A zot extension that speaks the last agent message aloud, using Gradium Phonon text-to-speech on your own machine. No network access, no account, no cloud voice.

Features · Requirements · Usage · Install · For developers · Acknowledgements · License · Design notes

Phonon is a 31 M-parameter model: natural speech with imperfect prosody, not a commercial studio voice. Output is mono 24 kHz through paplay. Speaking needs a live interactive session — a one-shot zot -p "…" tears the extension down before the audio finishes.

• Linux with paplay (PipeWire / PulseAudio).
• CPU only — no GPU, no CUDA, no Metal. That is the point: the GPU stays free for your local LLM.
• The model (99 MiB) and the voices (9 MiB) it uses must already be in your Hugging Face cache (helper scripts provided); this extension reads the cache and never downloads. Cache location: $HF_HUB_CACHE, $HF_HOME/hub, $XDG_CACHE_HOME/huggingface/hub, or $HOME/.cache/huggingface/hub.
• English only: phonon-31m-en is an English model, so only English catalog voices are offered.

The settings panel:

> Voice        Everly  (37 ready of 37 en voices)
  Volume       55%
  Temperature  0.30
  Auto speak   off

up/down move, left/right adjust the focused row, enter opens the voice picker, s speaks the last message in the highlighted voice, esc closes. In the picker, moving the cursor says each voice out loud (ready voices only), f pins a favourite, and voices whose embedding is not in your cache yet are listed greyed as not downloaded; the bootstrap section below gives the command that fetches them.

Auto-speak reads the agent's final answer aloud when a turn ends. It is scoped per project, and the row cycles off → here → everywhere.

Install once; zot runs it from then on. Nothing is ever downloaded at runtime.

Grab the binary from the latest Release named zot-speaks (x86-64 Linux) together with the SHA256SUMS file; the release notes print the same digest.

If you have just, two recipes do exactly this — they are plain curl + sha256sum + zot ext install, readable in the justfile:

git clone https://github.com/lovehumans/zot-speaks
cd zot-speaks
just install-release                    # or pin one: just install-release v1.7.8

Then run just bootstrap once, before your first /tts: zot-speaks only ever reads the Hugging Face cache, so speech stays silent — with a loud warning about the missing snapshot — until it is filled. See One-time bootstrap below.

Or fetch, verify and install by hand (no just needed):

git clone https://github.com/lovehumans/zot-speaks
cd zot-speaks
curl -fsSL -o zot-speaks https://github.com/lovehumans/zot-speaks/releases/latest/download/zot-speaks
curl -fsSL -o SHA256SUMS https://github.com/lovehumans/zot-speaks/releases/latest/download/SHA256SUMS
sha256sum -c SHA256SUMS        # must print: zot-speaks: OK
chmod +x zot-speaks
stage="$(mktemp -d)/zot-speaks"; mkdir -p "$stage"
cp extension.json "$stage/"; mv zot-speaks "$stage/"
zot ext install "$stage"

The clone supplies extension.json — the manifest zot needs — and the license; the release supplies the binary. Verify before installing: sha256sum -c fails loudly on a truncated or tampered download, and you can cross-check the printed hash against the release notes. To pin a version instead of taking the newest, replace latest with the tag in both URLs: releases/download/v1.7.8/….

Restart zot: zot ext doctor should list zot-speaks with /tts and /tts-settings active.

Updating. Download the new binary and swap it inside the installed directory; your settings (voice, favourites, per-project auto-speak) live in settings.json there and travel untouched. just update-release does the download, verification and swap.

