djalmajr Create audio
Produce a same-person spoken track from a script. Output is WAV/MP3. Picture is /create-video.
Initial context via slash: $ARGUMENTS
Parse --type= from $ARGUMENTS when present.
Project root
<project-root>/tmp/audio/<module>/. Do not commit large WAV/MP3 or API keys. Commit lines.json + DEFAULT.md.
Prompting
Ask these before generating, unless $ARGUMENTS already answers them.
| Decision point | Why structured | Suggested options |
|---|---|---|
| Type | Changes line length and deliverable | tutorial (short lines for a later video) · podcast (longer takes, one episode file) · voiceover (same as tutorial, named for mux) |
| How to create the voice | Changes stack, license, and cost | VoxCPM2 (local clone of an approved ref) · CosyVoice3 (local clone) · Grok / xAI Voice · OpenAI TTS · Other hosted · macOS say (timing only) |
| Identity | When local clone / design | Existing approved ref.wav · New CustomVoice take · Hunt with VoiceDesign then lock |
Free-form: language, tone, script, proof sentence, module name.
No-pause defaults: --type=tutorial, ask is skipped only if the user already named an engine; otherwise do not invent a hosted key or start a multi-GB download.
When to use
- Voiceover lines for
/create-video --type=screencast - Audio-only lesson
- Podcast / spoken newsletter
- Voice A/B for a brand
When not to use
- Recording a screen —
/create-video - Voice-in-product (user talks to the app)
Hard rules
- Ask how the audio will be created. Do not silently pick an engine.
- One identity for the whole set. Rank one locked speaker × one proof sentence. Never rank VoiceDesign folders of independent takes.
- Approve the reference first when cloning. A thin/nasal/"duck" take is almost always a bad ref, not a bad engine — the clone copies whatever you feed it.
- If the ref changes, regenerate every line.
- License before you like the take. Local defaults (Qwen3-TTS, VoxCPM2) are Apache-2.0. Skip research/non-commercial weights for anything that might ship. Hosted voices follow that vendor's ToS.
- macOS
saytimes a line. It is not a product voice.
Details: references/engines.md, references/gotchas.md.
Types
--type |
Script | Deliverable |
|---|---|---|
tutorial / voiceover |
Short lines, one idea each | default-vo/<id>.wav (+ optional concat) |
podcast |
Longer paragraphs, conversation allowed | episode.wav / episode.mp3 via templates/assemble-audio.mjs |
Same identity rules for every type.
Process
1. Type + engine
Resolve --type. Then ask how (table above). Record the choice in DEFAULT.md.
2. Script
Save templates/lines.json. proofId is the A/B line.
3. Generate
Follow the chosen engine in references/engines.md.
Local (the two engines that won the ear test: VoxCPM2 and CosyVoice3):
- Start from an approved
ref.wav(human take, or a Qwen3 CustomVoice take you actually liked — Qwen is a ref source, not the shipped engine). - Clone every line with VoxCPM2 (templates/generate-vo.py) or CosyVoice3.
- Listen: ref vs clone of the proof line, then the full set.
Do not blame VoxCPM/CosyVoice for a duck/nasal set if the ref already sounds like that.
Grok / OpenAI / other hosted: one voice id for the whole set, key only in env, still listen to the proof line before rendering everything.
4. Deliver
- Video later: point
/create-videoatdefault-vo/. - Podcast / aula:
node assemble-audio.mjs(optionalOUT_FORMAT=mp3).
Verification
node skills/create-audio/scripts/smoke.mjs
Then listen. Do not ship on spectrograms.
Chain
/create-audio --type=tutorial → /create-video --type=screencast.