local-tts

Make your coding agent talk to you!

A tiny command-line text-to-speech tool. It shells out to Kokoro-82M by default — eight languages from one small model — so speech is generated locally and offline.

$ tts "Hello from my terminal."
$ echo "Read this out loud." | tts
$ tts -f chapter1.txt -o chapter1.wav

Zero runtime dependencies. The package installs nothing but itself — no requests, no numpy, no audio libraries. Everything is Python's standard library plus binaries you already have (or install once, on your terms).

Make Claude talk like Jarvis. Or Cortana. In Spanish.

Your coding agent already writes your code. With an RVC voice model it can sound like whoever you want while it does — offline, on your own machine, no API key.

$ tts --lang en "Good evening. The tests are passing, sir."
$ tts --lang es "Buenas noches. Todas las pruebas pasaron."

One rvc server holds both voices resident and picks one per language:

tts config --set rvc.base_provider=kokoro          # Kokoro speaks; RVC changes who
tts config --set kokoro.language_voices.en=bm_george
tts config --set kokoro.language_voices.es=ef_dora
tts config --set rvc.language_models.en=jarvis     # your own trained .pth
tts config --set rvc.language_models.es=cortana-es
tts languages --set en=rvc
tts languages --set es=rvc

Now every agent on the machine that knows about tts speaks in that voice, in the right language, because the preference lives in one config file rather than in one chat.

It acts, too. Tone tags change the delivery per sentence:

$ tts "<calm>Deploy finished.</calm> <urgent>But staging is down.</urgent>"

And it answers fast. Fragments start playing as soon as the first one is synthesized, so a long passage begins in about a second instead of after the whole thing renders — measured on a 19-span story, 9.6s to first sound instead of 40.5s.

Bring your own .pth — local-tts wires up a voice model you already have or trained; it does not distribute voices, and cloning a real person's voice without their consent is not what this is for.

See Providers → rvc for the full setup.

Contents

• Make Claude talk like Jarvis. Or Cortana. In Spanish.
• Requirements
• Install
  • Install with an AI agent
• Updating
• Quick start
• Usage
• Background playback
• Coding-agent skills
• Status-bar hook
• Language memory
• Providers
• Configuration
• Audio playback
• Troubleshooting
• Development
• Thanks

Requirements

Installing a speech backend

local-tts does not bundle, build or ship any speech model. It drives a backend you install separately, and the default provider is kokoro — small, fully offline, and good for eight languages. Its setup lives with the provider itself:

Run kokoro and rvc as a persistent server. Both reload their model from disk on every call otherwise, which dominates the time you wait. See the server section — it is the difference between roughly a second per sentence and several.

Prefer to be walked through it? Install with an AI agent does the whole thing, asking before each step.

Speech models

Model weights are the backend's business, not this tool's: kokoro needs kokoro-v1.0.onnx plus voices-v1.0.bin (fetched once — see the kokoro section), piper needs a .onnx voice per language (~60 MB each), and llamacpp downloads its default OuteTTS weights on first use. Every one of them works fully offline afterwards.

Each backend points at its own weights through its own settings; see its section under Providers.

Install

Install the published package in an isolated environment:

pipx install agents-local-tts
# or, inside your own virtual environment:
python -m pip install agents-local-tts

The distribution is agents-local-tts; commands remain tts and local-tts, and imports remain localtts. Speech models are installed separately. For development, clone the repository and use the editable install below. Legacy source installations named local-tts should uninstall that distribution before installing the renamed package in the same environment, to avoid competing entry points.

Everything happens inside a virtual environment so nothing touches your system Python.

git clone <this-repo> local-tts
cd local-tts

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

python -m pip install --upgrade pip   # needs pip >= 24.2 for the package metadata
pip install -e .

That puts two equivalent commands on your PATH (while the venv is active): tts and local-tts.

# pipx keeps the tool isolated but always on your PATH
pipx install .

# or symlink the venv entry point somewhere on your PATH
ln -s "$PWD/.venv/bin/tts" ~/.local/bin/tts

Install with an AI agent

If you use an AI coding agent (Claude Code, Cursor, Copilot, …), the whole setup — including detecting what is already on the machine, installing the backends, and creating the global symlink — is scripted for it in AGENT_INSTALL.md.

Point your agent at this repository and say what you want:

"Install this with the link."

or, more specifically:

"Read AGENT_INSTALL.md and install local-tts with piper in Spanish, and make it global."

The agent will detect what you already have, ask once about anything missing, and validate the result with tts check plus a real synthesis. It is written to ask before installing anything, downloading a model, creating a symlink, or running sudo — so you approve each decision rather than discovering it afterwards.

Phrases it understands without further questions:

Prefer doing it yourself? Everything the agent does is the same set of commands documented in Requirements, Install and Providers below.

Verify the install

tts check

config file : /home/you/.config/local-tts/config.json (not created yet)
default     : kokoro

[ok] llamacpp  /usr/local/bin/llama-tts -> default OuteTTS (downloaded on first run)
[--] openai    https://api.openai.com/v1 (no api_key and $OPENAI_API_KEY is unset)
[--] piper     piper: 'piper' not found on PATH. ...
[ok] kokoro    /home/you/.local/bin/kokoro-tts
[--] rvc       rvc needs the python interpreter from the venv rvc-python is installed in: ...
[ok] command   espeak-ng -w {output} {text}

players     : ffplay, paplay  -> using ffplay
tone shaping: ffmpeg atempo (best quality)
phonetics   : no /IPA/ entries in `pronunciations` (plain respellings work everywhere)
streaming   : on -- each fragment plays as it is synthesized

Only the line matching your default provider has to say [ok]. The last three lines report what shapes the audio: tone shaping says whether ffmpeg is doing the retiming or the slower built-in fallback is, and phonetics says which backends can accept the /IPA/ entries in your dictionary — see Pronunciation dictionary.

Updating

For a PyPI installation, run pipx upgrade agents-local-tts (or python -m pip install --upgrade agents-local-tts in its environment). Source installations use the git workflow below. Both need tts skills --install and tts servers --refresh to refresh copied resources. There is no automatic package update.

cd local-tts               # the repo you cloned in Install
git status --short         # make sure there's nothing uncommitted first
git pull

# editable install (the default): src/ changes are live immediately. Rerunning this
# is still worth it — it's a no-op most of the time, but it's what picks up a
# pyproject.toml change (entry point, version, python floor), and it costs nothing
# since the package has zero runtime dependencies
pip install -e .