# zot's state dir: $ZOT_HOME, else $XDG_STATE_HOME/zot, else ~/.local/state/zot
installed="${ZOT_HOME:-${XDG_STATE_HOME:-$HOME/.local/state}/zot}/extensions/zot-speaks"   # or <project>/.zot/extensions/…
tmp="$(mktemp -d)"
curl -fsSL -o "$tmp/zot-speaks" https://github.com/lovehumans/zot-speaks/releases/latest/download/zot-speaks
curl -fsSL -o "$tmp/SHA256SUMS" https://github.com/lovehumans/zot-speaks/releases/latest/download/SHA256SUMS
( cd "$tmp" && sha256sum -c SHA256SUMS )
chmod +x "$tmp/zot-speaks"
cp "$tmp/zot-speaks" "$installed/zot-speaks.new"
mv -f "$installed/zot-speaks.new" "$installed/zot-speaks"   # atomic swap on one filesystem

A running zot keeps the old binary open; the new one takes effect at the next zot start.

git clone https://github.com/lovehumans/zot-speaks
cd zot-speaks
just install-ext      # cargo build --release, then zot ext install

cargo build --release
stage="$(mktemp -d)/zot-speaks"; mkdir -p "$stage"
cp extension.json "$stage/"; cp target/release/zot-speaks "$stage/zot-speaks"
zot ext install "$stage"

just update-ext        # same directory, new binary, settings.json untouched

# zot's state dir: $ZOT_HOME, else $XDG_STATE_HOME/zot, else ~/.local/state/zot
installed="${ZOT_HOME:-${XDG_STATE_HOME:-$HOME/.local/state}/zot}/extensions/zot-speaks"   # or <project>/.zot/extensions/…
cp target/release/zot-speaks "$installed/zot-speaks.new"
mv -f "$installed/zot-speaks.new" "$installed/zot-speaks"

zot-speaks never downloads: it only reads the Hugging Face cache. Filling that cache is the one network step, and it needs the upstream ptts CLI on your PATH — get it from the official releases.

just bootstrap   # the model (~99 MiB) plus all 37 English voices (~9 MiB), once
just cache       # what the extension can see right now: snapshot, and how many voices are ready

Or in pieces. just fetch-model fetches the model and one voice in a single run (just fetch-model VOICE=Aria chooses the voice). just fetch-voices fills the rest of the English catalog and skips whatever is already cached, so re-running it is also the coverage check: it prints fetched/skipped/failed counts. Both are thin wrappers over dev/prefetch-en-voices.sh and the ptts CLI, so ptts --voice Miles "warm up" does the first half by hand too.

You can listen to individual voices here.

If you skip this step, /tts warns you: it reports the missing snapshot, the path it looked under, and just fetch-model as the fix — and it never reaches for the network.

Do this before starting zot, or restart zot afterwards: a running extension has already read its catalog, so voices you fetch mid-session join the picker on the next zot start.

Building, testing and running from a clone — the recipes a fork or a pull request uses, and the same ones CI uses: CI produces the Release binary by running cargo build --release on every version tag, so just build and the Release asset are the same kind of artifact.

just build      # cargo build --release   ->  target/release/zot-speaks
just test       # the test suite
just smoke      # streaming smoke test: prints time-to-first-sound
just lint       # cargo clippy
just run CMD="/tts hello"   # one command against the built binary

just is a thin wrapper: every build recipe is the cargo command it names, so you can run those directly instead. Use the release build for live playback; the debug build generates audio slower than real time. If cargo cannot write ~/.cargo or cannot reach crates.io, pin CARGO_HOME and CARGO_TARGET_DIR to directories you can write, run cargo fetch once with network, and build with --offline afterwards.

How it works internally, the settings file and auto-speak scoping, the voice catalog and what it takes to build, install and test this extension: docs/DESIGN.md.

Thanks to:

• xn-ptts — Gradium's Phonon inference stack. This extension builds against its ptts crate, just bootstrap drives its ptts CLI to fill the model cache, and the checkpoints it serves — phonon-31m-en and phonon-voices — are theirs as well.
• zot — the agent this extension speaks for, and the extension protocol (extension.json, slash commands) it exposes.

MIT only — see LICENSE. This covers the sources in this repository. Third-party artifacts keep their own terms: the ptts crate this builds against is MIT OR Apache-2.0, and the Phonon model and voice embeddings you cache are upstream downloads under the terms on their Hugging Face pages (also MIT OR Apache-2.0).