digitalsamba/claude-code-video-toolkit / .claude/skills/elevenlabs
elevenlabs
VerifiedGenerate AI voiceovers, sound effects, and music using ElevenLabs APIs. Use when creating audio content for videos, podcasts, or games. Triggers include generating voiceovers, narration, dialogue, sound effects from descriptions, background music, soundtrack generation, voice cloning, or any audio synthesis task.
ElevenLabs Audio Generation
Requires ELEVENLABS_API_KEY in .env.
Text-to-Speech
from elevenlabs.client import ElevenLabs
from elevenlabs import save, VoiceSettings
import os
client = ElevenLabs(api_key=os.getenv("ELEVENLABS_API_KEY"))
audio = client.text_to_speech.convert(
text="Welcome to my video!",
voice_id="JBFqnCBsd6RMkjVDRZzb",
model_id="eleven_multilingual_v2",
voice_settings=VoiceSettings(
stability=0.5,
similarity_boost=0.75,
style=0.5,
speed=1.0
)
)
save(audio, "voiceover.mp3")
Models
| Model | Quality | SSML Support | Notes |
|---|---|---|---|
eleven_multilingual_v2 | Highest consistency | None | Stable, production-ready, 29 languages. Toolkit default; supports style + speed |
eleven_v4 | Most expressive | None (audio tags, IPA) | Newest (released 2026-09-28), 90+ languages, best clone accuracy. Stability + similarity only, no style/speed |
eleven_flash_v2_5 | Good | <break>, <phoneme> | Fastest, cheapest; supports pause/pronunciation tags |
eleven_v3 | Expressive | None | Previous generation — prefer v4 |
Choose: multilingual_v2 for reliability and speed/style control, v4 for maximum expressiveness and cloned voices (test a few takes), flash for SSML control.
Deprecated: eleven_turbo_v2_5 → use eleven_flash_v2_5 (same quality, lower latency; see deprecated models). The toolkit still accepts --model eleven_turbo_v2_5: it prints a one-line warning and sends eleven_flash_v2_5.
Eleven v4
Released 2026-09-28 (changelog, model page). The toolkit calls v4 through the Text to Dialogue API with one input (the narrator voice), as ElevenLabs recommends for content creation and long-form audio:
from elevenlabs import DialogueInput
audio = client.text_to_dialogue.convert(
inputs=[DialogueInput(text="[warm] Welcome to my video!", voice_id="JBFqnCBsd6RMkjVDRZzb")],
model_id="eleven_v4",
settings={"stability": 0.5, "similarity": 0.75}, # the only two settings v4 takes
)
save(audio, "voiceover.mp3")
uv run tools/voiceover.py --model eleven_v4 ...(andredub.py --model eleven_v4) do this for you, warn once that--style/--speedare ignored, and split text over ~2,000 chars per request (the reliable dialogue limit) at sentence boundaries. Pace with--max-wpminstead of--speed.client.text_to_dialogue.convert_with_timestamps(...)returns the same character alignment as TTSconvert_with_timestamps, soredub.py --syncworks with v4 and sends the same stability + similarity.- Pass
settingsas a dict: SDK versions before 2.69 type it with a stability-only model. The API readssimilaritythere since 2026-09-28 (it rejects an out-of-range value with a 422). Text to Dialogue needselevenlabs>=2.21. - Direct delivery with audio tags (
[whispering],[warm],[sigh]) and punctuation. Write tags that clearly describe the voice, because v4 can also read a tag as a sound effect. - A voice can change accent on one word. A British library voice read "thirty dollars" and "once a card is
on the account" with an American r, in every plain retake of those lines. A direction at the start of the line
held the accent in three takes of three:
[crisp British English accent] So what does thirty dollars buy?. v4 doesn't read the tag aloud. Put it in the text sent for voicing only, and keep the.txtbeside the MP3 as the spoken words, since captions are spelled from it. It can be measured before anyone listens: with the word times fromcaptions.py, an American r pulls the third formant under about 2,200 Hz at the end of "dollars" or "card" in a woman's voice, and a British ending leaves it near 2,700 (Praat,uv sync --extra prosody).templates/howto/scripts/accent.pydoes this for a folder of narration (--scenes DIR) and lists the lowest first. A word before a vowel reads low in British speech too ("your account" has a linking r), so listen to what it flags before re-voicing. - A pace direction slows a line a little. v4 ignores speed, but
[crisp British English accent, measured and unhurried]at the start gave takes about 8% longer than the same line without "measured and unhurried" (four takes of three lines). For an opening line, make three or four takes and letpick_takes.pychoose. Its wpm counts the words of the tag, so put the spoken words back in each take's.txtbefore picking, not only beforecaptions.py: with the tag still there it read 33 lines as 180 to 300 wpm and slowed 31 of them, where the true paces were 150 to 185 and 6 needed a touch. eleven_v4_turbois the real-time variant, available only over the Text to Dialogue WebSocket. The toolkit doesn't use it, since offline renders don't benefit from low latency.
