voice | pipesDownload
Voice pipelines for macOS

your voice, piped.

Hold a hotkey, talk, let go. Voice Pipes runs every take through a pipeline you built: transcribe on your Mac, fix your words, ask a model, paste at the cursor or read it back.

Download for MacSee a pipeline
Apple silicon · macOS 14 or later · signed, notarized, updates itselfOr install from a terminal, with the vp command: one line →

Illustrative demo: five takes

Shell (a track you build)⌃⌥T hold
REC 00:02PROC 584msOK 605ms

› “show me what changed in git”

~/voice-tools $ git status▌pasted at the prompt; Return is yours to press

parakeet 96 ms › llm · claude-haiku-4.5 488 ms › paste 21 ms · 605 ms

Clean dictation⇧⌥Space toggle
REC 00:06PROC 1,181msOK 1,203ms

› “uh hey sam so um the plan is we ship friday and I’ll send the notes tonight”

mai-transcribe-2 540 ms › fix-words 1 ms › claude-haiku-4.5 640 ms › paste 22 ms · 1,203 ms

Read aloud⌥R toggle
VOICE ···READ 12%READ 42%

› reads the selected paragraph

Pipelines you play from the keyboard. Hold a hotkey, talk, let go.Jev picked easy: plain prose, read as it is by a voice on your Mac

selection 2 ms › branch · easy 304 ms › speak · pocket tts 25 ms · 331 ms

Quick answer (recipe)your hotkey
REC 00:02PROC 960msREAD 30%

› “how many ounces in a cup”

quick → claude-haiku-4.5 · Jev 262 ms 100%A US cup is 8 fluid ounces.

parakeet 88 ms › jev route 262 ms › claude-haiku-4.5 610 ms › speak 25 ms · 985 ms

Note (a track you build)⌃⌥N hold
REC 00:03PROC 93msOK 233ms

› “remember to renew the domain before march”

Notes · new noteRenew the domain before March.

parakeet 92 ms › fix-words 1 ms › http request 140 ms · 233 ms

Sample text and illustrative timings, not a recording or a benchmark.

How it runs

one hotkey, one pipeline

Every track is a row of blocks. Your take goes in one end, text or speech comes out the other, and the HUD at the bottom of your screen says what's happening in one line. Nothing to click, nothing to switch to.

REC 00:04›PROC 118ms›OK 142ms
From the terminal

ride it from the terminal

vp, the command-line tool, runs your tracks, speaks, listens and transcribes, and edits your config, keys and vocabulary. It's built for agents too: compact output, next-step hints and no prompts, plus a skill that teaches Claude Code, Codex and others to use it, build a track from what you ask for, and read you what needs your attention.

install: app, vp and agent skill
curl -fsSL https://github.com/brancusi/voice-tools-releases/releases/latest/download/install.sh | bash

It checks the download is signed by Voice Pipes' developer and notarized by Apple, and asks nothing, so an agent can run it. macOS still asks you for Microphone and Accessibility on the first launch.

Install guideCLI referenceFor agents →
vp
# live state: the app, your config, your tracksvp
# speak, on your Macvp say "Tests passed."
# ask out loud, print the answervp ask "Staging or production?"
# send text through a trackvp run <track> --text "…"
# a file, on your Macvp transcribe memo.m4a
# validate config.tomlvp config check
# agents read you what needs youvp agents read-aloud attention
Tracks

three to start, build your own

A track is a hotkey plus a pipeline. Hold or toggle. Three come set up; change any block, or drag them into a new order. Quick answer is a recipe: build it from the same blocks and give it a hotkey.

Fast dictation
Mic›Parakeet›Fix words›Paste

Talk anywhere, text appears at the cursor. Runs entirely on your Mac.

⌥Space hold~140 ms
Clean dictation
Mic›MAI›Fix words›Haiku›Paste

Rambling in, clean prose out: filler and false starts gone, your wording kept.

⇧⌥Space togglecloud
Read aloud
Selection›Jev branch›Speak · Pocket TTS

Select text, press the key, listen; press again to pause. With Jev and OpenRouter keys, figures and code are rewritten for the ear first.

