---
title: "Phonon voice hardpoint"
description: "xn-ptts as the tts hardpoint in your own trained Phonon voice file: where the voice file, its pace and the loadout's tts_voice each live, the per-voice pace patch and its build, the launch line (-0.04 for the operator voice), and what the Admin's Hardpoints card shows."
type: "living"
source: "catalog/hardpoints/xn-ptts/README.md"
verified: "2026-11-14"
group: "Adapters and integrations"
---

> **Living doc**, checked against the code 2026-11-14

`catalog/hardpoints/xn-ptts/hardpoint.toml` binds xn-ptts's own Wyoming server (`ptts-openai-server`, MIT OR
Apache-2.0, github.com/gradium-ai/xn-ptts) into the gateway's `tts` slot with no shim. You build and run the server;
Vulpine does not install it. This page says how to serve a trained Phonon voice file (a 125 x 512 `speaker_wavs`
prompt) at its own pace, and how a loadout picks it.

## Where each piece lives

| piece | where | why there |
|---|---|---|
| the voice file | your private folder, beside nothing else: `<voices>/<name>.safetensors` | it is derived from someone's recordings; it never enters git, a loadout, the trace or the Admin |
| its pace (`duration_delta`) | the server's launch line, `--voice-condition <name>:duration_delta=<x>` | pace belongs to the voice and the runtime; the patch below makes it per voice |
| which voice the gateway asks for | your loadout, `[hardpoints] tts_voice = "<name>"` | a user's choice, not the shared, digest-pinned Catalog entry |
| on the wire | Wyoming's own `synthesize {text, voice: {name}}` (and `synthesize-start {voice}` when streaming) | the standard field; any Wyoming server that knows the name takes it |

The server loads every `*.safetensors` in `--voice-dir` as a voice named by its file stem, so
`<voices>/operator.safetensors` is the voice `operator`.

## Build (on the machine that serves)

1. Unpack xn-ptts at upstream commit `6f42148066dfe3640ab2099d5c2f088c826937d6`. From the tree's parent folder, apply
   the patch: `patch -p1 -d xn-ptts-main -i catalog/hardpoints/xn-ptts/patches/xn-ptts-voice-conditions.patch`.
   The patch adds `SynthApi::set_voice_conditions` to `ptts` and the server flag `--voice-condition VOICE:NAME=VALUE`.
   The flag repeats; `PTTS_VOICE_CONDITION` takes a comma list. A condition for a voice the server does not have stops
   the server at start.
2. `cargo build --release --locked -p ptts-openai-server` inside the tree. It needs Rust 1.98 or later. The tree's
   `.cargo/config.toml` adds `-C target-cpu=native`, so build on the machine that serves.
3. `cargo test --release --locked -p ptts-openai-server` runs the patch's parser tests with the server's own.

`python3 tools/speech/xn_ptts_voice.py --src <tree> build --patch <patch>` does all three.

## Run

```text
ptts-openai-server --addr 127.0.0.1:8880 --wyoming-addr 127.0.0.1:10220 --lang en --voice-catalog '' \
  --voice-dir <voices> --voice-condition operator:duration_delta=<x>
```

`<x>` is the voice's measured pace setting. For the operator voice file it is **-0.04**. Each voice file needs its
own: `tools/speech/xn_ptts_voice.py grid` renders a grid, and `python -m voicetune phonon pace-cal` from
`pipelines/voice-tune` fits the setting from it. Pace moves with a line's words about as much as with the knob, so fit
on 30 lines or more.


## Pick it in a loadout

```toml
[hardpoints]
tts = "xn-ptts"
tts_voice = "operator"
```

The gateway qualifies the binding by rendering `Hardpoint check.` in that voice. It does not play it. A server without
the voice answers `unknown voice 'operator'`. The bind then falls back to `qwen3-worker`, and the Admin's Hardpoints
card shows the server's words. The loadout refuses `tts_voice` beside a built-in, or a name outside 1 to 64 of
`[A-Za-z0-9_.-]` (`voice_invalid`). Check it first with `just catalog::hardpoint-check tts xn-ptts --voice operator`.

## What the Admin shows

While xn-ptts carries the slot, the Hardpoints card's `tts` row reads `xn-ptts (yours) · voice operator`, with its
qualification ms against the leg budget. When the fallback carries it, the row leaves the voice off: the built-in
speaks in the persona's voice. The voice check after the bind reports the named voice as `configured`.