Voice Settings by Style
| Style | stability | similarity | style | speed |
|---|---|---|---|---|
| Natural/professional | 0.75-0.85 | 0.9 | 0.0-0.1 | 1.0 |
| Conversational | 0.5-0.6 | 0.85 | 0.3-0.4 | 0.9-1.0 |
| Energetic/YouTuber | 0.3-0.5 | 0.75 | 0.5-0.7 | 1.0-1.1 |
voiceover.py's own defaults (stability 0.85, style 0.0) sit at the top of the first row: steady but flat. A
library voice picked for its preview usually sounds slower and duller at those defaults than in the preview.
Lower the stability first (0.45 on v4 brought Marilyn back to her preview's pace and pitch range), then
measure the takes with voice_prosody.py measure and listen.
Pauses Between Sections
With flash models: Use SSML break tags inline:
...end of section. <break time="1.5s" /> Start of next...
Max 3 seconds per break. Excessive breaks can cause speed artifacts.
With multilingual_v2 / v3 / v4: No SSML support. Options:
- Paragraph breaks (blank lines) — creates ~0.3-0.5s natural pause
- v3/v4: audio tags such as
[long pause] - Post-process with ffmpeg: split audio and insert silence
WARNING: ... (ellipsis) is NOT a reliable pause — it can be vocalized as a word/sound. Do not use ellipsis as a pause mechanism.
Pronunciation Control
Phonetic spelling (any model): Write words as you want them pronounced:
Janus→Jan-usnginx→engine-x- Use dashes, capitals, apostrophes to guide pronunciation
SSML phoneme tags (flash only):
<phoneme alphabet="ipa" ph="ˈdʒeɪnəs">Janus</phoneme>
IPA in v4: no tags needed. Wrap the IPA in slashes and double quotes in place of the word:
The term "/ˌbaɪoʊˈkemɪstri/" refers to the study of chemical processes.
Iterative Workflow
- Generate → listen → identify pronunciation/pacing issues
- Adjust: phonetic spellings, break tags, voice settings
- Regenerate. If pauses aren't precise enough, add silence in post with ffmpeg rather than fighting the TTS engine.
Speech to Text (Scribe)
with open("voiceover.mp3", "rb") as f:
result = client.speech_to_text.convert(file=f, model_id="scribe_v2", tag_audio_events=False)
words = [w for w in result.words if w.type == "word"] # .text, .start, .end (seconds)
scribe_v2 is current and the toolkit default. scribe_v1 is deprecated, with removal announced for 2026-07-09 (changelog), but still answers, so --stt-model scribe_v1 (redub) and --model scribe_v1 (align_captions) still run it, with a warning. scribe_v1_experimental, a 2025 preview of scribe_v1, is accepted with a warning and sent as scribe_v2. The toolkit uses Scribe in redub.py (transcribe + --sync word timing) and align_captions.py.
Voice Cloning
Instant Voice Clone
with open("sample.mp3", "rb") as f:
voice = client.voices.ivc.create(
name="My Voice",
files=[f],
remove_background_noise=True
)
print(f"Voice ID: {voice.voice_id}")
- Use
client.voices.ivc.create()(notclient.voices.clone()) - Pass file handles in binary mode (
"rb"), not paths - Convert m4a first:
ffmpeg -i input.m4a -codec:a libmp3lame -qscale:a 2 output.mp3 - Multiple samples (2-3 clips) improve accuracy
- Save voice ID for reuse
Professional Voice Clone: Requires Creator plan+, 30+ min audio. See reference.md.
