EchoVoice

EchoVoice™ — hear your AI partner

Milo speaks every reply, from every terminal agent, in his own window. Claude Code, Codex CLI, Gemini or Kimi finishes an answer in the terminal; EchoVoice reads it aloud, each partner in its own voice, so your eyes stay on the code while your partners talk. Free local voices, no account, no telemetry. Your words go nowhere.

Built for Claude Code, Codex CLI, Gemini / Antigravity, and Kimi Code — each in its own voice, so you can run several partners at once and always know who's talking. Part of the EchoTools suite by PiGON AI. No telemetry — nothing about you or your code is collected. Network use is limited to things you choose: a local engine's one-time, integrity-verified download, and ElevenLabs synthesis if you bring your own key.

▶ Watch the full tour — with Milo's voice — or hear it inside the extension: EchoVoice: Play Intro.

Better together: install EchoAvatar and every voice plays through Milo — a living avatar with mouth-sync, a dance, and an ear tuned to your voice. Add EchoMemory and Milo remembers. The three are one companion and install as a ladder: EchoVoice works on its own; EchoAvatar brings EchoVoice with it; EchoMemory brings both. VS Code keeps the ladder whole, so it won't uninstall EchoVoice while EchoAvatar or EchoMemory needs it, and it names which one.

Features

• 🔊 Speaks new responses automatically — watches your partners' conversation transcripts and reads each new reply aloud.
• 🎙️ Four voices to choose from. Best sound: ElevenLabs, with your own key. Best free and private, and the default we recommend: Kokoro. On a Mac the choices are Kokoro, ElevenLabs and the system voices; Piper is Windows and Linux.
  • Kokoro (recommended) — free, local, and the most natural of the free options. Twenty-nine voices (US / UK, male and female), plus three Halloween voices with a sound at the end (below). They share one model that downloads once, 92 MB, and EchoVoice asks before spending a byte; after that it's offline, and each extra voice is half a megabyte. Runs the same on Windows, macOS (Intel and Apple Silicon), and Linux. It works your CPU: on a modern machine it flows; on an older or smaller one, long replies can pause between sentences while the next one is made — see Kokoro and your machine in the FAQ before you decide it's broken. It isn't; it's the machine.
  • Piper (Windows and Linux) — free, local, neural, and much lighter on the machine: about 150 MB of memory while it speaks against Kokoro's 600 MB to 1 GB, and a sentence is made in a fraction of the time. Ten voices (US / UK / Scottish); the engine is 39 MB and each voice you pick is 63 to 121 MB, so on disk it is not the smaller choice — see What it costs below. The right pick for a slower machine. Not offered on macOS: the published Mac build never ran there (it is missing the libraries it needs), so a Mac speaks with Kokoro, and a Mac's own system voices are the light option.
  • System voices — free, offline, works out of the box (robotic on some systems). Also the automatic fallback if a local engine can't run, so you still hear your partner.
  • ElevenLabs — bring your own API key and voice ID for premium/cloned voices. The key is stored in VS Code's encrypted SecretStorage, never in settings files.
• 🎭 A voice per partner — give Claude Code, Codex, Gemini, and Kimi Code each their own voice, so you know who's speaking without looking.
• 🤖 Milo's own voice — the lines Milo speaks himself (a summary, a greeting, "Codex finished…") have their own row. By default he speaks in your computer's own voice, so he works everywhere with no download; give him any Kokoro, Piper, system or ElevenLabs voice, and he keeps it whatever engine your partners use. Kokoro on the partners and ElevenLabs on Milo works fine. He has his own mute row too.
• 🎨 One glance, one speaker — while someone talks, the status bar shows their mark and colour: Milo's robot, Claude Code's sparkle, Codex's terminal, Gemini's telescope, Kimi's rocket, in the same tints Milo's body glows in EchoAvatar. The name stays beside the mark, so colour is never the only cue.
• 🎤 Voice input for EchoAvatar's push-to-talk — hold the button in Milo's window, speak, and Milo answers. EchoVoice provides the speech-to-text pack (about 150 MB, once, after one card that names the size); EchoAvatar listens; what you say is transcribed on your machine. Prepare voice input in the menu downloads it ahead of time.
• 📖 Teach Milo a word — a name the voices say wrong, fixed in two questions from the menu; Milo says it back so you hear it before you keep it.
• ✂️ Sentence cap — off by default, so a spoken reply is read to the end; set N to hear only the first N sentences of long answers. (Which replies are spoken at all is the speak mode: important by default, which skips progress chatter.)
• 🔇 One-click mute — status bar toggle, always visible. It mutes this window until you reload it; to keep EchoVoice quiet for good, turn off echovoice.enabled in Settings.
• 📋 Copy last response — the full text, straight to your clipboard.
• 📄 Export transcript — turn any saved Claude Code conversation into a clean Markdown file.

