The DJ's leftovers: outro cues, what aired, and voices from the console

[robert-dean]

The DJ that actually says something

As of: 2026-08-08, when the station first played audio it was not given by a music provider.
Cut down: 2026-10-10, to what is still open. The sections describing the two pieces as designed
and as shipped were removed; the file as it stood before is in this discussion's edit history.

Status (2026-10-10)

Both pieces this discussion was written for are built, and so is most of what it left behind:

• Piece one, the station speaks in its own voice (2026-08-08), and piece two, something decides
  what to say (2026-08-12): BreakWriterRegistry keyed by segments.kind, the model writer with
  the deterministic TalkBreakWriter as its floor, and the forward-claim check
  (segments.claims_item_id).
• The talk-up limit (Breaks: talk up to the post, from where the singing starts #412): talking up to the post, from where the singing starts. Vocal ranges
  come from synced lyrics (lyrics/vocal.ranges.ts, lyrics/vocal.runway.ts, LRCLIB as the first
  source), an operator's correction wins (deadair.track_vocal_overrides, migration 0064), the
  writer is told the runway and the words that fit (director/talk.up.ts, breaks.talkUp, off by
  default), and the link becomes a talk-over on the order. The first sung word replaced the "first
  sustained vocal" this file asked source separation for.
• A measured segment duration (Render: write down how long a spoken segment decoded to #434): the analyzer's post-render measurement records the decoded
  length into segments.duration_ms, which was null on 160 of 161 rows.
• Break ratings: 0022_script_ratings.sql, script.ratings.repository.ts, PUT /scripts/{id}/rating.
• Chatterbox's native path: the expressiveness dials (exaggeration, cfg_weight) per break and
  per voice in chatterbox.delivery.ts (chatterbox: a break can be read hushed or frantic, and a voice carries its own exaggeration #80), on POST /tts.
• A cue's fired or missed outcome: markSpokenOver in playout/live.order.ts.
• A console for segments: the Segments tab on the Voice page (components/segments/segments.page.tsx),
  with upload (/segments/upload) and an inbox scan. Adding one to the running order is
  POST /director/air/segments, owned by the page that owns the order.
• Persona memory, which "What the station SAID" pointed at: scoped and built under
  personas.

What is left is six smaller things: outro cues, an aired-at record of what the station said, a
setting for writer order, re-writing a failed break, getting a clip or voicepack onto a speech server
from the console (with a Kokoro blend helper beside it), and a live microphone.

Outro cues

"Finish two seconds before the record ends" is the other half of what makes a station sound like a
station rather than a playlist with announcements: a break over a long fade. The talk-up limit (#412)
is the front of a record; this is the back.

Both inputs now exist. outro_start and cue_out are measured per track (playout/rundown.ts
carries outroStartMs, and playout/crossfade.ts sizes the blend from it), and since #434 a spoken
segment has a real duration_ms. The original note here said Liquidsoap could measure the length
(request.duration on the resolved request) as a better answer than trusting a tag or a header; the
analyzer's measurement of the whole rendered file is that answer, and it is already stored.

What is not built is everything between: a writer told how long the outro runs (the same
constraint-on-the-request the talk-up uses, since it is cheaper to constrain the text than to trim
rendered speech), a cue that arms itself against outro_start rather than being handed a time, and
the decision about how an outro talk-over and the blend into the next record share the same seconds.
The fallback is the talk-up's: a record with no measured outro gets no outro cue, and the break airs
in the gap as it does today.

Play history for what the station SAID

Segments are deliberately excluded from deadair.play_history, because the repeat window and artist
cooldown are reads of it and an ident has no artist. deadair.script_history (the "What it said" tab)
records what was WRITTEN, not what aired, so it answers "what has the station been saying" and cannot
answer "what did it say at nine o'clock". markSpokenOver records whether a talk-over fired, which
is one fact about one kind of break. An aired-at stamp is still its own decision, and it still wants
its own table rather than a relaxation of play_history.

This is the same gap as #449 (an as-run log the operator can trust) seen from the break side. Whoever
builds one should build it as the other.

Which writer wins is registration order, not a setting

director.module.ts ranks the writers and llm.breakWriter is the only switch. That is right for one
model and one floor; an operator who wants the templates ABOVE the model on some shows wants a
setting, and the registry would take one without changing shape. A setting is a string: read it
through the registry and settingIsOn, per apps/api/CLAUDE.md.

A break still cannot be re-written

A failed segment is not re-claimable for writing, on the grounds that the usual reason nothing could
be written is that there was nothing true to say. An operator who edits their phrasings and wants last
night's failures retried has no button for it. The Segments page says so in its own header: it cannot
re-render a failed segment because there is no route that does, and a client-side retry would mean
deleting and re-creating the row, which is a different thing wearing the same button.

