Kovati Docs
SpeechForge

Casting

Speakers, voice profiles, and the Cast page.

Documented in depthSpeechForge0.2.2Open source

Two assets, and the distinction between them is the whole model.

What it isAnalogy
Speaker (SP_)The character sheet — who is speakingThe role
Voice Profile (VP_)The instrument — provider, preset, model, settingsThe performer

A speaker references a voice profile. Several speakers can share one; a speaker can be recast by pointing at a different one, and every line that speaker owns follows.

The speaker sheet

Four sections, and the third one does more work than its size suggests.

The SP_Doric asset open in its own editor, headed 'Asset Type: Speech Speaker'. An Identity section holds Speaker Id 'Doric', Display Name 'Emil Doric' and an empty Description. A Voice section holds Voice Profile VP_Will___Relaxed_Optimist. A Rig section holds Rig Target Face_Archetype_Skeleton and Face Mesh SKM_MH_Demo_Person_2_FaceMesh, each with an asset thumbnail. A Links section holds External Bindings: one map element keyed 'NarrativePro' pointing at NPC_Doric.
One character, described once. The Voice section is what SpeechForge reads, the Rig section is what FaceForge reads, and the Links section is what the dialogue system reads — three plugins answered by one asset nobody has to keep in step.
SectionFieldWhat it decides
IdentitySpeakerIdThe handle. Searchable in the asset registry, so resolution is a tag lookup rather than a scan
DisplayNameWhat a person reads
DescriptionCasting notes. Genuinely used — this is what a director reads when choosing a voice
VoiceVoiceProfileThe reference to an instrument
RigRigTargetThe skeleton this character's face is animated on
FaceMeshThe head itself — used for preview and for coverage checking
LinksExternalBindingsA map from a framework name to an asset path

The Rig section is why a scene with two character standards sorts itself out. The Perform page prints a Rig per line and routes each to the face bank that bakes against that skeleton — and it knows which is which because the speaker sheet said so. Nobody assigns lines to face banks by hand.

In this scene the routing is exact: SP_Astronaut names Astronaut_Skeleton, the other three name Face_Archetype_Skeleton, and the Perform page's Rig column prints those two values against the right lines.

The Face Mesh is read separately — it is what the Face Bank panel previews on, and what coverage checking measures against.

External bindings let a speaker point at your dialogue system's own character definition without SpeechForge knowing what that is: the key is a framework name you choose, the value is any asset path. In the capture above, NarrativePro → NPC_Doric is what the Cast page's Links column counts.

The set of speaker assets is the cast list. There is no separate list, so there is nothing to fall out of step.

The voice profile

Display name, the provider's own voice name (cached), and the provider/preset/model/settings/provenance.

Caching the provider's voice name is worth noting: it means the Cast page can show you what a profile actually points at without a network call every time you open it.

The Cast page

Three things side by side.

The Cast page of the Speech Library. On the left, 'The cast' with a Show all speakers checkbox and a New speaker id field: four rows — Dorian Smith, Emil Doric, Ivo Renk and Player, each marked 'in scene' — with columns Voice, Lines, Links and an Open button. The three NPCs each have 2 lines and 1 link; Player has 1 line and a dash. On the right, 'Voice profiles - reusable, provider-agnostic' lists Callum - Husky Trickster, Lily - Velvety Actress, Roger - Laid-Back Casual Resonant and Will - Relaxed Optimist, each tagged (ElevenLabs, 1 speaker) with a play button, Assign and Open. Below, 'Browse provider voices' with an ElevenLabs picker and a Fetch Voices button, and explanatory text about the ElevenLabs Voice Library. A line at the foot reads: No casting notes on the sheet yet - Open to add some.
The cast on the left, the instruments on the right, and the vendor's catalogue underneath them. '(ElevenLabs, 1 speaker)' is the used-by count answering the question a profile list otherwise raises: is anyone actually using this?
  • The cast list — sorted by who appears in the scene, with create sheet offered for any speaker id that lines reference but no sheet exists for. · in scene marks a speaker this bank's lines actually reference; Show all speakers widens it to every sheet in the project.
  • The profiles — each with used by, and Assign.
  • A provider browser — the vendor's own voice library, with Save as Profile and Cast in one step.

That last one is the path most people take: find a voice you like in the vendor's library, and cast it onto a speaker in one action rather than creating a profile, naming it, and then assigning it.

The Cast page after pressing Fetch Voices. Under 'Browse provider voices' a line reads '21 voice(s) on the account.', followed by guidance: browse the ElevenLabs Voice Library and press Add on any voice — it lands in My Voices and appears here after Fetch; designing a voice from a description or cloning one from recordings lives in the same app, under Voices. A link reads 'Add voices on ElevenLabs'. Beneath, a list of stock voices — Roger, Sarah, Laura, Charlie, George, Callum, River, Harry, Liam — each marked (stock) with a play button, a Save as Profile button and a Cast button.
Twenty-one voices on this account, each with three things you can do to it: hear it, keep it as a profile, or give it straight to the selected speaker. The page also says plainly where voices come from, because 'why is my new voice not here' is otherwise a support question.

Two buttons, and the difference is whether you will use the voice twice. Save as Profile creates a VP_ asset you can assign to several speakers and recast later in one place. Cast skips that and gives the voice to the selected speaker directly.

Prefer the profile for anyone with more than a line or two — recasting a profile moves every speaker that references it, and recasting a speaker moves only that speaker.

Adapters seed sheets

An adapter for a dialogue framework does not resolve voices — it creates or updates speaker sheets and sets the binding.

Empty fields never clobber a cast voice. An adapter re-harvesting a dialogue will not blank a voice somebody chose, because a field it has no opinion about is left alone rather than written as empty.

This matters because harvesting tends to be re-run, and the first version that overwrites a director's casting is the last version anyone runs.

One concrete trap from the Narrative Pro adapter: an authoring template's speaker array is routinely empty, so the harvest reads speakers off the dialogue's class defaults instead.

Per-line overrides

A line can override its speaker's voice. It is the first step of the resolution walk, and the result says so.

Use it for the line where a character whispers, shouts, or is heard through a radio — not as a substitute for casting.

Edit on GitHub

Concepts

Banks, lines, handles, and the Speech Library.

Providers

ElevenLabs, and what a speech provider declares about itself.

On this page

The speaker sheetThe voice profileThe Cast pageAdapters seed sheetsPer-line overrides