# pipx install instead? pipx never re-reads the source directory on its own:
pipx install . --force

Three more things are snapshots taken at install time, not symlinks into the repo, so pulling doesn't update them on its own:

tts skills --install     # refreshes the skill copies every detected agent is reading
tts hooks --install      # only if `tts hooks --status` shows one is active
tts servers --refresh    # the kokoro/rvc server scripts, which live in their own venvs

Then tts --version and tts check to confirm it landed.

The server script is the one that goes stale in silence: it is a copy in the backend's venv, an older one answers /health perfectly well, and it drops any request key it never learned to read — so a new capability just quietly does nothing. tts servers compares what is installed against this version and says which is which.

If you use a coding agent, the whole thing — including finding the repo behind whatever install method you used — is the local-tts-update skill from Coding-agent skills; just say "update local-tts."

Quick start

# speak an argument
tts "The quick brown fox jumps over the lazy dog."

# speak a pipe
git log -1 --format=%s | tts

# speak a file, save the audio instead of playing it
tts -f notes.md -o notes.wav

# save and play
tts -o greeting.wav --play "Good morning."

# narrate a markdown document: syntax stripped, long text chunked and joined
tts -f README.md -o readme.wav

# see the exact backend command (and the chunk plan) without running it
tts --dry-run -f README.md

The first invocation is slow: it downloads the model. After that a short sentence takes a couple of seconds on CPU.

Usage

tts [options] [TEXT ...]
tts providers
tts check
tts config [--show | --path | --init | --set KEY=VALUE]

Input precedence is TEXT → --file → stdin. Without --output, audio goes to a temporary file that is played and then deleted (--keep keeps it).

Markdown is handled for you. Reading a .md file strips headings, emphasis, link URLs, bullet markers, tables and fenced code blocks before synthesis, so none of it gets read aloud. Override either way with --markdown / --no-markdown.

Long documents are handled for you too. Backends that need short prompts (llamacpp) get the text split at sentence boundaries, synthesized piece by piece, and joined into a single file with a short pause between pieces. Backends that manage long input themselves (piper) receive it whole. See max_words below.

Exit codes: 0 success, 1 error (with a one-line message on stderr), 130 interrupted.

Background playback

--background (-b) starts playback detached and returns straight away, so a long file does not block the shell — or an agent driving it. The file is kept and its path printed.

$ tts -b --lang es -f documento.md
playing in the background (pid 4123, 0:12) — `tts stop` to end it, `tts playback` for progress
/tmp/local-tts-a1b2c3d4.wav

Control it afterwards, with elapsed time tracked against the file's real duration:

$ tts playback
playing [###########---------] 0:03 / 0:05 (pid 4123): /tmp/local-tts-a1b2c3d4.wav

tts pause
tts resume
tts stop

Starting a new background playback stops the previous one in the same session, so voices never stack there. Separately, playback is also serialized machine-wide: if another session (a second agent, a second terminal) already has audio playing, a new -b call queues behind it instead of overlapping — only one file ever plays at a time, no matter which provider or session started it. The CLI call itself still returns immediately either way; only the actual audio start is deferred. tts playback shows 0:00 and holds there while queued, then starts advancing once its turn begins.

pause/resume use SIGSTOP/SIGCONT and therefore work on Linux, macOS and WSL; on native Windows they report that they are unsupported and stop is the control.

Coding-agent skills

local-tts ships five skills that teach a coding agent to use it, and installs them into whichever agents it finds on your machine:

• local-tts-speak — speak to the user. Triggers on "talk to me", "read this aloud", "narrate this file", "háblame", and so on. Instructs the agent to check the language memory first, to always use -b (and to run the command itself non-blocking), to play the whole thing regardless of length unless told otherwise, to offer stop/pause/resume, to always report the file path, and never to read secrets out loud.
• local-tts-configure — install, diagnose and configure: backends, voices for a new language, playback, and the per-language memory. Starts from tts check and asks before installing or downloading anything.
• local-tts-tune — make it sound better once it already works: pacing, pauses, emphasis, which voice reads a borrowed word, and the noise or robotic artifacts that come from the wrong player or a missing ffmpeg. Diagnoses by measurement first, changes one setting at a time, and asks for your ears only where a measurement cannot decide.
• local-tts-phonetics — get one word said properly: a name, a brand, an acronym, a borrowed term inside another language. Looks the transcription up rather than deriving it from spelling, tries it by ear with tts pronounce, reads back which phonemes the model actually has a token for, and keeps the winner in the dictionary — or wires a phonetics_hooks script when there are too many words to type.
• local-tts-update — update an already-installed CLI to the latest version. Locates the repo behind the running tts command, pulls it, reinstalls only if that's actually needed, and refreshes the skill/hook files that are copies rather than live links to the repo — including its own file: since the update process can change between versions, it refreshes and re-invokes itself right after pulling, rather than finishing the rest of the update under instructions that might already be stale. See Updating.

tts skills                       # what was detected, and what is installed
tts skills --install             # install into every detected agent
tts skills --install gemini      # or just one
tts skills --install --dry-run   # show the paths without writing
tts skills --uninstall           # remove them again
tts skills --print local-tts-update   # print one skill's current content to stdout

--print reads straight from this install, not the copy sitting in any agent's skill directory — useful for a host that won't reliably pick up a changed skill file mid-session (local-tts-update uses it on itself for exactly that reason, see Updating).

Restart the agent (or open a new session) afterwards so it picks them up.

<config> is %APPDATA% on Windows and ~/.config on Linux and macOS ($XDG_CONFIG_HOME wins on any platform when set). Detection only writes where the agent's directory already exists, so nothing is created for agents you do not use.

All listed agents receive native skills. Reinstalling all skills migrates older managed Codex, Cursor, Windsurf and Copilot instruction blocks, preserves unrelated content, and backs up the old file. Agent permissions and skill discovery remain controlled by the agent; restart its session after installing. See verified integration paths.

Status-bar hook

By default, speaking prints a status line in chat each time. Two coding agents can instead show live progress in their own status bar — verified against their actual settings schemas, not assumed:

tts hooks                # what's detected, installed, and why the rest can't do this
tts hooks --install      # install into every detected supported agent
tts hooks --status       # is a hook live right now? (exit 0/1; used by the skill)
tts hooks --uninstall    # remove it, restoring whatever status line was there before

$ tts hooks
supported : claude-code, qwen

[ok] claude-code  active
[  ] qwen         agent not detected

[xx] codex        not supported: no status line mechanism yet (open feature request upstream)
[xx] copilot      not supported: has one, but its config schema isn't documented solidly enough to target yet
[xx] cursor       not supported: would need a full VS Code extension, not a lightweight hook
[xx] gemini       not supported: footer settings are show/hide toggles only; no custom command
[xx] opencode     not supported: no status line mechanism yet (open feature request upstream)
[xx] windsurf     not supported: would need a full VS Code extension, not a lightweight hook

Only these two have a documented "run my command, show its stdout in the status bar" mechanism today. Gemini CLI's footer is hide/show toggles only (checked its shipped settingsSchema.js); Codex CLI and OpenCode both have open upstream feature requests for this, not yet shipped; Cursor and Windsurf are VS Code forks where a status-bar item means writing a real extension, not a lightweight hook; GitHub Copilot CLI has one, but its config schema isn't documented solidly enough to target without an install to test against.

