Skip to content

Global-AI-Group-Inc/content-generation-skills / chatgpt/overrides/skills/mira-audio

mira-audio

Verified

Voice, music and sound on Mira AI through its MCP server: text to speech with library voices and emotion tags in 70+ languages, multi-voice dialogue, music timed to the cut, a soundtrack the model writes by watching the clip, sound effects, voiceover and captions on a finished clip, transcripts, clean audio and stems. Use it for "voiceover", "narration", "text to speech", "TTS", "dialogue", "podcast", "music", "soundtrack", "jingle", "beat", "sound effect", "SFX", "captions", "subtitles", "transcript", "clean audio", "noise removal", "stems". NOT for the picture itself (mira-generate).

Mira audio

Every sound on Mira comes from ElevenLabs models behind the same MCP server that makes the pictures: Eleven v4 and v4 Turbo for speech, text-to-dialogue, Music v2.5, the sound-effects model, the voice isolator, and Scribe v2 for transcripts. A tool that makes something returns a generation id at once and wait_for_generation brings the result: a track is an mp3 in urls[0], an operation on a clip is a new clip, and stems, captions and transcripts arrive as files in files[], each with a role.

Pick the tool

The user wantsCallRead
one voice reading a text: a voiceover track, narration, an ad readgenerate_audio kind speechspeech.md
two or more people talking: a podcast, a sketch, an interviewgenerate_audio kind dialoguedialogue.md
music with a structure, lyrics or hits on given secondsplan_music → generate_audio kind musicmusic.md
a sound effect or an ambience bedgenerate_audio kind sfxsfx.md
the sounds of what happens on screen, for a silent clipadd_sound_effectssfx.md
music for a clip that already existsadd_soundtrackmix.md
a voice over a finished clip, captions optionalvoiceover_videomix.md
a library voice for the job: language, gender, age, accent, stylelist_voicesvoices.md
speech without the noise, music and echo around itisolate_voicemix.md
vocals and instruments on separate tracksseparate_stemsmix.md
subtitles burned in, plus an SRTadd_captionsdub-captions.md
a transcript, a silence cut, short clips, chapterstranscribe_video → analyze_transcriptdub-captions.md

One file per job. Do not load them all for one track.

Sources

An operation on existing media takes exactly one of two ids: sourceGenerationId for anything the account made on Mira (find it with list_generations, kind video or audio), or libraryItemId for a file the user attached to the chat.

ToolTakesLongest source
add_soundtracka clip600 s
voiceover_videoa clip300 s
add_captionsa clip30 min
isolate_voicea clip or a track30 min
transcribe_videoa clip or a track30 min
separate_stemsa track or a clip600 s

A longer source is refused (source_too_long) before a credit is spent: trim it first.

Credits

Everything that makes or changes sound spends credits; list_voices and plan_music do not. Never quote a price from memory: rates change and estimate_cost is the only source. Speech and dialogue bill by characters, music, soundtracks and stems by length, the operations on a clip by the source's seconds or minutes.

estimate_cost {"kind": "audio", "model": "speech", "chars": 740}
estimate_cost {"kind": "audio", "model": "music", "durationSeconds": 60}
estimate_cost {"kind": "soundtrack", "sourceSeconds": 15}
estimate_cost {"kind": "stems", "model": "6", "sourceSeconds": 180}

Workflow

  1. Name the job from the table. Say in one sentence what you will make and that it spends credits.
  2. When the user cares who speaks, call list_voices with filters and offer two or three voices by name with their preview links. Otherwise leave voiceId empty for the default voice.
  3. Write for the medium: text for the ear (speech.md), genre, tempo and instruments for music (music.md), a source in a space for an effect (sfx.md).
  4. Quote with estimate_cost before anything long: minutes of music, pages of text, a long voiceover.
  5. Call, then wait_for_generation until it finishes. Show the link. Do not narrate polling.
  6. Iterate by changing one thing: one tag, one section length, one style, one voice.

Common chains

  • Ad with a voice and music: generate_video → voiceover_video (original duck, captions on) → add_soundtrack with mix under, so the music dips under the new voice.
  • Clean up a creator clip: isolate_voice if the room is noisy → add_captions for viewers watching without sound.
  • Cut the picture to the music: plan_music → generate_audio music → the track's URL in referenceAudioUrls of generate_video on seedance-2-5, or into the user's editor.
  • Long talk to shorts: transcribe_video → analyze_transcript task clipper → cut the ranges in an editor → add_captions style bold on each short.

Rules

  • Speech uses the platform's library voices only. Never imitate a real person: no celebrity voices, no "sounds like" a named actor or singer. There is no voice cloning or voice design here; never offer it as a step.
  • Music prompts name no artists, bands or songs and quote no existing lyrics: the provider refuses them (music_prompt_rejected) and suggests a rewrite. Describe the sound instead.
  • Speech text is written in the language it will be spoken, lyrics in the language they will be sung, music and effect prompts in English. Parameter values stay English. Reply in the user's language.
  • Never show generation, voice or design ids to the user; show names and links.

When it fails

CodeWhat to do
text_too_longSplit at paragraph ends: v4 takes 10000 characters a call, turbo 20000, a dialogue 2000.
source_too_longThe source is over the tool's limit in the Sources table: trim it and send the part.
music_prompt_rejectedAn artist, a song or known lyrics were recognised. Take the suggested prompt or describe the sound.
composition_invalidA section is outside 3-120 s or over 30 lines, or there are more than 30 sections.
voice_not_foundThe voice is not available to this account: list_voices again.
text_moderatedThe text was refused. Rewrite it; do not resend it as is.
elevenlabs_disabledAudio is switched off on the platform at the moment. Say so and stop.
any other *_failedA failed job returns its credits. Retry once, then report the code in plain words.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall