Voice Types

Voice types

Every speaker in a script gets a voice card in the Voices tab. This page explains the voice types you can choose and when to use each.

The speech models

The app uses three Qwen3-TTS models:

model	ID	used for
CustomVoice	Qwen3-TTS-12Hz-1.7B	9 built-in voices that follow delivery instructions
Base	Qwen3-TTS-12Hz-1.7B-Base	cloning a voice from reference audio, and LoRA voices
VoiceDesign	Qwen3-TTS-12Hz-1.7B-VoiceDesign	creating a voice from a text description

Models load when first needed and stay cached. Only the model the current voice type needs is loaded into GPU memory.

The voice types at a glance

type in the Voices tab	model	best for	batched?
Custom Voice	CustomVoice	the narrator and main characters	yes
Built-in Voice	Base + a shipped LoRA	ready-made trained voices, no training needed	yes
Voice Clone	Base	a character who needs one specific voice	yes
LoRA Voice	Base + your LoRA	recurring characters who need a unique, lasting voice	yes
Voice Design	VoiceDesign	minor characters with 1–3 lines	no — one line at a time
Together	the members' own voices	a group speaking at once	—

Choosing a type

Is this a main character with many lines?
├─ Yes → Do you need a voice other than the 9 built-in ones?
│        ├─ Yes → LoRA Voice (train one) or Built-in Voice (use a shipped one)
│        └─ No  → Custom Voice
└─ No  → Does the character have more than 3–4 lines?
         ├─ Yes → Voice Clone (save a designed voice and use it as the reference)
         └─ No  → Voice Design (no setup needed)

Identity anchor and style timeline

Every voice card has a Character Style field. It is the voice's identity anchor: a fixed description (for example "Heavy Scottish accent" or "Refined aristocratic tone") that is sent with every line, so the per-line delivery instructions cannot change who the character sounds like.

• For Custom voices the anchor comes first and the line's own instruction follows it: "Heavy Scottish accent Excited, bright energy.". Measured on 120 real lines, this cut the narrator's pitch drift from 3.49 to 2.53 semitones and kept the voice closer to its own opening (ECAPA 0.737 → 0.771).
• For LoRA and Built-in voices the character style is added after the line's instruction: "Excited, bright energy. Heavy Scottish accent".

The style timeline lets a character change from a given line onward — "older after the time skip" — without becoming a different speaker. Each change point sets a new anchor from that line on.

Generate Personas (assign voices automatically)

Instead of setting up each speaker by hand, click Generate Personas:

1. The language model reads sample lines for each speaker.
2. It writes a voice description ("A gruff middle-aged man with a deep, gravelly voice") and a sample sentence.
3. The VoiceDesign model renders reference audio from each description.
4. Each speaker is assigned that audio as a Voice Clone.

Options:

• Scope — by default only characters without a voice yet; you can choose all characters.
• Save the current voices to the library first — keeps what you already have.
• Advanced — sets the batch size (default 40), for scripts with very large casts.

Notes:

• Persona voices are clone voices: they keep a consistent identity but a neutral tone, because clone voices do not follow delivery instructions.
• For more expressive voices, start with personas and then switch key characters to LoRA or Custom.
• The generated reference audio is saved in designed_voices/ and can be reused.

Suggest LoRA Voices does something similar with trained voices: it matches the LoRA voices you have to your characters.

Aliases

The Alias of dropdown on a voice card makes one speaker use another speaker's voice.

Use it for:

• name variants: "DR. SMITH" → "SMITH";
• age variants: "YOUNG ELENA" → "ELENA";
• titles: "KING MARCUS" → "MARCUS";
• cutting down the number of voices to set up.

How it works:

• When audio is generated, an aliased speaker uses the target speaker's voice.
• Aliases chain: if A → B → C, speaker A uses C's voice.
• A loop (A → B → A) is detected, logged as a warning, and stopped.
• The alias is stored in voice_config.json as the speaker's alias_of field.