Install never rewrites an existing status line — it appends into it. If your settings already point at a script (yours, or another tool's), that pointer is never touched; instead a small block is added to the end of that script file, so the original tool keeps owning its slot and keeps running exactly as it always did. Idle, output is byte-for-byte what it was before — our block only adds text while something is actually playing:

$ tts hooks --install claude-code
  claude-code  did append into /home/user/.claude/statusline-command.sh -- your existing
               status line is untouched, and picks this up on its very next refresh

This only appends into a plain path to a writable script file — not a one-liner, not a command with arguments, not something unwritable. If the existing command doesn't qualify, install refuses and shows you the exact block to add by hand, or you can pass --force to replace the pointer outright (the old chain-by-reference behavior — the existing command still runs, but the tool that owned it no longer does, which is a real tradeoff, not a free upgrade; only reach for it when appending genuinely isn't possible). Reinstalling replaces our block in place rather than duplicating it. --uninstall removes only our block and leaves the rest of the file untouched, or drops the settings key entirely if nothing was configured before we installed.

Appended mode takes effect on the very next status-bar refresh — no restart needed, since only the script's content changed, not anything Claude Code reads once at startup. A fresh install with nothing configured before (or --force) does need a restart, since those set statusLine.command/refreshInterval directly. When a hook is live, the local-tts-speak skill stops printing its own status line — tts hooks --status is what it checks.

Refresh cadence

With nothing else configured, a fresh install defaults to a real 2-second timer. When appending into an existing status line, the existing refresh cadence — timer or event-only — is left exactly as it was by default, since changing it also changes how often the other tool's own script re-runs, not just ours:

tts hooks --install claude-code --refresh-interval 2   # a real timer, ticks live
tts hooks --install claude-code --refresh-interval 0   # explicitly event-based, no timer
tts hooks --install claude-code                        # leave whatever cadence was already there

0 is a deliberate choice, not the same as omitting the flag — it removes refreshInterval outright (so the status bar only redraws on host events like a new message), whereas omitting the flag means "don't decide, leave it as configured." Changing the cadence (anything other than "leave it as configured") does write to settings.json — just the refreshInterval key, never command — so that one does need a restart.

Multiple sessions

Running more than one session at once (two terminals, two agent instances) works without one's audio stopping another's or its status bar showing the wrong progress. Pass --session with anything that identifies the run:

tts -b --session "$CLAUDE_CODE_SESSION_ID" "hello"
tts stop --session "$CLAUDE_CODE_SESSION_ID"

Playback state is stored per session; starting playback only stops a previous playback from the same session, and stop/pause/resume/playback only ever act on your own session's audio. What two sessions can't do is talk over each other: actual audio output is serialized machine-wide (see Background playback) — a second session's -b call queues rather than overlapping, it just doesn't stop or otherwise touch the first session's state while it waits. --session is auto-detected when omitted — currently from $CLAUDE_CODE_SESSION_ID, verified by capturing a live status-line payload from Claude Code and confirming it carries the exact same value in its session_id field, which is also how the status-bar hook knows which session's progress to show. Omit --session entirely and everything works exactly as before it existed — one shared slot.

Language memory

Which backend speaks which language is remembered in the config file, so the preference survives sessions and is shared by every agent rather than living in one agent's memory.

tts languages                                     # show what is recorded
tts languages --set es=piper:~/voices/es_MX.onnx  # provider + voice
tts languages --set en=llamacpp                   # provider only
tts languages --forget de

Then just name the language:

tts --lang es "Hola, ya terminé."
tts --lang es -f documento.md -o documento.wav

The lookup prefers the specific tag over the base one, so with both recorded, --lang es-MX picks the Mexican entry while --lang es picks the generic one. Explicit flags always win over the memory, and a recorded voice is only applied to the provider it was recorded for — a piper .onnx is never handed to llama.cpp.

This is what the agent skills write to when you give feedback like "use piper for Spanish" or "that accent is wrong, use the Mexican voice". You can also set it per shell with LOCALTTS_LANG_ES=piper:/path/voice.onnx.

Providers

tts providers

kokoro is the default. The others are there because each one wins at something the default does not do: a voice you trained yourself (rvc), a language kokoro reads badly (piper), a hosted endpoint (openai), or a binary you already have (command).

llamacpp — local, offline, four languages

Runs llama-tts. Not the default — it speaks English, Chinese, Japanese and Korean only, and reads every other language with English phonetics. Zero configuration: with no model set it passes --tts-oute-default and llama.cpp handles the weights.

local-tts calls the llama-tts binary; it does not bundle or build llama.cpp:

# macOS / Linux (Homebrew)
brew install llama.cpp

# Windows
winget install llama.cpp

# Prebuilt binaries for every platform
# https://github.com/ggml-org/llama.cpp/releases  (grab llama-<build>-bin-<platform>.zip)

# Or build from source
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release -j
# binaries land in build/bin — put that on your PATH, or see "Configuration" below

Verify it is reachable, and point local-tts at it if it is not on your PATH:

llama-tts --version
tts config --set llamacpp.binary=/path/to/llama.cpp/build/bin/llama-tts

On the first run it fetches its default OuteTTS weights plus the WavTokenizer vocoder into the Hugging Face cache (~/.cache/huggingface/hub, about 640 MB). Later runs use the cache and work offline.

Output is 24 kHz mono WAV. The default OuteTTS weights speak English, Chinese, Japanese and Korean; other languages come out with English phonetics, so use the default kokoro or piper for those. Quality also drops on long prompts, which is why max_words splits them — raise or lower it to trade continuity against reliability.