Halloween voices

Three voices for the season, under Halloween at the top of Per-partner voices: the Witch (thin and sharp, a flutter at the edges), the Pumpkin (low and hollow, with a growl) and the Ghost (wavering in a cold hall). Give one to a partner, or to Milo himself. Each is a blend of two Kokoro voices with its effect applied as it speaks, and each ends every reply with its sound: the Witch cackles, the Pumpkin laughs, the Ghost leaves a haunted hush. echovoice.halloween.endSound turns the sound off and keeps the voice. They were tuned by ear in a listening lab, not by numbers.

The free voices, and the work behind them

Kokoro and Piper are open models anyone can download. What you get here is not the download; it is the listening in between. We did not build a new voice — we took an open one and taught it to read developer text, our ROAR-U way: repurpose, optimize, adjust, recycle, upgrade before building anything new. Every line below carries a number because we measured it on real replies, not because it sounded good in a meeting.

• A dictionary built for Kokoro. The pronouncing dictionary that ships with the free engine knew 126,000 words and none of async, TypeScript, linter or middleware. Kokoro's is now built from the very transcriptions the model was trained on — misaki's, used under their Apache-2.0 licence, with thanks — over CMUdict: 254,000 words, every one checked against the model's own alphabet, plus a hand-written list of the names developers actually say: Supabase, Vercel, Kubernetes, nginx, Postgres, Vite, Prisma.
• Words it does not know, read as words it does. "EchoVoice" used to come out as "etch-uh-voice" and changelog as "chain-jel-og". A word missing from the dictionary is now read as two that are in it — echo and voice, hand and off, screen and shot, un and install — with the stress where English puts it.
• Reading lessons from 400 real replies. We surveyed what tripped the voices and taught them one lesson at a time: acronyms spelled or said (CPU, README), numbers with their units, commit hashes cut to seven characters, UUIDs as "an ID", filenames with versions, domains with their dots, plural acronyms (LLMs), Date.now(), and curly apostrophes, which had been turning "I'll" into "ill". A fourth round followed: words inside paths and members (src/ui is "source slash U I", JSON.parse is "jason dot parse"), a number with its unit read the way English says it ("a 30 second timeout"), initialisms like IIRC and AFAIK spelled, and 24 developer words (iterable, gRPC, zlib, mojibake…) each heard and kept by ear before it went in. On those replies, the words the engine had to guess at fell from one in 37 to one in 330.
• Pauses, measured. On a machine that synthesizes slower than it speaks, long replies paused between sentences. We measured the pipeline instead of guessing: Kokoro now uses more of a big machine's cores, starts from a short first sentence and keeps two ahead. On our test machine, dead air across a three-minute reply went from 9.8 to 2.6 seconds and the first word arrives 1.3 seconds sooner. Every reply still writes its own receipt to the log.
• Piper reads the same lessons through its own engine, and stays the light choice for a slower Windows or Linux machine.

None of this is a switch you flip. It is in the box, it is free, and every number above was measured before it was written down; ask us for the receipt behind any of them.

EchoMemory: Milo remembers

EchoMemory is Milo's memory. Save notes for yourself, project by project, and Milo remembers them: ask him what you noted, by voice or by typing, and he answers with your own words, where they came from, and when, in his own voice through EchoVoice. Ask him to sum up what you just told him, or one of your notes, and his version waits in your Inbox for your OK. Your notes are plain files in your project; the folder keeps itself out of git, and no AI coding tool is connected to them. When you talk to Milo, what you say and the notes he reads go to the brain you chose: your Copilot, Ollama on your own computer, or Ollama's cloud with your key. EchoMemory installs with EchoAvatar and EchoVoice, because the three make one companion.

Quick start

1. Install EchoVoice.
2. Open a Claude Code session in VS Code and ask it anything.
3. When the response lands, EchoVoice speaks. That's it.

