command-center/omp-docs/core/dankvoice.md

---
type: reference
title: dankvoice
description: omp extension that speaks assistant output with the fleet's cloned voices through frank-tts.
tags:
  - omp
  - extension
  - tts
  - voice
  - frank-tts
  - dankclaw
verified: 2026-10-19
timestamp: 2026-10-19
---

# dankvoice

`agent/extensions/dankvoice.ts` adds cloned-voice speech to the interactive
harness: `/say` opens the options menu, `/say <text>` speaks through the
speakers and `/say on` narrates the assistant's thinking and final answer
after every turn. The words are first rewritten into natural spoken prose by
the **@tiny** role model (exactly what the built-in vocalizer's
`speech.enhanced` does), and everything is served by **frank-tts**
(Dankclaw :7772, the `speech` model role in [agent/models.yml](/core/index.md)).

## Why it exists

Nothing built into the harness can pronounce Frank's cloned voices.

| Surface | Engine | Verdict |
|---|---|---|
| TUI speak-aloud (`speech.enabled`, `speech.voice`) | bundled Kokoro ONNX worker, `synthesizeStream(voice, speed, signal)` | Kokoro voice enum only; never consults `models.yml` or the speech role |
| `omp say [TEXT]` | same local backend (`--model` = *local* model key) | `omp say --voice narrator1 "test"` → `error: Expected --voice to be one of: af_heart, af_bella, af_nicole, af_kore, af_sarah, am_michael, am_fenrir, am_puck, bf_emma, bm_george, bm_fable` (verified 2026-10-18) |
| `tts` tool / `xd://tts` device (`modelRoles.speech`) | frank-tts over OpenAI `/v1/audio/speech` | cloned voices work, but only as a **file** — nothing plays them |

`dankvoice` is not a patch to that vocalizer (harness code, not ours); it is a
second transport that happens to speak, so Kokoro stays available for quick
local speech and the cloned voices carry the real narration.

## Commands

| Command | Effect |
|---|---|
| `/say` \| `/say menu` | options menu (`ctx.ui.select`): auto-speak toggle, voice, output device, player, spoken cap, thinking-narration toggle, humanize toggle, test phrase |
| `/say <text>` | synthesize once and play; interrupts whatever is playing |
| `/say on` \| `/say off` | arm/disarm auto-speak of the thinking blocks and final text at `agent_end`; the footer shows `🔊 <voice>` while armed |
| `/say voice <name>` | pick a voice; an empty name defers to the server's own default |
| `/say voices` | list what frank-tts reports |
| `/say stop` | cut the current utterance |
| `/say status` \| `/say help` | endpoint, voice, player, device, cap, thinking, humanize |

Menu choices persist to `~/.omp/cache/dankvoice.json` (runtime state, never
tracked); the environment overrides the file, the file overrides defaults.
"Speak a test phrase" synthesizes a short line through the current voice,
player and device — the audible proof whenever a session "has say on" but
stays silent.

Auto-speak is skipped in `print`/`json` mode so benches and pipelines stay
silent. Each spoken unit — thinking first when narrated, then the answer —
is first rewritten into spoken prose by the @tiny model; when the enhancer
(below) declines, the text goes to segmentation as-is, where fenced code and
tables vanish, links collapse to their label and bullets and table pipes
disappear. A long answer is never one monolithic synthesis request; see
[Segmentation](#segmentation) for how utterances are cut and played.

## Segmentation

Utterances are cut by a state machine ported from the harness's own
speak-aloud vocalizer (the minified `fHe` class in the agent bundle's
`dist/cli.js`), kept deliberately in lockstep with the stock path:

* A strong boundary is a run of `[.!?…]` followed by whitespace — so
  `!!!!!` and `...` are one boundary, never a pile of cuts.
* Boundaries after `e.g.`, `etc.`, `Mr.` and friends, or inside inline
  code, don't count; decimals (`3.5`, `1,5`) survive because a boundary
  requires trailing whitespace.
* Line starts are classified: table rows and fenced code are swallowed
  whole (their content is never read aloud), heading and bullet markers
  vanish, ordered-list markers speak their number ("1, …"), quote markers
  strip.
* Per segment, markdown is cleaned: links collapse to their label, URLs
  to their host, deep paths to their last segment; emphasis markup dies.
* Segments aim for one sentence; run-ons break at commas or word wraps
  before 280 characters.

Playback is strictly sequential: segment N+1 is synthesized only after
segment N's player process exited, so audio never stacks. The chain runs
detached from `agent_end` (fire-and-forget) and may legitimately outlive
the harness's 30 s handler timeout — the old awaiting design produced a
"handler timed out after 30000ms" toast on long answers. A new utterance,
`/say stop`, or disarming auto-speak retires the live chain through a run
token: the current player is killed and pending segments are skipped.
`DANKVOICE_MAX` is a per-answer character budget: the leading whole
segments that fit are spoken, the rest are dropped — the thinking unit
draws on it before the answer does.