Using your own models

tts config --set llamacpp.model=~/models/OuteTTS-0.2-500M-Q8_0.gguf
tts config --set llamacpp.vocoder=~/models/WavTokenizer-Large-75-F16.gguf

Or fetch them from Hugging Face at run time:

tts config --set llamacpp.hf_repo=OuteAI/OuteTTS-0.2-500M-GGUF
tts config --set llamacpp.hf_file=OuteTTS-0.2-500M-Q8_0.gguf
tts config --set llamacpp.hf_repo_vocoder=ggml-org/WavTokenizer
tts config --set llamacpp.hf_file_vocoder=WavTokenizer-Large-75-F16.gguf

Speed it up with your own hardware settings:

tts -s threads=8 -s gpu_layers=99 "offloaded to the GPU"

openai — any OpenAI-compatible endpoint

Speaks HTTP (POST /v1/audio/speech) using urllib — no SDK involved. It works with OpenAI itself and with local servers such as openedai-speech, Kokoro-FastAPI, or LocalAI.

tone is flat voice-style instructions sent with every call (model=gpt-4o-mini-tts only); auto_tone derives tone from ?/!/. where no <tag> is active in the text. See Tone and emotion tags below.

export OPENAI_API_KEY=sk-...
tts -p openai -v nova "Hello from the cloud."

# a local server needs no key at all
tts config --set openai.base_url=http://localhost:8880/v1
tts -p openai -o out.mp3 "Local, but OpenAI-shaped."

This is the only provider that writes formats other than WAV — the output extension picks the format (wav, mp3, opus, aac, flac, pcm).

piper — small, fast, offline, many languages

Piper runs neural ONNX voices on the CPU at roughly 7x realtime, with good models for ~40 languages. Reach for it when your language is not one of kokoro's eight — German, Dutch, Polish, Russian, Turkish and the rest live here, not there — when you want a different voice for a language kokoro does cover, or when you want one flat .onnx file per voice instead of kokoro's wrapper script.

Piper is distributed as a Python wheel (piper-tts, GPL-3.0). Install it in its own virtualenv so its ~200 MB of dependencies (onnxruntime, numpy) stay out of this project, then point local-tts at the binary:

python -m venv ~/.local/share/piper-venv
~/.local/share/piper-venv/bin/pip install piper-tts

# list every voice, then fetch the one you want (~60 MB for "medium", ~63 MB for "high")
mkdir -p ~/.local/share/piper-voices && cd ~/.local/share/piper-voices
~/.local/share/piper-venv/bin/python -m piper.download_voices            # list
~/.local/share/piper-venv/bin/python -m piper.download_voices es_MX-claude-high

tts config --set piper.binary=~/.local/share/piper-venv/bin/piper
tts config --set piper.model=~/.local/share/piper-voices/es_MX-claude-high.onnx
tts -p piper "Piper es muy rápido en una CPU."

Voice weights live at rhasspy/piper-voices (MIT/CC, no account or API token needed). Naming is <lang>_<REGION>-<speaker>-<quality>, where quality is x_low, low, medium, or high.

Piper splits long input into sentences by itself, so an entire document works in one call:

tts -p piper -f article.md -o article.wav

kokoro — the default: small, fast, offline, eight languages

This is the provider you get with no configuration at all. Kokoro-82M covers eight languages in a footprint comparable to piper's: English (US and UK), Spanish, French, Italian, Portuguese (Brazil), Hindi, Japanese and Mandarin — the set is fixed by the model's own voice prefixes (VOICE_LANGS in src/localtts/providers/kokoro.py). For anything outside it, use piper. There is no single official CLI, so the straightforward path is a minimal wrapper around the kokoro/kokoro-onnx Python package, in its own venv:

python -m venv ~/.local/share/kokoro-venv
~/.local/share/kokoro-venv/bin/pip install kokoro-onnx soundfile

mkdir -p ~/.local/share/kokoro-models && cd ~/.local/share/kokoro-models
# fetch kokoro-v1.0.onnx and voices-v1.0.bin, e.g. from
# https://github.com/nazdridoy/kokoro-tts/releases

Save this as ~/.local/share/kokoro-venv/kokoro_cli.py:

#!/usr/bin/env python3
import argparse, os, sys, warnings
warnings.filterwarnings("ignore")
MODELS = os.path.expanduser("~/.local/share/kokoro-models")

def main():
    p = argparse.ArgumentParser()
    p.add_argument("-o", "--output", required=True)
    p.add_argument("-v", "--voice", default="af_heart")
    p.add_argument("-l", "--lang", default="en-us")
    p.add_argument("-s", "--speed", type=float, default=1.0)
    p.add_argument("text", nargs="*")
    a = p.parse_args()
    text = " ".join(a.text).strip() or sys.stdin.read().strip()
    if not text:
        sys.exit("kokoro-tts: no text")
    import soundfile as sf
    from kokoro_onnx import Kokoro
    k = Kokoro(f"{MODELS}/kokoro-v1.0.onnx", f"{MODELS}/voices-v1.0.bin")
    samples, rate = k.create(text, voice=a.voice, speed=a.speed, lang=a.lang)
    sf.write(a.output, samples, rate)

if __name__ == "__main__":
    main()

And a shim on PATH so kokoro.binary's default (kokoro-tts) finds it:

mkdir -p ~/.local/bin
cat > ~/.local/bin/kokoro-tts <<'EOF'
#!/usr/bin/env bash
exec "$HOME/.local/share/kokoro-venv/bin/python" \
     "$HOME/.local/share/kokoro-venv/kokoro_cli.py" "$@"
EOF
chmod +x ~/.local/bin/kokoro-tts

Then configure a voice — IDs are per-language, e.g. af_heart for US English, ef_dora for Spanish:

tts config --set kokoro.voice=ef_dora
tts config --set kokoro.lang=es
tts -p kokoro "Prueba de voz con Kokoro."

Once this works, set up the persistent server — the wrapper above reloads the model on every single call, and the server is both much faster and the only way to get emphasis_lengthen and IPA phonetics, which live in the phonemizer the server holds.

Per-language voices

Kokoro names every voice by language — a/b English, e Spanish, f French, and so on — so one flat voice cannot serve two languages. Map them instead:

tts config --set kokoro.language_voices.en=bm_george   # British male
tts config --set kokoro.language_voices.es=ef_dora     # Spanish female

An exact tag beats its base language (es-MX before es), and the phonemizer language is taken from the chosen voice rather than from lang — otherwise switching languages leaves a stale lang behind and one language gets read with another's phonetics.

Emphasis, and kokoro's own pauses

tts config --set kokoro.emphasis_lengthen=2   # IPA length marks on the stressed vowel
tts config --set kokoro.sentence_pause=0.25   # kokoro's own within-utterance pauses
tts config --set kokoro.clause_pause=0.1

emphasis_lengthen is emphasis the way a phonetician writes it: N length marks on the vowel carrying primary stress, kˈasa → kˈaːsa. Kokoro has ː in its own vocabulary, so the model hears it — an isolated word measures 0.576s plain, 0.640s with one mark, 0.661s with two. It needs the persistent server, which is where the phonemizer lives, and 0 disables it.

rvc — voice conversion (not installed automatically)

rvc-python does audio-to-audio voice conversion only — it has no text input. This provider chains it: synthesize with another provider first (rvc.base_provider, piper by default), then convert that result to a target voice with a trained model.

Never installed automatically — it pulls in torch and is sizable. Set it up deliberately:

python -m venv ~/.local/share/rvc-venv
~/.local/share/rvc-venv/bin/pip install rvc-python
# GPU support needs a matching torch build; see the rvc-python README for the index URL.

tts config --set rvc.python=~/.local/share/rvc-venv/bin/python
tts config --set rvc.model=~/.local/share/rvc-models/<name>/<name>.pth
tts config --set rvc.index=~/.local/share/rvc-models/<name>/<index-file>   # optional
tts config --set rvc.base_provider=piper

tts -p rvc --dry-run "test"    # shows both steps: base synthesis, then conversion
tts -p rvc "Test of the converted voice."

There is no rvc.voice — the voice comes entirely from which .pth model is configured.

Recommended: a persistent server (kokoro / rvc)

For kokoro and rvc, this is the way to run them. Both reload their model from disk on every call otherwise, and that load — not the synthesis — is most of what you wait for: a sentence that comes back in about a second from a warm server pays several seconds, every single time, without one. rvc is the worse case, because the cost includes importing torch.

Three things only the server can do:

• Hold several voices at once. One process, one copy of torch, one GPU context, N voices — which is what makes a second language cheap rather than a second server.
• Phonetics and emphasis. kokoro.emphasis_lengthen and IPA pronunciations need the phonemizer, and the phonemizer lives in the server. The subprocess wrapper silently does without them.
• Keep the GPU context warm, instead of rebuilding it per sentence.

The cost is one background process holding RAM (and VRAM) — released on its own after five idle minutes, and restarted transparently on the next call.

Either provider talks to a small server that loads the model once and serves requests over localhost — this tool never runs that server itself, it's a short script you write into the provider's own venv (the local-tts-configure skill has the exact script for both, self-contained, stdlib http.server only):

tts config --set kokoro.server_url=http://127.0.0.1:8765
tts config --set 'kokoro.server_start=~/.local/share/kokoro-venv/bin/python ~/.local/share/kokoro-venv/kokoro_server.py --port 8765'
tts -p kokoro "test"     # auto-starts it on first use (a few seconds), fast after that

Auto-start polls server_url's /health until it answers or server_timeout elapses, then the request goes through as an HTTP POST instead of a subprocess call — voice, language and speed still travel per request, not fixed to whatever the server started with. tts check reports whether it's running or will auto-start, without starting it itself. It exits on its own after 5 minutes idle to release the model (a real cost for rvc's torch model in particular) — configurable via --idle-timeout SECONDS in server_start (0 disables it); the next call after it exits just starts a fresh one.

For rvc, one server can hold several models at once: start it with a repeatable --model NAME=PATH per voice, and each request names the one it wants — rvc.language_models picks it per language, with rvc.server_model as the flat default. A server started with a single --model and no names still works; it just answers every request with the one voice it loaded.

Keeping the script current. Neither server script is part of this package — each is a copy written into that backend's venv, so git pull and pip install -e . never touch one. tts servers reads the script out of the bundled local-tts-configure skill and compares it with what is actually on disk, wherever server_start points:

$ tts servers
[ok] kokoro  script is current -- ~/.local/share/kokoro-venv/kokoro_server.py (running)
[!!] rvc     script is STALE -- ~/.local/share/rvc-venv/rvc_server.py (not running)

`tts servers --refresh` rewrites it from the bundled template.

--refresh writes the current script, keeps the previous one as <name>.bak, and asks a running server to exit so the next call starts the new one. STALE only means differs from this version's template — a script you edited on purpose reads the same way, which is why the old copy is kept rather than replaced outright. A server installed before /shutdown existed cannot be stopped this way; --refresh says so, and that one ages out on its own idle timeout.

Tone and emotion tags

Wrap a stretch of text in <name>...</name> to mark its tone, e.g. <happy>Good news!</happy> <serious>One thing needs your review.</serious>. Any word works as a tag name; a built-in preset exists for common ones (anger, happy, joy, sad, fear, surprise, disgust, calm, excited, serious, whisper, sarcastic, urgent, gentle, confident, tired, playful, question, exclamation — see TAG_PROFILES in src/localtts/text.py), and anything else still works with a generic phrase, just without a hand-tuned preset. Tags can nest (<serious><question>...</question></serious> combines both) and escape a literal angle bracket with \< / \>.

tts -p openai --model gpt-4o-mini-tts "<happy>Good news!</happy> <serious>One thing needs your review.</serious>"
tts -p piper "<whisper>Very quiet now.</whisper> Back to normal."

What a tag actually does depends on the backend, since not every one has a real hook for it — see each provider's own table above (openai.tone/auto_tone, piper.auto_tone, kokoro.auto_tone) and the local-tts-configure skill for the full breakdown. On a backend with no hook at all (llamacpp, rvc), a tag is a safe no-op: it's always stripped before the text is spoken, never read out literally.

command — anything else

An escape hatch for any binary that can write a WAV file. {text} and {output} are substituted as single argv items, so text is never re-parsed by a shell.

tts config --set 'command.template=espeak-ng -w {output} {text}'
tts -p command "Whatever tool you like."

# macOS
tts config --set 'command.template=say -o {output} --data-format=LEI16@22050 {text}'

By default, a <tag> is stripped before {text} is filled in, like any provider above with no real tone hook. If your own script is written to parse the markup itself, opt in with tts config --set command.tone_tags=pass.

Also unlike every other provider, local-tts does not reshape what your command produced. Elsewhere its capabilities are known, so a tag's leftover speed and volume are safely applied to the rendered audio; here the script is yours and may already be varying its own delivery. Opt in only if it doesn't:

tts config --set command.audio_fx=true

If you wired up something through command that now has a real provider above (kokoro, rvc), tts config --detect-migrations finds it and prints the exact --set commands to switch — it never applies them, and never touches command.template itself:

$ tts config --detect-migrations
command.template runs something tts now supports natively as `kokoro`:
  command.template runs 'kokoro-tts', which local-tts now supports natively
  tts config --set kokoro.voice=ef_dora
  tts config --set kokoro.lang=es

Configuration

Settings are resolved in this order, later winning:

built-in defaults  <  config file  <  environment variables  <  CLI flags

Config file

tts config --path      # where it lives
tts config --show      # the effective configuration, defaults included
tts config --init      # write a file containing every default, ready to edit

Default location:

$LOCALTTS_CONFIG overrides the path entirely. The file only needs to contain what you change:

{
  "provider": "kokoro",
  "play": true,
  "providers": {
    "kokoro": {
      "voice": "ef_dora",
      "server_url": "http://127.0.0.1:8765"
    }
  }
}

Write to it from the CLI:

tts config --set provider=piper
tts config --set kokoro.voice=bm_george
tts config --set play=false

Top-level keys are provider, play (play by default when no --output), and player (force a playback command). Everything else is <provider>.<key>.

