Voice nodes (openpanel V4): stt, tts, wren reply/unhandled, socket publish

- `stt`: one realtime transcription session per utterance on the speech
  server's url, opened at `listening` true or the first chunk, committed at
  false; a whole clip goes to /v1/audio/transcriptions. Returns at once and
  publishes `partial` and `text` when the server answers.
- `tts`: one /v1/audio/speech call per sentence, cached by voice and text in
  node state; a reference is marked `append` while earlier ones still play.
- wren: `reply` says what was applied, `unhandled` carries the prompt when
  nothing was. The install hint names the git pin (PyPI's `wren` is another
  package).
- ai/client: transcribe, speak and transcribe_stream beside stream_completion.
- A device socket publishes scalar `in` ports as {"type": "publish"} text,
  in order with its binary frames.
- Pipeline.publish walks only what its values reach: a node downstream of a
  consumer that published nothing no longer runs again (wren re-acted on
  every microphone chunk).

docs/code/api.md

@@ -286,8 +286,12 @@ that family, and the whole message at most `MAX_FRAME_BYTES`. The frame is
 held in memory and published in one step, named after the message and credited
 to its device node. A frame that does not
 qualify is dropped, the reason logged once per socket, and the socket stays
-open. With `http` it uploads with `PUT /artifacts?volatile=true&name=<message>`
-and publishes with `POST /messages/<message>`. A device that is unpaired or
+open. A value of one of its `in` ports wired as `bool`, `int`, `float` or `str`
+goes as the text message `{"type": "publish", "name": "<message>", "value":
+...}`, taken in turn with the frames, so a microphone's `listening` false
+arrives after its last chunk. With `http` it uploads with
+`PUT /artifacts?volatile=true&name=<message>` and publishes with
+`POST /messages/<message>`. A device that is unpaired or
 deleted while connected is closed with 1008, at the latest with its next frame.
 
 ### Live view

docs/code/nodes.md

@@ -179,7 +179,10 @@ and ARM, so nothing has to be installed into the image.
 
 openpanel publishes a 16 kHz mono WAV on its `mic` port after its wake word,
 and plays any `audio` reference sent to its `speaker` port (WAV, MP3 or AAC).
-Two Function nodes close that loop with local models.
+The [Speech to text](../reference/node-types.md#speech-to-text) and
+[Text to speech](../reference/node-types.md#text-to-speech) nodes close that
+loop through a speech server on a GPU. Without one, two Function nodes close it
+with local models.
 [faster-whisper](https://github.com/SYSTRAN/faster-whisper) turns speech into
 text and [piper](https://github.com/OHF-Voice/piper1-gpl) turns text into
 speech.

docs/interface/dashboards.md

@@ -514,8 +514,10 @@ to 640×360. Re-encode other clips with the [clip recipe](../code/nodes.md#makin
 The device decides what its [port](#the-devices-own-ports) names
 mean. openpanel [reports](#what-a-device-reports) the ones it serves, such as
 `gpio.<n>`, `backlight`, `touch`, `mic` and `speaker`, with the settings each
-one reads, and ignores any other name. A browser screen serves none of them. The [speech recipe](../code/nodes.md#speech-to-and-from-a-panel)
-turns `mic` into text and text into `speaker`.
+one reads, and ignores any other name. A browser screen serves none of them. The
+[Speech to text](../reference/node-types.md#speech-to-text) and
+[Text to speech](../reference/node-types.md#text-to-speech) nodes turn `mic`
+into text and text into `speaker`, and `caption` shows what is being heard.
 
 ### The device's own ports
 

docs/reference/node-types.md

@@ -344,7 +344,9 @@ use a [batch flow](../concepts/runs.md) and a Function node.
 
