffic JesRecap generates is that one-time model download and license activation — no accounts, no telemetry, no audio, ever.
The honest limits
Local transcription is not magic. It costs you time after the meeting instead of somebody else's compute; a long recording is minutes of work, not instants. Very noisy audio with heavy crosstalk is hard for any system, cloud or local, and no model invents information it never heard. And you are responsible for your own backups — files on your disk are private precisely because nobody else has a copy.
Those seem like reasonable terms. In exchange, a recording of your most sensitive conversations becomes text without ever being handed to anyone.