Environment variables

Per-run overrides

-s/--set changes a setting for one invocation only:

tts -s threads=4 "just this once"
tts -p openai -s model=tts-1-hd -s speed=1.15 "faster, nicer"

Phonetics — a borrowed word, said properly

Real speech mixes languages. "Ya subí el pull request" read entirely with Spanish phonetics sounds wrong, because "pull request" is English. Give the dictionary its transcription and the same voice says it correctly:

tts config --set 'pronunciations.pull request=/pˈʊl ɹᵻkwˈɛst/'
tts --lang es "Ya subí el pull request al repositorio."

A value between slashes is IPA rather than a respelling. Slashes are the phonetician's own notation for a phonemic transcription, so the file reads the way the reference material does, and no real respelling starts and ends with one.

Any language works. IPA is not tied to one: /ˈkʁwasɑ̃/ for a French word inside Spanish is the same mechanism as an English one. What limits it is the backend, not this table — a model can only produce the phonemes its own vocabulary contains, so a sound it was never trained on comes out as the nearest thing it has.

Not every backend can use them. local-tts has no runtime dependencies and cannot transcribe text itself, so it passes the table to backends that have a phonemizer of their own. No extra install is involved: kokoro-onnx already requires phonemizer and espeakng-loader, and the server uses kokoro's own tokenizer, so a transcription matches what the model would have produced from the text itself.

tts check asks the server rather than assuming. A server_url says a URL was written down, not that anything is listening, and a server copied from an earlier version of the skill answers /health perfectly well while dropping a table it never learned to read. Today that is kokoro with server_url set and a current server script, and rvc when kokoro is its base. Anything else comes out as ignored, which tts check says outright rather than leaving a silent no-op:

phonetics   : 2 /IPA/ entries -> kokoro, rvc; ignored by llamacpp, openai, piper, command

An ignored entry is not an error — the word is still spoken, just the backend's own way.

Why this and not one voice per language. Synthesizing a borrowed word separately and splicing it in gives that word its own end-of-sentence intonation, which mid-sentence reads as an interruption, and leaves a seam at each edge. Transcribing the whole line and swapping in phonemes keeps one utterance, one voice and one intonation curve. Measured on a sentence with three English words: 4.651s spliced against 4.020s as one utterance, the difference being dead air at six fragment edges.

Older text may still contain <en>…</en> markup. It is recognized and removed rather than read aloud or mistaken for a tone tag.

Migrating from language spans

piper.language_tags, kokoro.language_tags, rvc.delivery.*.language_tags, foreign_voices and foreign_models are gone. An existing config file containing them still loads — unknown keys are ignored, nothing crashes — but tts config --set no longer accepts them, and anyone who had foreign_voices set loses that behaviour.

Replace each borrowed word with a dictionary entry:

# before: a span, a second voice, and a seam at each edge
tts --lang es "Ya subí el <en>pull request</en>"

# after: one entry, one voice, one utterance
tts config --set 'pronunciations.pull request=/pˈʊl ɹᵻkwˈɛst/'
tts --lang es "Ya subí el pull request"

language_voices and language_models are not affected: they pick the voice for the call's own language, which is the language memory (tts languages), a separate feature.

Pronunciation dictionary

Say these words this way. One table, two kinds of entry:

tts config --set pronunciations.jarvis="JAR-viss"          # respelling: every backend
tts config --set pronunciations.kubectl="cube cuddle"
tts config --set 'pronunciations.pull request=/pˈʊl ɹᵻkwˈɛst/'   # IPA: see Phonetics
tts config --set pronunciations.es:jarvis="yarvis"         # Spanish only
tts config --set pronunciations.jarvis=                     # empty value removes it

A plain value is a respelling, rewritten into the text before synthesis, so it works on every backend. A value between slashes is IPA, handed to the model as phonemes — see Phonetics for which backends accept it.

Keys match whole words, case-insensitively; the replacement is used exactly as written, because a respelling's own capitalization is often load-bearing. A bare key applies to every language, and <lang>:<word> applies to that one only — so a word said differently in two languages needs no nested structure. Tone-tag markup is left untouched: an entry for happy will not rewrite <happy>.

Hearing a transcription before you keep it

An IPA entry is not guessable from spelling, and a wrong one fails in the way that is hardest to notice: the word still comes out, just mangled. tts pronounce renders the word as it is said now and as a candidate would say it, plays both back to back, and says whether the model has a token for every phoneme you asked for:

$ tts pronounce "pull request" --lang es --ipa "/pˈʊl ɹᵻkwˈɛst/"
word       : pull request
language   : es
provider   : rvc
now        : 0.88s
candidate  : /pˈʊl ɹᵻkwˈɛst/
phonemes   : every one is in this model's vocabulary
with IPA   : 0.73s
  playing as it is now
  playing with the candidate
keep it    : tts config --set 'pronunciations.es:pull request=/pˈʊl ɹᵻkwˈɛst/'

The phonemes line is the part that saves time. A character the model has no token for is dropped in silence, so it is reported before you listen:

$ tts pronounce croissant --lang es --ipa "/kr'wasɑ̃/"
phonemes   : no token for "'" -- it is dropped, and the word comes out mangled rather
             than wrong-but-whole. Usually a typo: an ASCII letter where IPA wants its
             own symbol.

That is nearly always a real typo — ' for ˈ, r for ɹ, : for ː. Add --sentence "quiero un croissant" to hear the word in context, --no-play to render without playing, which retains both WAV files and prints their directory. With playback, use --keep to retain the files. Asking which phonemes exist needs a current kokoro server (tts servers); without one you still get both renders, with a line saying the vocabulary could not be checked.

If you use a coding agent, the whole loop — look it up, hear it, keep it — is the local-tts-phonetics skill: just say "it's saying my name wrong."

Phonetics hooks — transcriptions nobody wrote down

The dictionary is the right place for a handful of names and borrowed terms someone typed in by hand, and the wrong place for anything generated: a lexicon, a team's glossary, a real grapheme-to-phoneme transcriber. Those want to run at synthesis time, and they want to be somebody else's code — this package has no runtime dependencies and is not going to grow a transcriber of its own.