Sound Effects
Max 22 seconds per generation.
result = client.text_to_sound_effects.convert(
text="Thunder rumbling followed by heavy rain",
duration_seconds=10,
prompt_influence=0.3
)
with open("thunder.mp3", "wb") as f:
for chunk in result:
f.write(chunk)
Prompt tips: Be specific — "Heavy footsteps on wooden floorboards, slow and deliberate, with creaking"
Music Generation
10 seconds to 5 minutes. Use client.music.compose() (not .generate()).
result = client.music.compose(
prompt="Upbeat indie rock, catchy guitar riff, energetic drums, travel vlog",
music_length_ms=60000,
force_instrumental=True
)
with open("music.mp3", "wb") as f:
for chunk in result:
f.write(chunk)
Prompt structure: Genre, mood, instruments, tempo, use case. Add "no vocals" or use force_instrumental=True for background music.
Remotion Integration
Complete Workflow: Script to Synchronized Scene
VOICEOVER-SCRIPT.md → voiceover.py → public/audio/ → Remotion composition
↓ ↓ ↓ ↓
Scene narration Generate MP3 Audio files <Audio> component
with durations per scene with timing synced to scenes
Step 1: Generate Per-Scene Audio
Use the toolkit's voiceover tool to generate audio for each scene:
# Generate voiceover files for each scene
uv run tools/voiceover.py --scene-dir public/audio/scenes --json
# Output:
# public/audio/scenes/
# ├── scene-01-title.mp3
# ├── scene-02-problem.mp3
# ├── scene-03-solution.mp3
# └── manifest.json (durations for each file)
The manifest.json contains timing info:
{
"scenes": [
{ "file": "scene-01-title.mp3", "duration": 4.2 },
{ "file": "scene-02-problem.mp3", "duration": 12.8 },
{ "file": "scene-03-solution.mp3", "duration": 15.3 }
],
"totalDuration": 32.3
}
Step 2: Use Audio in Remotion Composition
// src/Composition.tsx
import { Audio, staticFile, Series, useVideoConfig } from 'remotion';
// Import scene components
import { TitleSlide } from './scenes/TitleSlide';
import { ProblemSlide } from './scenes/ProblemSlide';
import { SolutionSlide } from './scenes/SolutionSlide';
// Scene durations (from manifest.json, converted to frames at 30fps)
const SCENE_DURATIONS = {
title: Math.ceil(4.2 * 30), // 126 frames
problem: Math.ceil(12.8 * 30), // 384 frames
solution: Math.ceil(15.3 * 30), // 459 frames
};
export const MainComposition: React.FC = () => {
return (
<>
{/* Scene sequence */}
<Series>
<Series.Sequence durationInFrames={SCENE_DURATIONS.title}>
<TitleSlide />
</Series.Sequence>
<Series.Sequence durationInFrames={SCENE_DURATIONS.problem}>
<ProblemSlide />
</Series.Sequence>
<Series.Sequence durationInFrames={SCENE_DURATIONS.solution}>
<SolutionSlide />
</Series.Sequence>
</Series>
{/* Audio track - plays continuously across all scenes */}
<Audio src={staticFile('audio/voiceover.mp3')} volume={1} />
{/* Optional: Background music at lower volume */}
<Audio src={staticFile('audio/music.mp3')} volume={0.15} />
</>
);
};
Step 3: Per-Scene Audio (Alternative)
For more control, add audio to each scene individually:
// src/scenes/ProblemSlide.tsx
import { Audio, staticFile, useCurrentFrame } from 'remotion';
export const ProblemSlide: React.FC = () => {
const frame = useCurrentFrame();
return (
<div style={{ /* slide styles */ }}>
<h1>The Problem</h1>
{/* Scene content */}
{/* Audio starts when this scene starts (frame 0 of this sequence) */}
<Audio src={staticFile('audio/scenes/scene-02-problem.mp3')} />
</div>
);
};
Syncing Visuals to Voiceover
Calculate scene duration from audio, not the other way around:
// src/config/timing.ts
import manifest from '../../public/audio/scenes/manifest.json';
const FPS = 30;
// Convert audio durations to frame counts
export const sceneDurations = manifest.scenes.reduce((acc, scene) => {
const name = scene.file.replace(/^scene-\d+-/, '').replace('.mp3', '');
acc[name] = Math.ceil(scene.duration * FPS);
return acc;
}, {} as Record<string, number>);
// Usage in composition:
// <Series.Sequence durationInFrames={sceneDurations.title}>
Audio Timing Patterns
import { Audio, Sequence, interpolate, useCurrentFrame } from 'remotion';
// Fade in audio
export const FadeInAudio: React.FC<{ src: string; fadeFrames?: number }> = ({