 [Wren](https://github.com/stroblme/wren) is an optional local language model
 for device commands. Install it in the engine's environment with
-`pip install wren`; the node reports a missing dependency if it is absent.
+`pip install git+https://github.com/stroblme/wren@eae4d6ff9ded68e36624972fd3f4a8f8d210b0c2`.
+The `wren` package on PyPI is an unrelated one. Without the model the node
+reports a missing dependency.
 
 | Input | Type | Notes |
 |---|---|---|
@@ -353,7 +355,8 @@ for device commands. Install it in the engine's environment with
 | `outdoor` | `float` or `int` | °C; 20 without it, clamped to −5…30 |
 | `dark` | `bool` | without it, dark before 07:00 and from 18:00 |
 
-The outputs are `actions`, a list of records, and any number of **targets**. A
+The outputs are `actions`, a list of records, `reply`, `unhandled`, and any
+number of **targets**. A
 target is an output named `<area>_<kind>` and bound to the message its device
 listens on: `living_room_setpoint` bound to `home.setpoint` is the living
 room's heating. The wiring is the allowlist, so the model can write nothing
@@ -393,6 +396,10 @@ Each record in `actions` contains `message`, `value`, `applied`, `delay_s`,
 `call`, and `confidence`. Unwired calls have an empty `message` and are not
 applied. Keep `apply=false` while checking how the model interprets commands.
 
+`reply` is one sentence saying what was applied, such as "Kitchen light on."
+or "Garage light on in 5 minutes.". When nothing was applied, `unhandled`
+carries the prompt instead, for something else to answer.
+
 A call with a delay ("in 10m") is held by the engine's work queue, like the
 Debounce node's second value: it survives a restart where the queue is in
 Redis, and not otherwise.
@@ -408,6 +415,57 @@ They are reported but cannot receive writes unless wired.
 
 The chat's [Control tab](../code/agents.md#control) sends commands by hand.
 
+### Speech to text
+
+**`stt`** — turn a microphone's audio into text, through a speech server.
+
+| Input | Type | Notes |
+|---|---|---|
+| `audio` | `audio` | 16 kHz mono PCM16 chunks (`audio/L16`), or a whole clip such as a WAV |
+| `listening` | `bool` | true while the device listens; false ends the utterance |
+| `url` | `str` | the speech server, from a [Service](#service) node; read, never waited for |
+
+| Output | Type | Notes |
+|---|---|---|
+| `partial` | `str` | the words agreed on so far, as they come, then the final text |
+| `text` | `str` | what was said, once per utterance |
+
+The server is any that speaks OpenAI's audio API, such as `parrot serve`. Each
+utterance is one realtime transcription session (`{url}/v1/realtime`), opened
+when `listening` turns true or at the first chunk and committed when it turns
+false. A clip is sent whole to `{url}/v1/audio/transcriptions`. The node
+returns at once and publishes when the server answers, since a device's next
+chunk waits for it. An utterance that fails, or gets no chunk for 10 seconds,
+publishes nothing and is logged.
+
+`listening` has to hold a value before a chunk wakes the node. Declare it as a
+flow input with the initial value `false` when a device may never send it,
+such as one that sends whole clips.
+
+### Text to speech
+
+**`tts`** — speak text as audio for a speaker, through a speech server.
+
+| Input | Type | Notes |
+|---|---|---|
+| `text` | `str` | what to say, a whole reply or one sentence at a time |
+| `url` | `str` | the speech server, as for Speech to text |
+
+| Output | Type | Notes |
+|---|---|---|
+| `speech` | `audio` | one WAV reference per sentence |
+
+| Setting | Default | Notes |
+|---|---|---|
+| `voice` | `af_heart` | one of Kokoro's voices, such as `bm_george` |
+
+Each sentence is one call to `{url}/v1/audio/speech` and published as soon as
+it is spoken. A sentence spoken before in the same voice is reused rather than
+spoken again, for the last 256 sentences. A reference carries `"append": true`
+while the sentences before it are still playing, which queues it on a panel's
+`speaker`. The first sentence of a reply carries none and cuts off whatever
+was playing.
+
 ## Devices
 
 ### Device