Click the 🔊 EchoVoice item in the status bar to mute or unmute this window (a reload brings back your Settings choice, echovoice.enabled). Unmuting also clears an "off" saved in your Settings by an older version, so EchoVoice doesn't start muted again.

Take the guided tour — a five-step walkthrough with a narrated intro by Milo (just under three minutes, full transcript included). It opens by itself on a fresh install; to revisit it any time, open the Command Palette and run "Welcome: Open Walkthrough…" → Get Started with EchoVoice — or jump straight to the narration with "EchoVoice: Play Intro".

Using your own ElevenLabs voice

1. Run EchoVoice: Set ElevenLabs API Key from the Command Palette and paste your key (input is masked; stored encrypted). If an ELEVENLABS_API_KEY environment variable is present, it is used as a fallback.
2. Set echovoice.elevenlabs.voiceId in Settings to the voice you want.
3. Set echovoice.provider to elevenlabs.

If ElevenLabs is unreachable or misconfigured, EchoVoice falls back to your system voice instead of going silent.

Commands

Settings

How it works (and what it doesn't do)

Coding partners save conversations as transcripts on disk. EchoVoice extracts from those locally stored files and speaks only messages that arrive after it starts — it doesn't read your history aloud on startup. Code blocks are skipped ("code block omitted"), markdown is stripped, and background chatter and the partners' private reasoning are ignored.

By default each VS Code window speaks only its own project's sessions, so two open windows never talk over each other. Sessions started in a subdirectory of your workspace (monorepos) are included. Set echovoice.scope to all if you want one window to voice everything.

EchoVoice runs locally by design (extensionKind: ui) — in SSH, WSL, or Dev Container sessions it stays on your machine, where your speakers and your Claude Code transcripts actually live.

Platform notes:

• Windows / macOS — works out of the box (SAPI / say). On macOS the voices offered are Kokoro, ElevenLabs and System; Piper is not offered there.
• Linux — system voice needs spd-say (install your distro's speech-dispatcher); ElevenLabs playback needs mpv. If a binary is missing, EchoVoice tells you which one instead of failing silently. The voice input engine needs glibc 2.35 or newer (Ubuntu 22.04, Debian 12, Fedora 36 and later).
• Voice input is built for Windows x64, Linux x64 and Apple silicon Macs. On a Mac the first download takes a little longer to say it's ready: EchoVoice runs the new engine once, so your first words aren't kept waiting. On other machines the menu says so in one sentence instead of downloading anything.

EchoVoice contains no telemetry and no analytics code — that's checkable in this package, and DISCLOSURES.md itemizes every byte that can ever leave your machine. The complete network surface is: (1) the one-time Kokoro or Piper model download from pinned, SHA-256-verified sources — only if you choose that voice, and only after you agree to it; (2) the one-time voice input pack (a speech-to-text engine and its model), from the same pinned, verified release — only when you hold EchoAvatar's talk button or run Prepare voice input, and only after you agree; (3) the synthesis request to ElevenLabs — only if you choose ElevenLabs with your own key — carrying the prepared text (code blocks are stripped before it is sent), your key, and normal API metadata. System voices make no network calls at all. Voice input itself never leaves your machine: EchoAvatar runs the engine locally and EchoMemory receives text only.

What it costs

Measured, not guessed — on a 24-core laptop running Windows 11, on 2026-09-23, with the same runner the extension uses. A smaller machine is slower, not different; every number below has a receipt, and we will show it if you ask. EchoAvatar's README has the same section for Milo's window.

Disk

Plus a few kilobytes of bookkeeping: which replies have already been spoken (forgotten after a day) and how far each conversation was read.

Memory

The extension itself is a file watcher and a small pipeline; your editor will not notice it.

CPU and time

• Kokoro made a 25-second paragraph in 19 seconds — 0.75× real time — on 8 threads, and is ready about a second after it wakes (model plus dictionary). It uses up to 8 threads and leaves your editor a core on any machine with more than two. On a machine that synthesizes slower than it speaks, long replies pause between sentences; the FAQ entry Kokoro and your machine explains the receipt it leaves in the log.
• Piper (Windows and Linux) made a 28-second paragraph in 2.6 seconds including starting its process — 0.10× real time. That is why it is the pick for a slower machine there; on a Mac the light pick is the system voice.
• On a small machine (we tested a mini PC) the same work is slower: Kokoro ran