The template lives in the clock
An AI link starts as a slot in the clock template with a template: introduce the next song, lead into the news, read a listener's request, give a time check, say something about the artist that just played. The template says what, not how. It carries the presenter's persona: name, tone, daypart, and the station's rules, such as whether a jingle may sit next to it.
The script is written late
Only when the hour is prepared and the planner knows which songs really stand there does a language model write the script. It gets the title and artist of the song that just played and the one coming up, this hour's weather, the news headlines, the time, and the persona. A presenter saying 'good morning' at 23:30 is a script written too early; that is why the window is short and the script is rewritten on any change.
For Spanish stations: Castilian only, the Spanish of Spain. No celular, no ustedes. That sits in every prompt and is checked in code, because an instruction alone does not hold.
The voice
The script goes to a voice model with a fixed voice per presenter. Different voices have different strengths; some take direction such as warmth or pace, others do not, and we measure that per voice instead of assuming it. The raw render is too quiet and too uneven for air, so it goes through a fixed treatment: high-pass, loudness normalisation, compression and a limiter, to 24-bit. That is the only place where voice is processed; after that nothing touches the volume.
A render is checked before it becomes usable: too short, silent or broken means render again, not broadcast. A broken render that did go out gave us twelve seconds of silence after a short jingle; since then the check sits before saving.
Placement to the millisecond
Once the voice is ready, the system knows how long it is. Then it is placed: it starts on the extro of the outgoing song, talks over the tail, and whatever is left goes over the intro of the next song until the vocals begin. If the successor is not a song but a news bed or jingle, the voice overlaps at most three hundred milliseconds, so the news never starts through the link. That check happens at air time against the real successor, because a scheduled song may have been skipped at the hour boundary.
Phone calls and requests
A phone call is a script with two voices: the presenter and a caller, each with its own voice model, edited into one conversation with pauses and a telephone tone on the caller. A listener request comes in via the website or the app, is checked for sense and read in the next suitable slot, with the requester's name if they wish. Both are the exceptions to the rule that the clock decides everything: an event from a listener or operator may insert something, provided it is within the current hour and has a verifiable reason.
What an AI DJ is not
Not a replacement for a live show with a human carrying the conversation of the day. It is a presenter who is on time twenty-four hours a day, names the right title, gives this hour's weather and never says 'good morning' at midnight. For most stations it is a combination: people on the hours that matter, AI on the rest, with the same clock and the same jingles.
Frequently asked questions
Does the AI voice sound the same every hour?
No. The script is written per link with what is playing then, and the form varies: introduction, fact about the artist, time check, weather, request. The tone follows the daypart too.
Can the AI DJ use my own voice?
A voice is fixed per presenter. Cloning your own voice is possible with the owner's consent and enough clean source material; we discuss that during setup.
How do you stop the AI saying nonsense?
The script only gets facts the system knows: titles, artists, time, weather, headlines. It may not invent anything about the artist that is not in the source, and scripts are filtered for words and constructions that do not belong on air.