Piper TTS — local spoken replies

What it is

Piper turns short spoken lines into audio without sending reply text to a cloud TTS vendor. The primary path is a one-shot spawn of piper.exe for each speak request. That is not a long-running Piper process: there is no Start / Stop button, and VibeFuse does not keep a Piper server running in the background.

If local piper.exe (or the selected ONNX voice) is unavailable, VibeFuse falls back to an OpenAI-compatible HTTP API: POST /v1/audio/speech on port 4123. The factory-stored Piper endpoint (ttsBase) is http://zrowton.ddns.net:4123. The field placeholder in Settings is http://127.0.0.1:4123.

Primary

Spawn piper.exe per speak

Fallback

POST /v1/audio/speech on :4123

Stored default

http://zrowton.ddns.net:4123

UI placeholder

http://127.0.0.1:4123

How a reply is spoken

1. Jarvis (widget id com.vibefuse.custom-agent) or the voice dock finishes a turn with on-screen text and a short spoken speech line.
2. If Use local Piper is on, VibeFuse runs piper.exe against the ONNX file for the selected Voice and plays the WAV through the output device in Settings → Voice.
3. If local spawn cannot run, the same line is posted to the Piper endpoint as /v1/audio/speech.
4. If Use local Piper is off, VibeFuse never spawns piper.exe and never calls a local Piper API — speech uses the remote endpoint only.

Settings → Local Servers · Piper TTS

Open Settings (Assistants group) → Local Servers. The first section is Piper TTS, with this copy:

Local text-to-speech server for Jarvis replies. Voice, speed, and volume are in Settings → Voice.

Voice, speed, and volume are not on this tab. They are under Settings → Voice → Spoken output. MCP connections are on Settings → Extensions, not Local Servers.

Settings → Voice · Spoken output

Section title Spoken output, description: “Piper voice, speaking rate, and loudness for Jarvis replies.”

Output device hint: “Piper voice replies, wake-word audio, video widgets, and browser webviews.” Test Piper is not on this tab — only on Local Servers. There is no Test voice button here. Whisper URLs: “Piper and Whisper server URLs live in Settings → Local Servers.”

Launch greeting

Section Launch greeting: “A short bundled Jarvis greeting played once when VibeFuse finishes starting up.”

• Toggle Greet me on launch (default on).
• Plays Good morning (5:00–11:59), Good afternoon (12:00–16:59), or Good evening by local time.
• Uses the output device and volume from Spoken output.
• These are bundled WAVs, not a live Piper synthesize. If the browser blocks autoplay, the clip plays on your first click or key press instead.

Status bar chips

The workspace status bar shows a Voice chip (selected TTS voice name) and a Piper chip (Ready / Down). Both open Settings → Local Servers. The Speech chip next to Piper is Whisper health — Test Piper reports that same pair as Speech … · Piper ….

The Mode chip opens Settings → Voice. That is the voice dock, not Piper.

Bundled acknowledgements (not Piper)

Two short Jarvis clips are shipped as WAV files and played instantly — no piper.exe spawn and no /v1/audio/speech round-trip:

If those play and later replies are silent, the problem is live Piper (binary, voice file, toggle, or endpoint) — not the acknowledgement path.

Mid-task speech and the finished JSON

While Jarvis is working, it can call the play_tts tool (aliases speak_tts, say, speak) for a short first-person update. That is live Piper. The finished on-screen answer and the final spoken line arrive as JSON:

• text is what you read on the canvas. speech is what Piper speaks for the completed result.
• play_tts is for mid-task status only — not to announce that a tool is about to run.
• Spoken lines are short, first person, and start with I am / I have / I will / I did.

Mute Piper on the Jarvis widget

The canvas chat widget is Jarvis (id com.vibefuse.custom-agent). The speaker button in its top bar is Mute Piper voice replies / Unmute Piper voice replies. Mute stops live Piper for that chat; it does not turn off Use local Piper for the rest of the app, and it does not remove the bundled wake / tool WAVs from the voice dock.

Paths and the installer

The Windows installer has a step named Piper voice. It checks for the Jarvis voice and local speech engines. If the download is skipped, VibeFuse can still fetch jarvis.onnx on the next speak attempt (Jarvis only).

Which voices download

Only Jarvis auto-downloads. Lewis, Michael, Adam, Heart, Bella, and Emma need their ONNX files already on disk under %LOCALAPPDATA%\VibeFuse\models (for example bm_lewis.onnx, am_michael.onnx, am_adam.onnx, af_heart.onnx, af_bella.onnx, bf_emma.onnx). Selecting a voice that is not installed will not fetch it.

HTTP fallback

Used when local spawn cannot produce audio and Use local Piper is still on (local API on :4123), or when the toggle is off and the stored Piper endpoint is called instead.

Port 4123 is the local/API fallback port. The factory URL points at http://zrowton.ddns.net:4123; a machine that hosts its own API typically uses the placeholder http://127.0.0.1:4123.

Troubleshooting

FAQ

Related