voice-line

/voice-line — Generate TTS Voice for NPCs, Narrator, Dialogue

Overview

The hardest part of TTS isn't the line — it's picking the right voice. ElevenLabs ships hundreds; pick wrong and a tough warlord NPC sounds like a podcast host. This skill walks the voice-id discovery flow (summer_list_models family audio-voice), narrows by gender / accent / age / energy, presents 4–5 finalists with sample URLs for the user to audition, and locks the pick before generating.

Then it generates one line (or a multi-turn dialogue) with the right stability / style / similarity / speed for the delivery — flat for narration, expressive for a bark, fast for an excited shout.

When to use

• NPC bark, narrator line, dialogue between two characters, tutorial voice, menu announcer.
• Multi-line conversation with two or more voices → use text_to_dialogue capability.

When NOT to use

• Non-verbal vocal SFX (grunt, scream) → audio/sound-effect with prompt male grunt of pain, mid-thirties, sharp, 400ms.
• Pre-recorded voice from a real actor — import the .mp3 / .wav with summer_import_from_url and skip TTS.

Steps

1. Read the audio bible and any character notes

Read .summer/audio-bible.md
Read .summer/memory/casting/voices.md   # preferred cast memory, if present
Read .summer/voice-cast.md              # legacy cast memory, if present
Read .summer/characters.md   # if present
Glob .summer/characters/*.md
Glob .summer/memory/characters/*.md

If a character bible exists, the voice should match the character's age / gender / regional origin / energy. If it doesn't, ask:

Tell me the character: gender, rough age, accent or regional flavor, and energy (calm / measured / excited / gruff). Or just give me a reference — "sounds like the captain in Mass Effect" works.

2. Browse the voice catalog

summer_list_models(family="audio-voice")

Returns a list with id, name, accent, gender, age, description, and previewUrl. Filter mentally to 8–12 candidates by the character's hard constraints (gender + age + accent), then pick 4–5 finalists with maximally distinct character (gruff vs warm vs neutral vs theatrical).

Present them to the user like this:

Four finalists for "old gruff dwarven warlord":

1. Brogan — male, 50s, rough Scottish, gravelly. Sample: <previewUrl>
2. Hodge — male, 60s, neutral, weathered baritone. Sample: <previewUrl>
3. Marrick — male, 40s, rough English, mid-energy. Sample: <previewUrl>
4. Old Drum — male, 70s, deep, slow speech. Sample: <previewUrl> Audition and pick one. Or ask for four more.

Wait for the user. Do not pick for them — voice is the highest-stakes audio decision in the project.

3. Character-to-voice decision tree (when the user wants a recommendation)

If the user says "you pick", lean on this tree:

Character archetype	Look for
Old gruff warlord	male, 50–70, rough or weathered, low pitch, slow speech, regional accent for color
Wise mentor	male or female, 50+, neutral or warm, measured cadence, mid-low pitch
Excited young hero	18–28, high energy, mid-high pitch, your own region's dominant accent
Cold villain	30–50, neutral or theatrical, low-mid pitch, slow with sharp consonants
Mischievous trickster	varies age, theatrical, light pitch, fast cadence
Calm scholarly NPC	30–50, neutral, mid pitch, even cadence, no regional color
Scared child	child voice if available; otherwise high-pitched young adult with high stability
Robotic / AI	flat / synthesized voice; or any voice with stability=1.0 and style=0.0 for monotone
Narrator (cinematic)	mid-deep male or female, neutral or theatrical, even cadence, very stable
Narrator (intimate)	warmer, mid pitch, lower stability for breathy nuance

4. Lock the voice id

After the pick:

Locked voice: Hodge (id 21m00Tcm4TlvDq8ikWAM). I'll use this for all <character> lines unless you say otherwise.

Ask before writing:

May I update .summer/memory/casting/voices.md with this locked voice assignment?

