The writing around the code, dictated
You can hold a whole system in your head and still stall on the commit message. Waver takes the part of the job that is prose — commit messages, PR descriptions, docstrings, design docs, incident write-ups, the Slack reply you have been avoiding since standup — and lets you say it instead of typing it. Then it hands back something you can paste.
The part engineers notice first: Waver seeds the transcriber with proper nouns pulled from your own notes, so your service names, your libraries and your teammates come back spelled the way you spell them.
Your stack, spelled right
Generic dictation hears your architecture and writes it phonetically. Kubernetescomes back as "Cuban Aities". Your internal service becomes three words that are not words. You spend longer fixing the transcript than you would have spent typing it.
Waver fixes that before the audio is ever sent. Speech models accept a short block of context that they condition on as if those words had just been spoken — so Waver fills it with actual names, not instructions. Three sources, in priority order:
- The name you gave at onboarding.
- Terms you added yourself in Settings — repo names, internal services, and lowercase tools like kubectl or pnpm that no harvester would ever spot.
- Proper nouns harvested from your own recent notes, ranked by how often you actually say them.
That context window is small — a couple of hundred characters — so Waver spends it on the highest-signal names and drops the rest. Words that only ever appear capitalised at the start of a sentence are filtered out, because a capital after a full stop tells you nothing. The practical effect: the longer you use Waver, the better it gets at your particular codebase, without you configuring anything.
What engineers actually dictate
| You say | Waver hands back |
|---|---|
| Why this function is weird | A tight docstring or comment block explaining the why, not the what. |
| What you just changed | A commit message: subject line, then a body that survives a git blame in two years. |
| The context behind a branch | A PR description with what changed, why, and how to test it. |
| A ten-minute architecture ramble | A structured design doc or RFC with headings, tradeoffs and open questions. |
| A half-formed answer to a thread | A short, readable Slack reply instead of the message you kept not writing. |
| Narration while you debug | A running incident log you can turn into a postmortem timeline. |
| A standup or planning call | Speaker-separated notes with decisions and action items pulled out. |
| A whiteboard after the meeting | The diagram and scrawl scanned into text you can search later. |
| Notes on an unfamiliar codebase | Flashcards and quizzes on a spaced-repetition schedule, from your own notes. |
The writing is not the hard part. Starting it is.
You already know what the PR does. You know exactly why the retry logic has a comment on it. The cost is not thinking — it is the context switch out of the code and into prose, holding a paragraph in your head while your fingers catch up, and rewriting the first sentence four times.
Speaking skips all of that. You explain the change the way you would explain it to the reviewer standing behind you — messily, with an aside in the middle — and the AI rewrite turns it into something structured. You edit down instead of building up, which is a much cheaper mental operation.
Commit messages and PR descriptions
Two minutes of talking covers what you changed, what you deliberately did not change, the one weird constraint that forced the approach, and what a reviewer should look at first. Ask for a commit message and you get a subject line and a body. Ask for a PR description and you get the same material shaped for a reviewer, with a section on how to test it. Same recording, different output — you are not dictating twice.
Design docs that get written at all
The RFC that never ships is usually the one that was never started. Talk the whole thing through on a walk — the problem, the two options you rejected, the migration you are worried about — and come back to a draft with headings instead of a blank file. The parts you left vague stay visibly vague, which is honest, and easier to fix than a blank page.
Calls, standups and incidents
Meeting recording separates speakers, so a design review comes back as a transcript where you can tell who committed to what. During an incident, narrate as you go — every hypothesis, every thing you ruled out, every timestamp. Afterwards you have the timeline instead of trying to reconstruct it from your shell history at 2am.
Ramping onto a new codebase
First fortnight on a new team, you are learning fifty proper nouns you have never heard before. Record yourself explaining each subsystem back to yourself after you read it — the classic test of whether you actually understood it. Waver turns those notes into flashcards and quizzes on a spaced-repetition schedule, and every name you say gets added to your vocabulary, so the next recording spells it correctly.
Where it sits in your setup
There is no IDE plugin. Waver runs as its own app on macOS, Windows, iPhone and iPad — beside your editor, not inside it. Recording is offline-first on mobile, so a basement office or a flight does not lose the thought; transcription and AI run when you reconnect. Notes sync across every device, so the idea you dictated on the walk home is on your desktop when you sit down.
On your recordings
You are going to say internal service names, customer names and unshipped architecture out loud. Recordings are encrypted, never sold, and never used to train models. Delete any note — or your entire account — whenever you want. No ads on any plan, including free. Check the privacy policy before you record anything your security team would care about.
How it works
- 01Add a few termsDrop your repo names, internal services and lowercase CLI tools into your personal vocabulary. Takes a minute, once.
- 02TalkHit record and explain the change, the design, or the incident. Length does not matter. Neither does coherence.
- 03Pick the shapeCommit message, PR description, design doc, summary, email, Slack reply. One tap each, from the same recording.
- 04PasteCopy it into your editor, your terminal, or the PR box. Waver keeps the note, and the names in it feed the next transcription.
Stop retyping what you already know
Unlimited transcription and 10 AI actions a month, free, with no card. Premium unlocks unlimited AI from $4.13/mo.
Try Waver free →FAQs
How does Waver get library and service names right?
Two ways. You can type terms into your personal vocabulary in Settings — repo names, internal service names, lowercase CLI tools like kubectl or pnpm. And Waver harvests capitalised words that appear mid-sentence in your own recent notes, ranked by how often you actually say them. Both feed the transcriber as decoder context before it hears a word, so it is already leaning toward Kubernetes rather than guessing phonetically.
Why not just tell the model to 'spell proper nouns correctly'?
Because that is not how the transcriber's prompt field works. It is a context prefix, not an instruction — the model conditions on it as if those words had just been spoken. An instruction sentence biases it toward the words in the instruction. The actual names are what move the needle, which is why Waver sends names instead of orders.
Is there a VS Code or JetBrains plugin?
No. Waver runs as its own app on macOS and Windows, so it sits beside your editor rather than inside it. You dictate, the cleaned text lands in Waver, you paste it where it belongs.
Can it handle a rambling architecture monologue?
That is the intended use. Talk for ten minutes, contradict yourself, backtrack, restart a sentence. The AI rewrite turns it into a structured document with headings — you edit down rather than starting from an empty file.
What about standups and incident calls?
Meeting recording separates speakers, so you get who said what, plus a summary with decisions and action items. Useful when you are the one who has to write the postmortem afterwards.
Does it work if I do not speak English at work?
Yes — 120+ languages, transcribed in their own script, with live translation if your PR description needs to land in English but your thinking does not happen there.
What happens to my recordings?
They are encrypted, never sold, and never used to train models. You can delete any note, or your whole account, whenever you want. Waver is ad-free on every plan.
What does it cost?
Free tier: unlimited transcription and 10 AI actions a month, no card. Premium unlocks unlimited AI from $4.13/mo.
Say it once. Paste it in.
120+ languages · macOS, Windows, iPhone, iPad · your vocabulary, not a generic one.
Get Waver free →