## Endpoint resolution

The base URL is **read from `agent/models.yml`**, from the provider block whose
model row carries `api: openai-speech` — the same row the `speech` role
resolves, so no host address and no credential appears in this file. From
`baseUrl: http://192.168.0.143:7772/v1` it derives
`POST …/v1/audio/speech`, `GET …/v1/audio/voices` and `GET …/health`.

| Environment | Default | Meaning |
|---|---|---|
| `DANKVOICE_URL` | models.yml | speech base URL (`…/v1`), bypasses the YAML read |
| `DANKVOICE_MODEL` | the id in that block (`frank-tts`) | model named in the request body |
| `DANKVOICE_CONFIG` | `~/.omp/cache/dankvoice.json` | persisted menu choices (voice, player, sink, cap, auto) |
| `DANKVOICE_VOICE` | empty (server default) | default voice name |
| `DANKVOICE_AUTO=1` | off | start with auto-speak armed |
| `DANKVOICE_MAX` | `700` | spoken-character budget per answer (whole leading segments, shared by the thinking and answer units) |
| `DANKVOICE_PLAYER` | detect | player binary, skipping detection |
| `DANKVOICE_MODELS` | `~/.omp/agent/models.yml` | YAML to read the endpoint from |
| `DANKVOICE_REWRITE_URL` | the @tiny role | chat base URL (`…/v1`) for the enhancer, bypassing role resolution |
| `DANKVOICE_REWRITE_MODEL` | the @tiny role's model | chat model named in the rewrite body |
| `DANKVOICE_REWRITE=0` | enhancer on when a route resolves | speak mechanical text |
| `DANKVOICE_THINKING` | settings file, default on | `0`/`1` force thinking narration off/on |
| `DANKVOICE_OMPCONFIG` | `~/.omp/agent/config.yml` | YAML to read the `modelRoles.tiny` role from |

Player detection order is `pw-play` → `paplay` → `ffplay` (PipeWire first:
the workstations run WirePlumber). A saved output device routes the player
explicitly — `pw-play --target <sink>` or `paplay --device=<sink>`; ffplay
has no device concept and ignores it — while an unset device leaves the
player's own default. Sinks are enumerated with `wpctl list audio sinks`;
if wireplumber is unreachable the menu offers only "system default". Each
utterance is written to `$TMPDIR/dankvoice-*/say-<ms>.wav` and the file plus
its directory are removed when the player process exits.

## Humanized narration (the @tiny role)

The built-in vocalizer never reads answers aloud verbatim when
`speech.enhanced` is on: a small model rewrites each utterance into words a
person would say. `dankvoice` ports that enhancer — same system prompt
(verbatim from the bundle), same budgets (1536 max tokens, 6 s timeout,
4000-char input window), same contract (a failed rewrite falls back to the
mechanically cleaned text; an empty rewrite skips the unit as "nothing to
say"). The target model is the harness's `@tiny` role, resolved the way the
harness resolves it: `modelRoles.tiny` in `agent/config.yml` reads
`provider (host)/model:effort`; the provider block in `agent/models.yml`
supplies the `baseUrl`; the request is `POST {base}/chat/completions`. On
this fleet that is **frank-zealot** (`llamacpp (dankclaw-3090)`,
`:7771`) — the 6 s abort keeps a slow tiny model off the critical path.

Thinking blocks ride the same path: at `agent_end` the last assistant
message's true `thinking` parts are narrated (rewritten first, as their own
unit, before the answer), while blocks flagged `narration: true` — which
providers use to address the user through the thinking channel — count as
answer text. The menu, `/say status`, and `DANKVOICE_THINKING` toggle it;
the choice persists in the settings file.