Save it in .summer/memory/casting/voices.md:

---
id: casting.voice.main-cast
type: casting
status: active
priority: locked
stable: true
providers:
  - elevenlabs
---

# Voice Cast

| Character | Voice name | Voice id | Notes |
|---|---|---|---|
| Captain Hodge | Hodge | 21m00Tcm4TlvDq8ikWAM | Calm, weathered, low-mid |
| Bram the Smith | Brogan | <id> | Gruff Scottish, mid energy |

This is the cast bible. Every future voice line uses these ids — consistency is the point.

If legacy .summer/voice-cast.md already exists, read it first and either keep using it for this project or ask to migrate it into .summer/memory/casting/voices.md. Do not maintain two conflicting cast files.

5. Tune stability / similarity / style / speed for delivery

ElevenLabs voice options:

Param	Range	Effect
stability	0.0–1.0	Higher = monotone, consistent. Lower = expressive, variable.
similarity_boost	0.0–1.0	Higher = closer to the original sample. Lower = more freedom (can drift).
style	0.0–1.0	Higher = more theatrical / exaggerated. 0 = neutral.
speed	0.7–1.2	Speech speed multiplier.

Defaults that work as a starting point:

Delivery	stability	similarity_boost	style	speed
Calm narration	0.75	0.75	0.0	1.0
Cinematic narration	0.5	0.85	0.4	0.95
NPC bark (combat shout)	0.3	0.75	0.6	1.1
NPC bark (calm idle)	0.6	0.75	0.2	1.0
Excited youth	0.35	0.7	0.55	1.1
Cold villain	0.65	0.85	0.3	0.92
Robotic	1.0	0.85	0.0	1.0
Scared / breathy	0.25	0.7	0.5	1.05

Ask:

Delivery flavor? Calm idle / combat shout / cinematic / robotic / scared / cold? I'll set stability and style to match.

6. Build the line

For a single line, the prompt is just the text. Punctuation and formatting matter:

• Periods = clear pauses (~300ms). Use them.
• Commas = short pauses (~100ms).
• Em dashes = sharp cut. (Note: avoid em dashes in user-facing UI copy per project standard, but they're fine in voice prompts as a delivery cue.)
• All-caps word = stress: You SHALL not pass.
• Ellipsis = trailing pause: I... I don't know what to say.
• Parenthetical stage direction is ignored — (angrily) is read as the word "angrily". Use stability/style instead.

Example lines:

"Hold the line. They're coming through the south gate. Ready arrows."
"You shall not pass. Turn back."
"I... I don't know if I can do this."
"BREACH! BREACH! Fall back to the inner courtyard!"

7. Confirm and call

summer_generate_voice(
  text="Hold the line. They're coming through the south gate. Ready arrows.",
  voiceId="21m00Tcm4TlvDq8ikWAM",
  modelId="eleven_multilingual_v2",
  stability=0.6,
  similarity_boost=0.75,
  style=0.2,
  speed=1.0,
  outputPath="audio/voice/hodge_hold_the_line.mp3"
)

modelId defaults to eleven_multilingual_v2. For very short / latency-sensitive in-game barks, eleven_turbo_v2_5 is faster; for highest fidelity, eleven_multilingual_v2 is the default.

8. Multi-line dialogue (two or more speakers)

Use summer_generate_audio with capability: "text_to_dialogue":

summer_generate_audio(
  capability="text_to_dialogue",
  speakers=[
    { name: "Hodge",   voiceId: "21m00Tcm4TlvDq8ikWAM" },
    { name: "Brogan",  voiceId: "<brogan-id>" }
  ],
  script=[
    { speaker: "Hodge",  text: "They're at the gate." },
    { speaker: "Brogan", text: "Aye. Let 'em come." },
    { speaker: "Hodge",  text: "Steady. On my mark." }
  ],
  outputPath="audio/voice/dialogue_gate_warning.mp3"
)

This produces a single MP3 with both voices, with dialogue-aware pacing the model handles internally. For very long scripts, split into chunks of ~6 turns and concatenate.

