Configuring for speech-dispatcher — $VOICE not set

[ysalmon]

I am trying to use piper with speech-dispatcher (and Firefox). I installed it with pipx install piper-tts and I can invoke it.

I followed this old post and adapted it (had to add --raw to pw-play to match --output-raw in piper).

My /home/yann/.config/speech-dispatcher/modules/piper.conf thus has

GenericExecuteSynth "export XDATA=\'$DATA\'; echo \"$XDATA\" | sed -z 's/\\n/ /g' | piper --cuda --model /home/yann/.local/share/piper/${VOICE} --output_raw | pw-play --raw --rate 22050 --channel-map LE - "
AddVoice "en-GB" "male1" "en_GB-alan-medium.onnx"
DefaultVoice "en_GB-alan-medium.onnx"

However, when using spd-say Hello, it appears that the $VOICE variable is not set during speech-dispatcher invocation of piper, which therefore fails.

I added a export VOICE=${VOICE:-en_GB-alan-medium.onnx} ; at the beginning to make it work, but I remain counfounded by the need to do so (Arch's doc also uses this variable).

This is probably more a speech-dispatcher thing than piper, but maybe other users here have been having the same problem ?

[MLuc24]

The confusion makes sense once you know that $VOICE isn't a shell variable at all — it's a placeholder that speech-dispatcher substitutes into the command string before handing it to the shell. Its generic module does a plain literal text replacement over the GenericExecuteSynth line, one token at a time:

e_string = string_replace(e_string, "$LANGUAGE", ...);
e_string = string_replace(e_string, "$PUNCT", generic_msg_punct_str);
if (generic_msg_voice_str != NULL)
    e_string = string_replace(e_string, "$VOICE", generic_msg_voice_str);
else {
    char *default_voice = module_getdefaultvoice();
    if (!default_voice)
        default_voice = "no_voice";
    e_string = string_replace(e_string, "$VOICE", default_voice);
}

Note the token it searches for is exactly $VOICE. Your config writes ${VOICE}, and that string doesn't contain $VOICE as a substring — the brace breaks it. So the replacement never fires, the text passes through untouched, and by the time the shell sees ${VOICE} it's just an unset shell variable expanding to nothing. Hence the empty path and the failure.

Dropping the braces should be all it takes:

GenericExecuteSynth "... | piper --cuda --model /home/yann/.local/share/piper/$VOICE --output_raw | ..."

Your export VOICE=... workaround works for a completely different reason — it creates a real shell variable so that ${VOICE} finally expands to something. That's why it feels like it shouldn't be necessary: you were fixing the shell's view of the world when the problem was that speech-dispatcher had already declined to fill the slot. It also means your workaround pins the voice, since the shell value wins regardless of what was requested, which is probably not what you want long term.

The Arch wiki page you found writes it without braces for exactly this reason. And if it ever does come through empty after the fix, that branch above tells you what to expect — a missing voice becomes the literal string no_voice rather than nothing, so a path ending in /no_voice in the error is speech-dispatcher telling you the AddVoice mapping didn't match, not that substitution failed.