So a hook is any executable you name. It receives the resolved table for one utterance on stdin as JSON and prints the table to use:

tts config --set phonetics_hooks='["~/bin/lexicon.py"]'
tts config --set phonetics_hook_timeout=5

#!/usr/bin/env python3
import json, sys
call = json.load(sys.stdin)
#   {"text": "ya subí el pull request", "lang": "es", "provider": "kokoro",
#    "phonetics": {"pull request": "pˈʊl ɹᵻkwˈɛst"}}
table = call["phonetics"]
for word in ("croissant", "déjà vu"):
    if word in call["text"].lower():
        table[word] = my_lexicon[word]          # bare IPA, or /between slashes/
print(json.dumps({"phonetics": table}))

Hooks run in the order listed, each seeing what the previous one returned, and may add, rewrite or drop entries. They cannot change the text — a misbehaving hook can never alter what gets said, only how a word in it is transcribed, and words it does not name come out exactly as they would with no hook at all.

Everything that can go wrong — a non-zero exit, output that is not JSON, a script that hangs past phonetics_hook_timeout — leaves the table as the previous hook left it and prints one line to stderr. Speech that is slightly wrong beats speech that does not happen. A .py without the executable bit is run with the current interpreter rather than refused. tts check's phonetics line counts the hooks it will run.

Two things worth being clear about: a hook is a program you name in your config, so it runs with your privileges — the same trust you extend to server_start. And it is unrelated to tts hooks, which installs the status-bar hook.

Delivery: pacing and pauses

How each language is delivered, on top of whatever a tone tag asks for:

tts config --set 'rvc.delivery.es={"speed": 1.0, "pause_ms": 45, "pause_tone_ms": 130, "emphasis_lengthen": 2}'
tts config --set 'rvc.delivery.en={"speed": 1.0, "pause_ms": 60, "pause_tone_ms": 160}'

"*" applies to any language not named. Spanish runs faster with shorter gaps than English, which is why this is per-language rather than one number — and why the previous behavior, a hardcoded 350 ms between every fragment, read as a stall rather than a breath.

The pause is padded onto the fragment itself rather than inserted while joining, so streamed playback and the saved file are the same sound.

trim_ms is what makes pause_ms mean anything. Every fragment arrives with its own lead-in and tail — the synthesizer's padding, plus whatever conversion adds at the edges. Joined, eight fragments carry eight lots of it, so the real gap would be that dead air plus your pause. Each fragment is trimmed to trim_ms of margin first, which puts the configured pause back in charge of the rhythm. On one mixed-language sentence this cut 7.24s to 5.79s without changing a word.

Fragments come from tone changes now, not from language: a borrowed word is handled by the pronunciation dictionary's IPA entries, inside the same utterance, so there is no edge to trim there at all. See Phonetics.

Audio playback

There is no audio library to install.

sudo apt install ffmpeg     # Debian/Ubuntu
brew install ffmpeg         # macOS

tts --player mpv "use this one instead"
tts config --set player=ffplay

If nothing is found, the file is kept and its path printed instead of vanishing.

On WSL, installing ffmpeg does not change your player. Windows' own player is the default on both Windows and WSL, because it is always present and it is the native way out of either — and WSL's Linux audio bridge is frequently the noisier of the two, so letting a freshly-installed ffplay take over would be a silent downgrade. Name a Linux player explicitly if you want one:

tts config --set player=ffplay      # or mpv, paplay, ...
tts config --set player=windows     # back to Windows' own player

Tuning a player for one machine

Audio stacks differ per box, and the fix is nearly always a flag or an environment variable rather than a code change — so both are configuration:

tts config --set 'player_args.ffplay=-af aresample=48000'   # inserted before the file
tts config --set player_env.SDL_AUDIODRIVER=pulseaudio      # set for the player only
tts config --set player_args.ffplay=                        # empty value removes it

player_args is keyed by player name and inserted just before the file argument; player_env is applied to the player process alone, so nothing leaks into your shell profile. tts check echoes both back when set.

ffmpeg is worth installing even if you already have a player. Tone tags change a span's pacing, and on any backend without its own rate control that retiming happens to the rendered audio: with ffmpeg through atempo, without it through a built-in WSOLA stretch that is listenable but measurably noisier. tts check prints a tone shaping: line naming which one you are getting.

Streaming playback

Each fragment plays as soon as it is synthesized, rather than the whole text being rendered and joined first — so the first words arrive in about a second instead of after however long the full text takes. The joined file is still written either way.

tts config --set stream=false    # render everything first, then play one file
tts --no-stream "..."            # same, for one run only

Fragment boundaries are the tone-tag and chunk boundaries that already existed; nothing is split that was not split before. Playback stays serialized machine-wide, and the runner holds the lock across the whole stream, so another session cannot cut in mid-sentence.

Troubleshooting

llama-tts: 'llama-tts' not found on PATH Install llama.cpp, or point at the binary: tts config --set llamacpp.binary=/full/path/to/llama-tts.

llamacpp.model is set but llamacpp.vocoder is not llama-tts needs two files: the TTS model and the WavTokenizer vocoder. Set both, or clear model (tts config --set llamacpp.model=) to fall back to the defaults.

llamacpp can only write .wav files Only the openai provider produces other formats. Render WAV, then convert: ffmpeg -i out.wav out.mp3.

"no audio player found" Install ffmpeg, or use --output and open the file yourself.

First run hangs for a long time It is downloading ~640 MB of weights. Run with --verbose to watch the progress.

The backend failed and I want to know why --verbose streams the backend's own stderr; --dry-run prints the exact command so you can run it by hand.

Development

python -m venv .venv
source .venv/bin/activate
pip install -e .

python -m unittest discover -s tests -v   # no test dependencies either

Layout:

src/localtts/
├── cli.py            argument parsing and the four subcommands
├── config.py         defaults, config file, env vars, precedence
├── text.py           markdown stripping and sentence-aware chunking
├── audio.py          playback autodetection and wav joining
├── skills.py         agent detection and skill installation
├── agent_skills/     the skill markdown shipped to agents
├── errors.py         TTSError -> a clean one-line message
└── providers/
    ├── base.py       Provider contract + subprocess helpers
    ├── kokoro.py     Kokoro-82M backend (the default)
    ├── llamacpp.py   llama.cpp backend
    ├── openai.py     OpenAI-compatible HTTP
    ├── piper.py      Piper ONNX voices
    ├── rvc.py        voice conversion over another provider
    └── command.py    user-defined template

Adding a provider: subclass Provider, implement synthesize(text, out_path, voice) and check(), register it in providers/__init__.py, and add its defaults to config.DEFAULTS["providers"]. A test asserts those two stay in sync.

