ComfyUI-ElevenLabs-Pro

Full-featured ElevenLabs API integration for ComfyUI. 26 nodes: TTS, Speech-to-Speech, Sound Effects, Music, Voice Cloning / Design, Multi-speaker Dialogue, Speech-to-Text, Audio Isolation — plus 11 pure-Python audio utility nodes (subtitle export, concat, normalize, trim, channels, save, cost estimator, …).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-ElevenLabs-Pro.git
pip install -r ComfyUI-ElevenLabs-Pro/requirements.txt

Then restart ComfyUI.

API key

Set the ELEVENLABS_API_KEY environment variable, or paste your key into the ElevenLabs Pro - API Key node. Keys are scrubbed from any raised error message so they never end up in ComfyUI logs or saved workflow PNG metadata.

Nodes (26 total)

Voice — 6 nodes

| Node | Purpose | |------|---------| | Voice Selector | Pick a predefined voice OR pass a custom voice_id. | | Fetch Voices | List your account's voices (premade / cloned / generated / professional) with search + gender filter. Cached 5 min. | | Get Voice By Name | Look up a voice_id by name (exact / contains / starts-with). | | Voice Clone | Instant voice clone from up to 8 audio samples. create=False by default — flip to True to actually upload (prevents accidental slot/credit burn on workflow re-queue). | | Voice Design | Generate 3 voice previews from a text description. Returns all 3 IDs; pick one with preview_index. | | Voice Create | Save a designed preview as a permanent library voice. Also gated by create=True. |

TTS — 4 nodes

| Node | Purpose | |------|---------| | Text to Speech | Full TTS with every API parameter (stability, similarity, style, speed, speaker boost, language, normalization, seed, previous/next text, PVC-as-IVC, enable_logging, pronunciation dicts). Auto-validates eleven_v3 normalization, turbo language requirement. | | TTS with Timestamps | TTS + word/character-level alignment JSON. | | Text to Dialogue | Multi-speaker (up to 10) eleven_v3 dialogue. Total text length validated. | | Voice Tag Inserter | Pure-text helper — dropdown of v3 expression tags ([whispers], [laughs], [excited], …) injected into your prompt. |

Audio (API) — 4 nodes

| Node | Purpose | |------|---------| | Speech to Speech | Transform an AUDIO input to another voice. | | Sound Effects | Text → SFX with duration / prompt_influence / loop / seed. Uses correct eleven_text_to_sound_v2 model id. | | Audio Isolation | Remove background noise / isolate vocals. | | Speech to Text | Transcribe AUDIO via scribe_v2 or scribe_v1, with diarization, audio-event tagging, keyterms, no-verbatim, temperature, seed. |

Music — 1 node

| Node | Purpose | |------|---------| | Music Generation | Text-prompt OR composition_plan JSON. 3 – 600 s. Force instrumental, respect section durations, C2PA signing. |

Audio (pure-Python utilities) — 7 nodes

| Node | Purpose | |------|---------| | Subtitle Export (SRT/VTT) | Convert TTSTimestamps alignment or STT words into SRT or VTT. Cue grouping by char count + seconds. | | Audio Concat | Stitch up to 8 clips. Resamples mismatched SRs (torchaudio or linear-interp fallback). Broadcasts mono → stereo. Optional silence gap. | | Audio Normalize | Peak / RMS normalization to a target dBFS. | | Audio Metadata | Read duration_seconds, sample_rate, channels, samples. | | Audio Channels | to-mono / to-stereo / left-only / right-only / swap. | | Audio Trim | Trim by start / end seconds. | | Save Audio | Write AUDIO to disk under ComfyUI/output as wav / mp3 / flac / ogg. Sanitized filename + subdir (no .. traversal). |

Utilities — 2 nodes

| Node | Purpose | |------|---------| | Cost Estimator | Estimate credit cost from text + model before the API call. | | Voice Settings Preset | Build a voice_settings JSON blob for reuse across workflows. |

Config — 2 nodes

| Node | Purpose | |------|---------| | API Key | Hold the key and pass it to other nodes. password=True masks it in the UI. | | Account Info | Subscription tier + character usage. PII-redacted by default — email, name, billing IDs are stripped before being returned so saved workflows don't leak your identity. |

Models

TTS

• eleven_v3 — flagship, expressive, supports [whispers]/[laughs]/v3 tags (5 000 char limit, auto-forces normalization=off)
• eleven_multilingual_v2 (10 000 chars)
• eleven_flash_v2_5, eleven_turbo_v2_5 (40 000 chars, require explicit language)
• eleven_flash_v2, eleven_turbo_v2 (English-only, 40 000 / 30 000)
• eleven_multilingual_v1, eleven_monolingual_v1 (legacy)

Speech-to-Speech

• eleven_multilingual_sts_v2, eleven_english_sts_v2

Speech-to-Text

• scribe_v2, scribe_v1

SFX

• eleven_text_to_sound_v2

Music

• music_v1

Output formats

All standard ElevenLabs formats supported: mp3 (44.1k @ 192/128/96/64/32, 24k @ 48, 22k @ 32), wav (44.1k / 48k / 32k / 24k / 22k / 16k / 8k), pcm (44.1k / 48k / 32k / 24k / 22k / 16k / 8k), opus (48k @ 192/128/96/64/32).

The PCM path is decoded directly as int16 little-endian mono so it works without an MP3 decoder. WAV / MP3 / Opus use soundfile first, fall back to torchaudio.

Security

This v2.1 release went through a top-to-bottom audit. Highlights:

• API key never leaks. scrub_secrets() removes the key from every raised exception, error message, and the IS_CHANGED hash (so the key never enters workflow PNG metadata).
• AccountInfo PII redacted. Email, name, billing IDs stripped before returning. Uses /v1/user/subscription by default (no PII).
• VoiceClone / VoiceCreate require explicit create=True. v2.0 burned a voice slot on every queue prompt; v2.1 default is dry-run.
• All paid API nodes cache by input hash. v2.0 used IS_CHANGED = NaN which forced every queue to re-bill the API.
• Download size cap (500 MB default). No more OOM from a malicious endpoint.
• Multipart uploads survive retries. File pointers reset to 0 between attempts (v2.0 sent empty bodies on retry).
• SaveAudio rejects path traversal. Subdirectory and filename are sanitized so a malicious workflow can't write ../../etc/passwd.

Tests

cd ComfyUI/custom_nodes/ComfyUI-ElevenLabs-Pro
python -m pytest tests/ -c pytest.ini

101 / 101 passing. All HTTP mocked — no real ElevenLabs credits consumed during testing.

License

MIT.