## Verified facts (2026-10-18, extended 2026-10-19)

* **Extension discovery** — any `.ts` file in `~/.omp/agent/extensions/` loads
  at startup: a probe file there printed `loaded 2026-10-19T02:36:55Z` from a
  fresh `omp -p` run. `omp --help` documents `--no-extensions` (discovery off,
  explicit `-e/--extension` and `--hook` still load) and `plugin install|link`.
* **Endpoint** — resolving the live `agent/models.yml` yields
  `speechUrl http://192.168.0.143:7772/v1/audio/speech`,
  `voicesUrl …/v1/audio/voices`, `healthUrl …/health`, `model frank-tts`.
* **Roster** — `GET /v1/audio/voices` answers the bare JSON array
  `["narrator1","narrator2","narrator3"]`; `/health` answers
  `{"status":"ok","backend":"omni-http","voices":[…]}`. `GET /v1/voices` and
  `GET /v1/models` are **404** on this shim, and an unknown voice answers
  `400 unknown voice 'zzz'; available voices: …` with `available_voices` in the
  body — which is the fallback the extension uses if the roster route ever
  disappears.
* **Omitting `voice` is legal** — frank-tts fills its `DEFAULT_VOICE` since the
  2026-10-18 voice-pack work, so an unset voice synthesizes instead of failing.
* **Speech** — `/say "dankvoice live smoke test"` produced a 99,884 B / 2.08 s
  WAV and `/say voice narrator2` + `/say "cloned voice check"` a 69,164 B /
  1.44 s WAV (durations from `ffprobe`), each handed to the player; temp
  directories were gone afterwards.
* **Auto-speak** — verified in a real TUI session: `/say on` armed the footer
  (`🔊 narrator1`), the next finished turn synthesized and played the answer
  (1.20 s WAV), and an unknown voice name answers with the roster instead of
  speaking. An `agent_end` carrying `willContinue` (the harness scheduled an
  auto-retry) is never narrated — only the terminal settle speaks.
* **Menu** — verified via driver scenarios 2026-10-18 (Refs #23): the menu
  lists auto-speak/voice/device/player/cap/test-phrase; a voice chosen in the
  menu is used by later utterances and survives a fresh load via the settings
  file; a device chosen in the menu appears in the player invocation as
  `--target <sink>`.
* **Segmented playback** — driver scenarios 2026-10-19: "First stop here.
  Second one goes!!! Really? Third ... done" produced exactly 3 sequential
  requests and 3 plays; "Model 3.5 ships today. Version 1,5 too." produced
  exactly 2 (decimals never split); the markdown fixture's fenced code and
  URL never reached the fixture endpoint; with `DANKVOICE_MAX=60` only the
  leading segment was spoken.
* **Humanize + thinking (2026-10-19, Refs #27)** — driver scenarios cover the
  rewrite round-trip (fixture chat route), the 404 → mechanical fallback,
  thinking narrated as its own unit before the answer, `DANKVOICE_THINKING=0`
  silencing it, and `narration: true` spoken once as the answer; the endpoint
  status line shows the resolved humanize model from the @tiny role. Live
  smoke against the real services: a markdown `/say` became 3 sequential
  prose segments (mechanical cleanup of that text yields at most 2), and a
  thinking+answer `agent_end` played thinking first. 22 green in
  `tests/dankvoice.bats`.

## Limits

* The TUI's own speak-aloud feature still narrates in Kokoro: its voice list is
  a compile-time enum in the harness bundle, and no setting routes it to
  frank-tts. Cloned voices live only on this file-producing/speaking path and
  on the `tts` tool. See [Handy](/tools/handy.md) for the STT half.
* `agent_end` narrates thinking blocks and the final message only — mid-turn
  tool chatter is never spoken. Thinking covers the last assistant message;
  tool-loop gaps in earlier messages are skipped by design.
* Humanized narration is serial: audio starts only after the @tiny rewrite
  returns (≤ 6 s per unit, then frank-tts synthesis). Say `/say off`-grade
  latency is the price of prose; `DANKVOICE_REWRITE=0` trades it back for
  instant mechanical speech.
* frank-tts is LAN-only: no Caddy vhost fronts :7772, so this extension is
  usable on the home network.
* Deployment, packs and the voice registry are owned by the
  `serverAppConfigs` repo (`serverAppConfigs-docs/Dankclaw/Frank TTS.md`), not
  by this one.

## Citations

* Live probes on the workstation, 2026-10-18: `omp say --voice narrator1 …`,
  `curl http://192.168.0.143:7772/health`, `/v1/audio/voices`,
  `/v1/audio/…/speech` (invalid voice), `ffprobe` on the produced WAVs.
* `omp --help` (extension discovery flags, `say`, `plugin`) and
  `@earendil-works/pi-coding-agent` `dist/core/extensions/types.d.ts`
  (`ExtensionAPI.on/registerCommand`, `ExtensionMode = "tui" | "rpc" | "json" | "print"`).
* The stock speak-aloud implementation — vocalizer pipeline `I6o`, utterance
  segmenter `fHe` and speech enhancer `Qyr` with its prompt `T6o` — in
  `@oh-my-pi/pi-coding-agent`'s `dist/cli.js`, read off the minified bundle
  2026-10-19: dankvoice ports the mechanical and the enhanced halves, so an
  LLM rewrite does happen before frank-tts.
* Gitea issue #21 on `dank/command-center` (the work item this page ships with).
* Gitea issue #23 on `dank/command-center` (the options-menu work item).
* Gitea issue #27 on `dank/command-center` (the enhanced-narration work item).

## See Also

* [Core Documentation](/core/index.md) — the live harness configuration.
* [Handy](/tools/handy.md) — offline speech-to-text, the input half of voice.
* `serverAppConfigs-docs/Dankclaw/Frank TTS.md` — the frank-tts service runbook.