The one retry that does exist is narrower and should not be mistaken for this: a render that failed
because its engine was unavailable is released for retry (releaseForRetry in
segment.repository.ts, from #29). That is a render retry with the words kept, not a re-write.

Getting a clip or a voicepack onto the server from the console

Every engine wants this and the station has none of it. Chatterbox reads reference WAVs out of a
directory and Kokoro reads .pt voicepacks out of one, and in both cases the operator's only route
is a file copy on the host. (Rhapsode's server takes an upload through its own API, which is the
server's feature, not the station's.) The previous station had it as a PUT with the raw bytes as
the body and the name in the query string, hand-written rather than generated because a binary upload
does not fit the JSON-only contract codegen. /segments/upload has since shown the station can carry
an upload route, so that reason is weaker than it was. The compose file already mounts
./.docvol/voices for exactly this and seeds the image's own voicepacks into it, so the deployment
half is done and only the route is missing.

Chatterbox's clone voice mode is the related idea: the plugin asks for predefined and says a
reference clip the operator pushed is "a second way to name a voice", which belongs with the mapping
table if it is ever wanted. An upload route is what would make it wanted.

A blend helper

Kokoro takes af_bella(2)+af_sky(1) as a voice and also has a /v1/audio/voices/combine that mints a
persisted one. The expression is already typeable in the voice table's engine cell
(kokoro.manifest.ts keeps that cell a string rather than a select for this reason), so what is
missing is only the affordance for building and hearing one before committing it. That is a console
feature over an engine capability that works today.

The harbor mount

input.harbor("dj", ...) was removed when the voice queue replaced it, and radio.liq keeps the
harbor port only for the /control/* endpoints. Bring it back as a second arm of the voice source if
something genuinely needs to stream live audio in; a real microphone is the honest case.

Related

director-and-lineups for the rest of the
director's deferred half, and from-v1 for what
the previous station did here.

---

Moved from docs/todo/dj-voice.md. The file as it last stood: https://github.com/robert-dean/deadair/blob/71e431d45ce7b35a189c7d226832ec4a6247e9dc/docs/todo/dj-voice.md

[robert-dean]

Status check, 2026-09-22 (main at 23b983e)

Done: break ratings (0022_script_ratings.sql, script.ratings.repository.ts); Chatterbox's expressiveness dials (exaggeration / cfg_weight in chatterbox.delivery.ts, #80); a cue's fired/missed outcome (markSpokenOver in playout/live.order.ts).

Still open:

• The talk-up limit: introEndMs is measured, but there is no first-sustained-vocal field, no cue armed against the record, and no length limit handed to the writer.
• Outro cues and a measured segment duration.
• A console for managing segments, and an aired-at history of what the station said.
• A setting for writer order, and re-writing a failed break.
• Voice upload or voicepack routes, and a Kokoro blend helper.
• A live microphone harbor: radio.liq keeps the harbor only for /control.

[robert-dean]

Status, 2026-10-23. The aired-at log is covered by the as-run log (#765). deadair.as_run has a row for every break and programme as well as every record, with when it started and ended, read through GET /history/as-run and Check-up → As run. The others listed here are still open: outro cues, the writer-order setting, rewriting a failed break, voice uploads, the Kokoro blend helper, and a live mic (see #727).

[robert-dean]

Status, 2026-10-24. Three more are done in #774.

• Which writer wins. llm.breakWriterOrder ("Who is asked first", Settings, Words) puts the station's own words ahead of the model for every kind of break. A schedule slot or PutOnAirInput can choose its own with writerOrder (migration 0116); absent follows the station.
• Re-writing a failed break. POST /director/air/items/{itemId}/rewrite (the pencil on a break's row on the Desk) sends a break still to come back to planned, failed ones included, through SegmentRepository.reopenForRewrite. It refuses recordings, carried programmes and one turn of a production. A break whose records have already aired is deliberately not covered: it has nowhere true to go.
• Outro cues. breaks.talkOut (off) says a link over the end of the record before it when it fits between the last sung word and the measured cue_out, and that record then hard-joins the next: the talk-over owns the boundary. The writer is told the tail's length as an offer. This needed the cue clock fixed first: on_air_elapsed ran about 3.5% slow, and a cue due 170s in landed at 173.0s. It now measures elapsed time, and lands at 170.03s (stream/voiceclock.check.liq).

Still open: voice uploads from the console, the Kokoro blend helper, and a live mic (#727).