Custom Voice

Model: CustomVoice. Best for: the narrator and main characters.

Choose one of 9 built-in voices, with full control over emotion and delivery through the line's instruction.

• Voices: Aiden, Dylan, Eric, Ono_anna, Ryan, Serena, Sohee, Uncle_fu, Vivian.
  • Most are not native English speakers; Ryan and Aiden are.
• Settings: the voice, and the Character Style (the identity anchor).
• Batching: yes — lines are batched and split into sub-batches by length.

Built-in Voice

Model: Base with a shipped LoRA adapter. Best for: a ready-made voice without training.

The repository ships a set of trained voices in builtin_lora/. They download from Hugging Face the first time you use one, and behave like LoRA voices.

Voice Clone

Model: Base. Best for: a character who needs one specific voice.

Clones a voice from a 5–15 second reference recording.

• Settings:
  • Designed Voice — pick a voice saved in the Designer tab (fills in the audio path); or
  • Custom Path — the path to any WAV file;
  • Reference Text — the exact transcript of the reference audio.
• Tips:
  • Use clean audio with no background noise.
  • References shorter than about 7 seconds give an unstable voice; the import check normalises and measures them.
  • The Designer tab can create reference audio from a text description.
  • Uploaded references record their transcript, source and rights basis.
• Delivery instructions are ignored for clone voices: the voice comes entirely from the reference.
• Batching: yes — lines are grouped by speaker and batched together.

LoRA Voice

Model: Base with a LoRA adapter. Best for: recurring characters who need a unique, lasting voice that follows delivery instructions.

A LoRA adapter fine-tunes the Base model on a small set of voice samples. The result is more consistent than a clone, and it responds to emotion and delivery instructions.

• Settings: the adapter (from the Training tab, or a built-in one), and the Character Style.
• How it works:
  1. Train an adapter on your voice samples (see the Training guide).
  2. At generation time the adapter is loaded on top of the Base model.
  3. A reference clip from training provides the voice prompt.
  4. The anchor plus the line's instruction control the delivery.
• Clone or LoRA?
  • Clone — quick, no training; the voice comes from one reference clip.
  • LoRA — more consistent over many lines and follows emotional direction, but needs training.
• Batching: yes — lines are grouped by adapter; each adapter's group is batched on its own.

Voice Design

Model: VoiceDesign. Best for: minor characters with 1–3 lines.

Creates a voice on the fly from a text description. Each line gets its own freshly generated voice, so there is no lasting identity between lines.

• Setting: Description — the character's base voice ("Young strong soldier", "Old gravelly bartender").
• How the description and the line's instruction combine:
  • description: "Young strong soldier";
  • line instruction: "Shouting in rising desperation, with hints of panic";
  • sent to the model: "Young strong soldier, Shouting in rising desperation, with hints of panic".
• Tips:
  • Keep the description about the voice itself (age, gender, vocal quality), and leave emotion and delivery to the instructions.
  • Avoid contradictions such as "calm gentle voice" with "screaming furiously" — they can cause audio glitches.
  • For a character with more than 3–4 lines, save a designed voice and use it as a clone.
• Batching: no — each line may have a different description, so these lines are rendered one at a time after all batched types.

Together

For a line spoken by several characters at once. Each member speaks the line in the voice they already have, and the clips are mixed together. The clips are aligned to the longest one, so the result sounds like a chorus rather than exact unison.

How consistent is each type?

type	consistency	why
Custom	high	the same built-in voice every time (with a seed)
Clone	high	the same reference audio gives the same voice
LoRA	high	a trained identity, stable across lines
Voice Design	low to medium	the same description keeps voices similar, not identical

For Voice Design, consistency depends on how similar the combined description and instruction are from line to line. If the description stays the same and only the emotion changes, the voices sound close enough not to be jarring.

The Editor's Check Voices measures this directly: it compares each rendered line with its speaker's reference and flags lines that drifted.