See CONTRIBUTING.md before opening a pull request — in particular the no-runtime-dependencies rule, which is not negotiable.

Thanks

• @Belseck — phonetics in the pronunciation dictionary (#2). A /IPA/ value in pronunciations now travels to kokoro's server as phonemes, so a borrowed word keeps its own sound inside a sentence in another language — "ya subí el pull request" — said by the same voice, with no fragment boundary. It replaced the old <en>…</en> language spans, which cut the line up and handed the pieces to a second voice. The server is asked whether it understands phonemes rather than assumed to, so an older copy of the script is never silently sent a table it would drop.

License

MIT

Repeatable language calibration

Run tts calibrate to save two speech samples for every remembered language and a report.json with render time, first-fragment readiness, duration, edge silence, levels, and hard clipping. It never plays audio or changes your preferences. Models may start automatically, just as when speaking. The second run helps separate initial loading costs from repeated requests; the first run may already have a warm server.

tts calibrate --provider kokoro --provider rvc
tts calibrate --lang es --text 'Revisé el pull request y el despliegue está listo.'

Use --lang and --provider repeatedly to select comparisons, --runs for repeat count, and --output for the parent directory of a new results folder. Languages without a built-in sample require --text. The measurements require PCM16 WAV; unsupported formats or synthesis failures are recorded as errors with a nonzero exit. First-fragment readiness excludes player startup and queue time. These statistics cannot establish pronunciation accuracy or naturalness: compare two recordings by ear before choosing a voice or dictionary entry. Keep the sentence and all other settings fixed during each comparison.

Regional pronunciation entries override base-language entries regardless of their order in the config file. Language codes are case-insensitive and accept underscores: en_US, EN-us, and en-US select the same pronunciation and provider voice mappings.

Streaming starts the first available fragment immediately. Subsequent contiguous fragments already available in the same WAV format are played together, avoiding a player launch between each short span. Batching adds no silence and never waits for another fragment to finish synthesizing.

Codex skills use the native user skill directory documented by OpenAI. Reinstalling all five skills migrates the old local-tts block out of ~/.codex/AGENTS.md after the native files have been written. Unrelated instructions stay intact and the old file is backed up as AGENTS.md.local-tts.bak. --dry-run previews this without changing files.

Pronunciation trials use the requested language scope even when the dictionary already contains a regional respelling. Sentence matching requires a whole word, and the Kokoro server supports IPA for isolated words as well as sentences. Overlapping entries prefer the longest phrase. After updating, use tts servers --refresh to apply server fixes to an existing installation.

Vowel and consonant regression samples

tts calibrate --suite sounds --runs 1 renders connected speech covering Spanish R/RR, vowels, stops, fricatives, nasals, clusters, stress and questions, and analogous English contrasts. It uses the remembered profiles, including regional English, and never changes settings or plays audio. Other languages still use the basic sample or a custom --text; an unsupported sound suite fails explicitly.

RVC can be tuned per language with tts config --set 'rvc.conversion.es={"index_rate":0.65,"protect":0.2}'. This example is a trial setting, not a universal quality recommendation. Updated servers accept these parameters per request and restore their startup values afterward. Refresh an older server with tts servers --refresh; unsupported tuning is reported instead of silently ignored. Scoped pitch zero explicitly disables a startup pitch shift.

For optional unattended intelligibility checks, tools/evaluate_speech.py reads a calibration report and uses faster-whisper in a separate environment. It does not change runtime dependencies, upload recordings, play audio, or alter preferences. Word error rate can reveal lost words, but does not establish clear trills, accent, voice similarity, or naturalness. tools/run_sound_sweep.py renders controlled local comparisons from identical source waveforms.

For an ending that feels cut short, tts config --set ending_silence_ms=350 adds 350 ms of silence after the final WAV fragment. It also pads saved WAV output, but does not add gaps between fragments or delay the first streamed fragment. The default is 0 (disabled); accepted values are 0–5000 ms. Compressed output is unchanged. This gives playback a longer ending; it cannot restore speech already missing from synthesized audio.

Faster startup, audio caching and settings UI

tts settings                                  # terminal editor; changes save immediately
tts config --set cache_enabled=true            # existing CLI remains supported
tts config --set cache_ttl_hours=36 --set cache_max_mb=256
tts config --set cache_policy=lfu               # or lru
tts cache status                               # also prune / clear
tts warm --lang en --lang es                    # load models, prefill sample phrases
tts warm --lang en --text "The build is ready." --keep-alive 1800
tts --lang en --no-cache "Render this fresh."

The disk cache skips synthesis for matching requests. It keys on text, language, voice/model, local asset size/mtime, output format, synthesis settings, pronunciation entries, ending padding, platform and implementation. Dynamic phonetics hooks bypass it. Use cache_revision or tts cache clear when a remote model changes at the same URL. Files have a fixed lifetime (36 hours by default), not a sliding expiration. The 256 MiB budget includes entry headers and audio; least-used entries go first, with oldest access breaking ties. Entries larger than the limit are not cached. Pruning runs on cache activity or tts cache prune; an idle directory may retain expired files until the next operation, but they are never served after expiration. The limit covers this cache, not explicit exports, retained background recordings, model weights, filesystem allocation overhead, or model RAM/VRAM. Transfers are buffered; the cache is not loaded into memory as a whole.

Model warm-up reduces the cold-start cost for new phrases. Cache hits do not need resident models. --keep-alive runs visibly for a bounded period, never plays audio, and exits with Ctrl-C; it does not install a hidden daemon. Holding models warm uses RAM/VRAM. PowerShell process startup and queued playback still contribute latency.

The terminal editor supports arrows/j/k, Enter to edit, / to filter, a to add a CLI-style assignment, and q to quit. All top-level settings, provider settings and language mappings are exposed. Lists/maps use JSON. API keys are masked. It reloads external edits automatically; the CLI reloads configuration on every request. Saves are atomic and serialized across processes. Active utterances keep their snapshot; new requests use new voices/settings. A running warm session reloads its plan. Backend startup-only settings (loaded models, port/device/start command) apply when that server next starts; editing them never kills an active utterance.

See the complete settings reference, agent support, and v2.0.0 migration/release notes.

Resident local servers report the model assets and startup arguments they actually loaded. If those disagree with current settings/files, caching is bypassed until the server has restarted; old-model audio is not stored under the new identity. Opaque custom-server assets need explicit configuration or caching stays disabled.