src,
fadeFrames = 30
}) => {
const frame = useCurrentFrame();
const volume = interpolate(frame, [0, fadeFrames], [0, 1], {
extrapolateRight: 'clamp',
});
return <Audio src={src} volume={volume} />;
};
// Delayed audio start
export const DelayedAudio: React.FC<{ src: string; delayFrames: number }> = ({
src,
delayFrames
}) => (
<Sequence from={delayFrames}>
<Audio src={src} />
</Sequence>
);
// Usage:
// <FadeInAudio src={staticFile('audio/music.mp3')} fadeFrames={60} />
// <DelayedAudio src={staticFile('audio/sfx/whoosh.mp3')} delayFrames={45} />
Voiceover + Demo Video Sync
When a scene has both voiceover and demo video:
import { Audio, OffthreadVideo, staticFile, useVideoConfig } from 'remotion';
export const DemoScene: React.FC = () => {
const { durationInFrames, fps } = useVideoConfig();
// Calculate playback rate to fit demo into voiceover duration
const demoDuration = 45; // seconds (original demo length)
const sceneDuration = durationInFrames / fps; // seconds (from voiceover)
const playbackRate = demoDuration / sceneDuration;
return (
<>
<OffthreadVideo
src={staticFile('demos/feature-demo.mp4')}
playbackRate={playbackRate}
/>
<Audio src={staticFile('audio/scenes/scene-04-demo.mp3')} />
</>
);
};
Error Handling
import { Audio, staticFile, delayRender, continueRender } from 'remotion';
import { useEffect, useState } from 'react';
export const SafeAudio: React.FC<{ src: string }> = ({ src }) => {
const [handle] = useState(() => delayRender());
const [audioReady, setAudioReady] = useState(false);
useEffect(() => {
const audio = new window.Audio(src);
audio.oncanplaythrough = () => {
setAudioReady(true);
continueRender(handle);
};
audio.onerror = () => {
console.error(`Failed to load audio: ${src}`);
continueRender(handle); // Continue without audio rather than hang
};
}, [src, handle]);
if (!audioReady) return null;
return <Audio src={src} />;
};
Toolkit Command: /generate-voiceover
The /generate-voiceover command handles the full workflow:
/generate-voiceover
1. Reads VOICEOVER-SCRIPT.md
2. Extracts narration for each scene
3. Generates audio via ElevenLabs API
4. Saves to public/audio/scenes/
5. Creates manifest.json with durations
6. Updates project.json with timing info
Popular Voices
- George:
JBFqnCBsd6RMkjVDRZzb(warm narrator) - Rachel:
21m00Tcm4TlvDq8ikWAM(clear female) - Adam:
pNInz6obpgDQGcFmaJgB(professional male)
List all: client.voices.get_all()
Voice library (a particular accent or character)
The premade voices are mostly American. For another accent, search the shared voice library: free to search and to hear, and every voice has a preview MP3.
curl -s -G https://api.elevenlabs.io/v1/shared-voices -H "xi-api-key: $ELEVENLABS_API_KEY" \
--data-urlencode "search=irish" -d gender=female -d page_size=30 # each voice has voice_id, accent, preview_url
- Search with
search=. Theaccent=filter alone came back empty forirish,scottishandbritish. - Using a library voice through the API needs a paid plan. A free key gets 402
paid_plan_required. The cheapest tier (Starter) is enough; the voice ID then works directly in--voice-idor a brand. - Don't clone a preview into Qwen3 or another model. The owner licensed the voice through ElevenLabs (and is paid per use there), not for cloning elsewhere.
- Qwen3 VoiceDesign is not a substitute for an accent. Irish, Scottish and Manchester briefs all came out American. An accent has to come from a real recording.
- Voices already picked by ear are in
assets/voices/library-shortlist.json. Start there.
Brand voices
A brand's voice.json can hold several narrators (voices map plus default), one per video:
voiceover.py --brand video-toolkit --voice bren. The brand's model, settings (stability,
similarityBoost, style, speed) and maxWpm apply wherever the command line left its default. See
docs/creating-brands.md.
For full API docs, see reference.md.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementation | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |
| using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |
| verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |