ElevenLabs audio API in Nebula — text-to-speech, sound-effects-from-text, speech-to-speech voice changing, audio isolation (denoise), translated dubbing, and speech-to-text transcription, all as audio-gen nodes. Activate when the user configures any elevenlabs-tts, elevenlabs-sfx, elevenlabs-sts, elevenlabs-isolation, elevenlabs-dubbing, or elevenlabs-stt node, or asks about ElevenLabs in Nebula. Sourced from the official ElevenLabs API reference (elevenlabs.io/docs/api-reference) cross-checked against backend/handlers/elevenlabs.py + backend/data/node_definitions.json, and the Nebula audit guide docs/api-guides/elevenlabs.md, on 2026-06-05.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
ElevenLabs audio API in Nebula — text-to-speech, sound-effects-from-text, speech-to-speech voice changing, audio isolation (denoise), translated dubbing, and speech-to-text transcription, all as audio-gen nodes. Activate when the user configures any elevenlabs-tts, elevenlabs-sfx, elevenlabs-sts, elevenlabs-isolation, elevenlabs-dubbing, or elevenlabs-stt node, or asks about ElevenLabs in Nebula. Sourced from the official ElevenLabs API reference (elevenlabs.io/docs/api-reference) cross-checked against backend/handlers/elevenlabs.py + backend/data/node_definitions.json, and the Nebula audit guide docs/api-guides/elevenlabs.md, on 2026-06-05.
ElevenLabs Skill
ElevenLabs is the audio studio inside Nebula: turn text into lifelike speech, generate sound effects from a description, swap one voice for another, strip background noise out of a recording, dub a clip into another language, and transcribe speech to text — six drag-and-drop nodes, all in the audio-gen category.
When to use
User configures an elevenlabs-tts, elevenlabs-sfx, elevenlabs-sts, elevenlabs-isolation, elevenlabs-dubbing, or elevenlabs-stt node
User wants text-to-speech / narration / voiceover, sound effects from a prompt, re-voicing a recording, denoising/isolating speech, translated dubbing, or transcription (speech → text or subtitles)
User asks which ElevenLabs node to use, what params are valid, how to set a voice, or how to chain audio nodes
User asks for music, multi-speaker dialogue, voice design, or voice cloning — so you can tell them those are not Nebula nodes (see Capability boundaries)
Universal rules
Auth header + env var. Every node authenticates with the xi-api-key: <key> HTTP header (note: xi-api-key, notAuthorization: Bearer). The backend reads the key from the ELEVENLABS_API_KEY environment variable. If it is missing, the node fails with ELEVENLABS_API_KEY is required. Get the key from the ElevenLabs dashboard (Profile → API Keys).
Execution pattern = sync (request/response) for all six nodes. From the graph's perspective every node is executionPattern: "sync" — you wire it, run it, and it returns one output (an audio file, or text for STT). There is no streaming or WebSocket node. Caveat: elevenlabs-dubbing is sync to the graph but the handler internally does submit → poll the job every 5s (up to ~10 min) → download the result, so a dubbing node can legitimately take minutes before its audio output appears. The other five return in seconds.
Request encoding differs by node. TTS and SFX send a JSON body (Content-Type: application/json). STS, isolation, dubbing, and STT send multipart/form-data with the source audio file attached (STS/isolation under form field audio; dubbing and STT under form field file). You don't set this in Nebula — the handler picks the right encoding — but it explains why audio-input nodes need a real file on disk, not a URL string.
Status / error codes. A non-200 from ElevenLabs surfaces as ElevenLabs API error <status>: <body> (or the per-node variant, e.g. ElevenLabs SFX error …, ElevenLabs Dubbing submit error …). Common cases: 401 bad/missing key, 422 invalid params (bad voice_id, out-of-range value, unsupported output_format), 429 rate limit / quota, 400 malformed request. Dubbing-specific terminal states: the job can return status: "failed" (→ ElevenLabs Dubbing failed: <error>) or never finish (→ ElevenLabs Dubbing timed out after the poll cap).
Input-URI / input-file rules. TTS and SFX take Text input (a string — wire a Text node or any Text output into the text port). STS, isolation, dubbing, and STT take Audio input: the upstream audio value must be a local file path that exists on disk (the handler does Path(value).read_bytes() and raises Audio file not found: <path> otherwise). Feed them the audio output of another audio node (TTS, SFX, isolation) or an audio upload node — not a bare remote URL. STT is the one node that goes the other way: Audio in → Text out (its text output is a string you can wire into text-driven nodes), whereas the rest output Audio.
Key gotchas.
PCM is saved as .pcm. When you pick a pcm_*output_format (TTS/SFX/STS), ElevenLabs returns raw PCM samples with no WAV header, and the handler saves them with a .pcm extension to keep the extension honest. Many players/preview surfaces can't open a headerless .pcm. Prefer an mp3_* format unless the user specifically needs raw PCM for their own pipeline.
Isolation and dubbing always output MP3, regardless of any format setting — isolation has no params at all, and dubbing's download is hard-coded to MP3.
voice_id is free text, not a dropdown. Nebula cannot list or search voices. Default is Rachel (21m00Tcm4TlvDq8ikWAM). To use another voice, paste its ElevenLabs voice ID (from the Voice Library / your Voices page) into the field. An invalid ID fails at request time with a 422/404.
style is only sent when > 0 (TTS and STS). Leaving it at the default 0 omits it from voice_settings.
seed is only sent when set (TTS and STS). Blank = random each run; set an integer for reproducibility.
SFX model_id is fixed. The node exposes no model selector; the handler hard-codes the sound-effects model (eleven_text_to_sound_v2). Don't tell the user they can choose an SFX model.
Pick the right node
One row per real Nebula node. IDs, ports, endpoints, and params all match node_definitions.json + backend/handlers/elevenlabs.py.
Nebula node (display name)
Node ID
Endpoint / model
In → Out ports
Key params
ElevenLabs TTS
elevenlabs-tts
POST /v1/text-to-speech/{voice_id} · model from model_id enum
Types, ranges, enums, and defaults below are exactly what the node exposes. "(handler)" notes describe behavior in backend/handlers/elevenlabs.py the user can't see in the panel.
elevenlabs-tts (Text → Audio)
voice_id — string. Default 21m00Tcm4TlvDq8ikWAM (Rachel). Free-text ElevenLabs voice ID. Goes into the URL path.
(handler) style and use_speaker_boost may be forwarded inside voice_settings if present on the node, but the STS node panel does not expose style/use_speaker_boost controls — treat them as TTS-only knobs.
elevenlabs-isolation (Audio → Isolated Audio)
No params. Submit audio, get back denoised MP3. Output port labeled "Isolated Audio".
elevenlabs-dubbing (Audio → Dubbed Audio)
target_lang — enum, required, default es. Options: enesfrdeitptjakozharruhinltrplsv. The download endpoint uses this code, so it must be one the dubbing job produced.
source_lang — enum, default auto (Auto-detect). Options: autoenesfrdejazh. (handler) only sent when not auto.
num_speakers — integer 0–10, optional (placeholder "Auto-detect"). (handler) only sent when > 0; 0/blank = auto.
drop_background_audio — boolean, default false. (handler) only sent when true. Drops music/ambience, keeps speech.
disable_voice_cloning — boolean, default false. Default (off) preserves the original speaker's voice in the dub; turn on to use a generic voice instead. (handler) only sent when true.
elevenlabs-stt (Audio → Text)
model_id — enum. Default scribe_v1. Options: scribe_v1 (Scribe v1) · scribe_v2 (Scribe v2). Both verified working against the live API.
language_code — string, optional. Blank = auto-detect the spoken language. Set an ISO language code to force a specific language.
diarize — boolean, default false. Labels which speaker said what.
num_speakers — integer 1–32, optional. Hint for how many distinct speakers to expect; leave blank to let the model decide.
transcript_format — enum. Default text. Options: text (plain transcript) · srt · vtt (subtitle files). (handler) plain text returns the transcript string in the text output port; srt/vtt return the subtitle content instead.
(handler gotcha) requesting srt or vttforces diarize=true + timestamps_granularity=word automatically, because the ElevenLabs API returns HTTP 400 ("Requesting additional formats must have diarization and timestamps enabled") otherwise. You don't set this — the handler does — but it means a subtitle export is always diarized + word-timestamped regardless of how you left diarize. (Caught via live smoke 2026-06-05.)
Recipes
1. Script → narrated voiceover (single node).
Add a Text node (or any Text output) with the script.
Wire it into elevenlabs-ttstext port.
Set voice_id; pick model_id (eleven_multilingual_v2 for quality, eleven_flash_v2_5 for speed); tune stability / style / speed.
Run → the audio output is the narration MP3.
2. Narration + matching ambience (branch into two nodes).
Branch a Text/prompt into:
elevenlabs-tts for the spoken line, and
elevenlabs-sfx with a prompt like "soft rainstorm with distant thunder", loop on, duration_seconds ~15 for a bed.
Take both audio outputs downstream into a mix/composite step to lay voice over ambience.
3. Clean up, then re-voice or translate a clip.
Start from an audio upload node or an elevenlabs-ttsaudio output.
If the source is noisy, feed it through elevenlabs-isolation first (audio → "Isolated Audio").
Then either:
same language, new voice: into elevenlabs-sts — set the target voice_id; optionally enable remove_background_noise (redundant if you already isolated).
new language, same speaker: into elevenlabs-dubbing — set target_lang (e.g. ja), leave source_lang on auto, keep disable_voice_cloningoff so the original voice is preserved.
Start from an audio upload node, or the audio output of any audio node (TTS, SFX, isolation, dubbing).
Wire it into elevenlabs-stt; leave transcript_format on text for a plain transcript.
Wire the node's text output into any text-driven node — e.g. a chat/LLM node (claude-chat, gpt-4o-chat) to summarize/translate, or Combine Text to assemble a script.
For subtitles instead: set transcript_format=srt (or vtt) — the text output then carries the subtitle content (the handler auto-enables diarization + word timestamps for these formats; see the param note).
In the nebula_nodes context
Node IDs:elevenlabs-tts, elevenlabs-sfx, elevenlabs-sts, elevenlabs-isolation, elevenlabs-dubbing, elevenlabs-stt. All category: "audio-gen", all apiProvider: "elevenlabs", all envKeyName: "ELEVENLABS_API_KEY".
Handler file:backend/handlers/elevenlabs.py — one handler per node (handle_elevenlabs_tts, _sfx, _sts, _isolation, _dubbing, handle_elevenlabs_stt). Shared _save_audio / _audio_extension helpers map output_format → file extension (mp3→.mp3, pcm→.pcm, wav→.wav, default .mp3) for the audio-output nodes; handle_elevenlabs_stt does not call _save_audio — it returns a text string, not a file.
Ports:
elevenlabs-tts: in text (Text, required) → out audio (Audio).
elevenlabs-sfx: in text (Text, required) → out audio (Audio).
elevenlabs-sts: in audio (Audio, required) → out audio (Audio).
elevenlabs-isolation: in audio (Audio, required) → out audio (Isolated Audio).
elevenlabs-dubbing: in audio (Audio, required) → out audio (Dubbed Audio).
elevenlabs-stt: in audio (Audio, required) → out text (Text).
Chaining rules: Text-input nodes (tts, sfx) accept any Text output. Audio-input nodes (sts, isolation, dubbing) accept any Audio output, but the upstream value must resolve to a local file path that exists — chain them off another ElevenLabs audio node or an audio upload, not a raw URL. Output of one audio node → input of the next is the normal way to build isolation → sts/dubbing and tts → sts chains.
How outputs render: each node writes its bytes to the run directory (get_run_dir()) as <12-hex>.<ext> and returns {"audio": {"type": "Audio", "value": "<path>"}}. The frontend serves/plays that file as an audio result. .mp3 plays everywhere; a .pcm (raw, headerless) generally won't — another reason to prefer MP3.
Dubbing internals (why it's slow): submit returns a dubbing_id; the handler polls GET /v1/dubbing/{id} every 5s up to 120 times (~10 min), tolerating up to 5 consecutive poll failures, then downloads GET /v1/dubbing/{id}/audio/{target_lang} and saves MP3. So a single dubbing node may run for minutes — that's expected, not a hang.
Capability boundaries — what ElevenLabs offers but Nebula does NOT expose
Never promise these via a node; there is no node for any of them (per the audit gap table in docs/api-guides/elevenlabs.md):
Music generation (/v1/music/compose, stems, video-to-music) — prompt-to-music and stem separation. No node.
Text-to-Dialogue (/v1/text-to-dialogue) — multi-speaker scenes in one call. No node; elevenlabs-tts is single-voice only.
Voice Design (/v1/text-to-voice/design|create|remix) — invent a brand-new voice from a description. No node.
Voice Cloning (/v1/voices/add IVC, /add/professional PVC) — clone a voice from samples. No node.
Voice management / library (GET /v1/voices, /shared, /similar) — Nebula takes a voice_id string but cannot list, search, or pick voices in-app. The user must supply the ID.
Models list (GET /v1/models) — model choices are hard-coded enums, not fetched live; SFX has no model selector at all.
Dubbing extras — the node submits/polls/downloads only. It does not expose transcripts, per-segment editing, speaker-by-speaker rendering, or watch/CSV-script dubbing.
Forced Alignment, Pronunciation Dictionaries, History, Conversational AI/Agents, Studio/Projects, Audio Native, and all streaming/WebSocket variants — none are nodes.
TTS output is MP3 or raw PCM only — no WAV/OGG/Opus container choice from the node; raw PCM lands as .pcm.
Roughly ~33% of the ElevenLabs API surface is wired into Nebula (6 nodes). If the user wants anything in the list above, the honest answer is "not a Nebula node — you'd call the ElevenLabs API directly."
