config.example.toml

# voxcaster configuration
#
# Copy to ~/.config/voxcaster/config.toml (or point $SPEAKD_CONFIG at it).
# Every key is optional — the built-in defaults work out of the box on CPU.
# Environment variables (SPEAKD_*) override anything set here.

[tts]
# Kokoro voice id. A few favourites:
#   af_heart, af_bella, af_nicole  (American female)
#   am_adam, am_michael            (American male)
#   bf_emma, bf_isabella           (British female)
#   bm_george                      (British male)
voice = "af_heart"

# Speech-rate multiplier (1.0 = normal).
speed = 1.0

# Kokoro language code: "a" = American English, "b" = British English, ...
lang_code = "a"

# Strip markdown (code fences, URLs, headers, bold/italic, table pipes, HTML)
# and extract an embedded <!-- TTS_SUMMARY ... TTS_SUMMARY --> before synthesis.
# Makes raw LLM/CLI output safe to pipe straight into `speak`. Default: true.
markdown_preprocess = true

[device]
# "auto" — dynamic offload (recommended): load on CPU, hop to the GPU for
#          synthesis bursts, release VRAM after `keepalive_seconds` idle.
# "cpu"  — never touch the GPU.
# "gpu"  — move to the GPU on first use and stay there.
policy = "auto"

# Idle seconds before the model is offloaded from GPU back to CPU.
keepalive_seconds = 180

[daemon]
# Unix socket the daemon listens on.
# Default: $XDG_RUNTIME_DIR/voxcaster.sock (falls back to /tmp/voxcaster-<uid>.sock).
#socket_path = "/run/user/1000/voxcaster.sock"

# Octal permissions applied to the socket file.
socket_mode = "600"

# Treat every request as an interrupt: each new line drains the queue and cuts
# off in-flight playback, so only the latest narration is ever spoken. Handy
# for Claude Code hooks where stale lines aren't worth hearing. Default: false.
always_interrupt = false

# Where an auto-spawned daemon writes its log.
# Default: ~/.local/state/voxcaster/daemon.log
#log_file = "/home/me/.local/state/voxcaster/daemon.log"

[audio]
# Playback volume 0-130 (mpv scale; 100 = nominal, >100 = software gain).
volume = 100

# Kill playback of a single line after this many seconds (safety net).
max_playback_seconds = 120

# Audio player argv template. "{file}" is the wav path, "{volume}" the
# current volume. Swap in any player you like, e.g. ffplay:
#   player = ["ffplay", "-nodisp", "-autoexit", "-volume", "{volume}", "{file}"]
player = ["mpv", "--no-terminal", "--volume={volume}", "{file}"]

[fallback]
# Spoken when Kokoro or audio playback fails (and by the client when the
# daemon cannot be reached). "{text}" is the line to speak.
# Set to [] to disable the fallback entirely.
command = ["espeak", "-s", "160", "-v", "en-us", "{text}"]

[client]
# Socket connect/send timeout (seconds). Keep small: callers should never block.
connect_timeout = 0.5

# How long `speak --blocking` waits for the spoken-ack.
ack_timeout = 300.0

# How long the client waits for an auto-spawned daemon to come up.
spawn_wait = 4.0