⌥R togglevoice on your Mac
Quick answerrecipe
Mic›Parakeet›Jev route›Speak

Ask out loud, hear the answer. Jev picks a fast, web or deep model for each question.

your hotkeycloud
Building blocks

snap the pipes together

Each block says what it takes and what it gives; + Add step lists what fits next first, and the editor won't let you connect two that don't fit. Every model is a choice in one picker: on your Mac first, then anything on OpenRouter, with what a paragraph costs, its speed on your Mac and a quality rating side by side.

Input
Microphone — → audioText selection, page, clipboard
Transcribe
Parakeet v3 on your MacAny OpenRouter transcription model
Transform
LLM your instructionsRoute · Jev picks the modelBranch · Jev picks the pathFix words your vocabularyHTTP request · Template
Output
Paste at cursorCopy to clipboardSpeak local or cloud voicesShow in HUD
Route · Jev

the right horse for the trail

A spelling question shouldn't wait on a big model, and a hard one shouldn't get a small one. Jev reads what you asked and picks a route in about a quarter of a second. You write the routes; the HUD shows which one won.

A Branch goes further: Jev answers a question about the text and picks a whole path of blocks. Read aloud starts with one, so code, links and figures are rewritten for the ear before they're read.

quickclaude-haiku-4.5"How many ounces in a cup?"
webperplexity/sonar"Who won the Champions League final this year?"
deepclaude-sonnet-5.5"Postgres or SQLite for a local-first app that syncs?"
READ 42%quick → claude-haiku-4.5 · Jev 262 ms 100%
~0.1 sfrom letting go to text, with Parakeet on your Mac (40–140 ms for 4–33 s of speech)
0.26 sfor Jev to pick a route; 15 of 15 test questions went where we'd send them
~25 msto the first sound when Pocket TTS reads aloud, streaming as it speaks
1,000takes kept in History, searchable, so a dictation that missed its window isn’t lost
On your Mac

home on the range

The fast path never leaves your machine. Cloud models are there when you pick them, with your own keys.

LOCALParakeet v3 transcription, Pocket TTS and Supertonic-3 voices, Fix words, the HUD. Downloaded once, then run offline.
KEYSOpenRouter and Jev keys live in your login Keychain, under your Developer ID-signed app. Nothing to sign up for here. Where to get each key →
DATATracks and vocabulary are commented TOML in ~/.config/voice-pipes; history is JSON in Application Support. Edit them, back them up, clear them.
CLOUDOnly the blocks you choose: an OpenRouter model, a Jev route or branch, an HTTP request you wrote. Read aloud's branch asks Jev only once you've added a Jev key.
Vocabulary

brand your words

Say a word five times; Voice Pipes replays each take 30 ways through Parakeet and rounds up every way it got it wrong, and with a Jev key, Jev flags which ones are safe to replace everywhere. Fix words then swaps them on your Mac, instantly. No model is retrained.

i asked cloud code Claude Code to set up open router OpenRouter
History

no steer left behind

Clicked away mid-sentence? The paste went nowhere? Your last 1,000 takes stay in History. Open one for every step: its time and output, which branch Jev picked, and what the cloud steps cost. Search them, copy one, paste it again.

Fast dictation · 2 min ago · 142 ms · paste had no fieldLet's ship the HUD changes in the next release.
Copied ✓
Read along

hold the reins

Anything read aloud, from your selection to an agent's summary, can show as text you follow, a sentence per line with the spoken word underlined. Click a sentence to jump there. The HUD takes the keys without taking you out of your app.

Model pickers

size up every model

Every picker shows what a paragraph costs, how fast the model has been on your Mac, and a 1–5 quality rating from Artificial Analysis's public leaderboards. Sort by best value, quality, speed or cost; the models on your Mac stay on top.

Changelog

new tricks every release

Voice Pipes checks for updates every few minutes and offers Install… when one is out. Your permissions and keys carry over.

1.8.2Every cloud voice now plays, whatever audio it sends back.
1.8.1New defaults: hands on the keyboard, ears on what matters. Change it by asking.
1.8.0Model pickers now show what a paragraph costs, how fast each model is on your Mac, and a quality rating from Artificial Analysis. Agents read to me unasked.
1.7.4Ask an agent to build or change a pipeline in plain words. Agents read aloud what you want them to.

saddle up.

Download, drag to Applications, hold ⌥Space and say something.
Download for MacApple silicon · macOS 14 or later