9. Wire the line as AudioStreamPlayer on the Voice bus

summer_add_node(parentPath="/root/Game/NPC", type="AudioStreamPlayer", name="VoiceLine")
summer_set_prop(path="/root/Game/NPC/VoiceLine", property="stream", value="res://audio/voice/hodge_hold_the_line.mp3")
summer_set_prop(path="/root/Game/NPC/VoiceLine", property="bus", value="Voice")
summer_set_prop(path="/root/Game/NPC/VoiceLine", property="volume_db", value=0.0)
summer_set_prop(path="/root/Game/NPC/VoiceLine", property="autoplay", value=false)

For positional NPCs, AudioStreamPlayer3D with max_distance=15.0 so distant NPCs aren't audible.

The bible's mix rule says voice ducks music: when this player starts, fade the music bus by -6 dB. Wire this in the bus setup or in code with AudioServer.set_bus_volume_db().

10. Save to the cast bible

After generation, append to .summer/memory/casting/voices.md so future skills find existing lines and don't regenerate. Include line text + filename + voice id.

If the assignment is marked priority: locked, do not replace the voice id automatically. If a scene, script, import config, or user prompt conflicts with the locked memory, stop and ask whether to update memory or fix the implementation.

Reference card — voice-pick by archetype

Old gruff warlord:        male, 50-70, rough/weathered, low pitch, slow
Wise mentor:              male/female, 50+, warm, measured, mid-low pitch
Excited young hero:       18-28, high energy, mid-high pitch
Cold villain:             30-50, neutral/theatrical, low-mid, slow with sharp consonants
Mischievous trickster:    theatrical, light, fast cadence
Calm scholar:             30-50, neutral, mid, even cadence
Scared / breathy:         high stability=0.25, style=0.5
Robotic AI:               stability=1.0, style=0.0
Narrator cinematic:       mid-deep, neutral/theatrical, even, very stable
Narrator intimate:        warmer mid pitch, lower stability for breath

Anti-patterns

• Picking a voice without auditioning samples. The user must hear it.
• Reusing a voice id across two characters. Players notice instantly. Lock the cast.
• Changing a locked voice without approval. A cast voice is project memory, not a throwaway generation option.
• Stage directions in the text. (angrily) Get out! is read literally as "angrily get out". Use stability/style.
• One long blob for a multi-turn scene. Use text_to_dialogue so pacing is dialogue-aware.
• Voice on Master bus. Must be on Voice bus so it can duck music per the bible.
• No ducking on dialogue. Music drowns voice; players miss lines. Always duck.
• Picking a voice that doesn't match the character bible. A noble queen with a cockney accent breaks immersion unless that's the joke.

Edge cases

• Line longer than ~1000 characters. ElevenLabs caps per-call. Split at sentence boundaries, generate each, concatenate with AudioStreamPlayer.queue or chained finished signals.
• Mispronunciation of a fictional word. Spell it phonetically: Caelthorne → KEL-thorn or KAYL-thorn. Or use SSML if the model supports it.
• Voice drifts in long lines. Increase similarity_boost toward 0.9; or split.
• Need a non-English line. eleven_multilingual_v2 supports 29 languages. Just write in the target language; the same voice id works.
• User wants the line shorter on regen. Edit the text first; don't try to compress with speed past 1.15 (artifacts).
• NPC bark variants. Generate 3–5 variants of the same intent ("They're at the gate!", "Enemy approaching!", "Look out!") and pick at random in code so the bark doesn't repeat.

Fallback (no MCP)

Print the call (text, voice id, params, target path). User runs via the Summer dashboard, then summer_import_from_url the result.

Handoff

Line hodge_hold_the_line.mp3 wired to NPC/VoiceLine on the Voice bus. Cast bible updated. Next:

• Generate 3 bark variants of the same intent so the NPC doesn't sound like a recording.
• For the upcoming gate scene, run /voice-line with text_to_dialogue for the multi-turn exchange.
• Set up music ducking on the Music bus when Voice is active (one Tween or a BusEffect Compressor sidechained to Voice).

See also

• audio/audio-direction — Voice bus and ducking rules
• audio/sound-effect — non-verbal vocal SFX (grunt, scream)
• audio/adaptive-music — ducking integration