Make a voice

Every character ships with a voice, but the stock voices are somebody's guess at what a character sounds like. This page makes one that fits: design it from a description, clone it into the speech server, and point the character at it. The example uses Rhapsode with its Chatterbox engine, which clones from a short clip, and ElevenLabs connected to Claude for the design. Any engine that clones from a clip works the same way. Only the upload step changes.

Three names

A character never names a sound file. It names a station voice, and the speech plugin's Voices table says what that name is on its engine:

Where	What it holds	Example
The character's Voice (Voice → Characters, edit)	a station voice name	latenight
Rhapsode's Voices table (Settings → Plugins → Rhapsode)	that name, an engine and an engine voice	latenight, chatterbox, latenight
The Rhapsode server	the cloned voice and its reference clip	latenight

Every built-in character's Voice is already its own key (classic, latenight, countdown and so on). Use the same word all three times, the character's key, and the wiring is one table row. Breaking that rule is allowed, but it is the first thing to check when a character speaks in the wrong voice.

Designing it

In Claude, with the ElevenLabs connector added, ask it to design a voice. It needs two things from you:

• A description. Work from the character's own sheet: how they talk, where they are from, how old they are, what they sound like at three in the morning. Describe an adult. A description of a child's voice is refused, and "a young man, around nineteen" gets a young voice where "a teenage boy" does not.
• A line for it to speak, at least a hundred characters. Write it in the character's manner, or take one from their phrasings, because the clip is the whole of what the clone learns from: a voice read flat in the design is flat on air.

ElevenLabs answers with a few previews of 13 to 15 seconds each. Listen to them all and pick one.

Use Voice Design, which invents a voice, and not a voice from ElevenLabs' library: those are recordings of real people, licensed to that marketplace, and cloning one onto your station is cloning a person. For the same reason, don't name a real singer or presenter in the description. Describe the manner you want instead. Check that your ElevenLabs plan lets you use what it generates in the way you mean to.

If the station speaks through the ElevenLabs plugin, there is nothing to clone. Save the preview you picked to your ElevenLabs account, and it appears by name in that plugin's Voices table. Give it the character's voice name there and skip to wiring it to the character.

What survives cloning

A designed voice is cloned a second time when it reaches your server, and not every voice comes through it. Breathy, whispery and husky designs come out thin, with the low end gone and too much upper midrange: it sounds like the character is on the phone. Ask for a full, clear, chest-resonant voice, "clear rather than whispery", and they clone with their body intact. If two previews sound equally right, the fuller one will sound better on air.

Cloning it into Rhapsode

The preview you picked is already the right length for Chatterbox, which takes a reference of about 5 to 20 seconds. Download its audio, and keep the line it speaks: that is the transcript.

Creating a voice writes a file on the Rhapsode server, so Rhapsode treats it as a management route and answers it only from the server's own machine, or from anywhere with its management token. The simplest way is Rhapsode's own page: under Try it, choose Chatterbox and clone from the clip. From another machine, with management.token set in rhapsode.config.json:

curl -X POST https://rhapsode.example.com/engines/chatterbox/voices \
  -H "Authorization: Bearer $RHAPSODE_TOKEN" \
  -F id=latenight \
  -F label="Sonny, late nights" \
  -F transcript="The words spoken in the clip, exactly as they are spoken." \
  -F reference=@latenight.mp3

A 201 answers with the new voice, including a spec. Ask GET /engines/chatterbox/voices/latenight/preview for a short sample.

Keep the clip. Rhapsode has no way to give a reference back, and its preview is a few seconds long, too short to clone from. If you ever want this voice again after replacing it, your copy is the only one.

Wiring it to the character

Under Settings → Plugins → Rhapsode, add a row to Voices: Station voice latenight, Engine chatterbox, Engine voice latenight. The engine voice cell offers what the server reports; a voice cloned a moment ago may need a refresh to appear, and can be typed in the meantime. Leave Variant empty to use whichever Chatterbox model the server has loaded, or name one (turbo, original, multilingual). Turbo performs laughs and sighs; the other two have dials for how theatrical a voice is, and no build does both.

If the character's Voice is already latenight, that is all. Otherwise set it on the character. Voice → Voices plays a sample of every station voice through the same path a break airs on, so listen there with a line that is not the one in the clip, which tests the clone rather than replaying its reference. The next break the character writes is in the new voice.

Changing it later

Cloning to the same id again replaces the voice, and its spec changes to say so. Nothing on the station needs to change: the next thing it speaks in that voice is the new one.

If it goes wrong

• The clone answers 403 forbidden. It came from another machine without the management token, or from a web page whose origin Rhapsode does not trust. Use the token, or Rhapsode's own page.
• It sounds like a telephone. The design was breathy or whispery. Design again and ask for a full, clear voice.
• The character speaks in the plugin's default voice. Check the three names. A character whose Voice has no row in the table falls back to the default, and the station's log says "no mapping for this voice" when it does.
• The character has gone quiet, and the sample on Voice → Voices fails. The row names an engine voice the server does not have, so Rhapsode refuses it and the break is skipped rather than aired late. Check the id against GET /engines/chatterbox/voices.

Next

Write a character to give the new voice somebody to speak for.