| # Prompt-Pirate/comfytts-models — Parrot Engine weights mirror |
|
|
| Generated file — do not hand-edit. The source of truth is |
| `functions/engine-distribution/engines.json` in the Parrot Engine repository; |
| this file is produced by `build-artifacts.js` from that data plus |
| `files.lock.json` (sha256 + byte size per mirrored file). |
|
|
| ## Policy |
|
|
| Gated weights are never re-hosted here, even where the licence would technically |
| permit it — gated engines carry instructions for accepting the upstream gate |
| instead. |
|
|
| ## 1. Mirrored here |
|
|
| These engines' weights are re-hosted in this repository under `engines/<kind>/`, |
| with every file sha256- and byte-size-verified against `files.lock.json` before |
| publish (see `MANIFEST.json` for the per-file hashes). |
|
|
| | Engine | Dest | Licence | Verified | Origin | |
| |---|---|---|---|---| |
| | `chatterbox` | comfyui | MIT | HF API cardData ResembleAI/chatterbox, fetched 2026-08-18 | ResembleAI/chatterbox | |
| | `cosyvoice` | comfyui | Apache-2.0 | HF API cardData FunAudioLLM/Fun-CosyVoice3-0.5B-2512, fetched 2026-08-18 | FunAudioLLM/Fun-CosyVoice3-0.5B-2512 | |
| | `dia2` | comfyui | Apache-2.0 | HF API cardData nari-labs/Dia2-2B, fetched 2026-08-18 — identified by searching HF for author=nari-labs (the manifest itself does not name the weights repo, only the GitHub pack) | lord-lethris/ComfyUI-lethris-dia2 (pack) + nari-labs/Dia2-2B (weights, identified via HF search this task — the manifest names only the GitHub pack repo) | |
| | `dots` | comfyui | Apache-2.0 | HF API cardData, all three repos (dots-studio/dots.tts-base, -mf, -soar), fetched 2026-08-18 | dots-studio/dots.tts-base (+ dots.tts-mf, dots.tts-soar) | |
| | `firered2` | comfyui | Apache-2.0 | FireRedTeam/FireRedTTS2 cardData (firered2-clone-v1 manifest notes) | FireRedTeam/FireRedTTS2 | |
| | `gptsovits` | runtime | MIT | TTS ENGINE REPORT.md §11: 'MIT / MIT' | GPT-SoVITS official v2ProPlus prezipped package (RVC-Boss/GPT-SoVITS) | |
| | `kitten` | comfyui | Apache-2.0 | HF API cardData KittenML/kitten-tts-mini-0.8, fetched 2026-08-18 | KittenML/kitten-tts-mini-0.8 | |
| | `kokoro` | pack | Apache-2.0 | HF cardData lineage via onnx-community, recorded in kokoro-tts-v1 manifest notes (v25.34.0) | onnx-community/Kokoro-82M-v1.0-ONNX | |
| | `longcat` | comfyui | MIT | longcat-clone-v1 manifest: 'Pack and weights MIT' | meituan-longcat/LongCat-Audio-Codec (AudioDiT) + google/umt5-base | |
| | `moss` | comfyui | Apache-2.0 | HF API cardData, all four repos (MOSS-TTS, MOSS-Audio-Tokenizer, MOSS-TTS-Local-Transformer, MOSS-TTSD-v1.0), fetched 2026-08-18 | OpenMOSS-Team/MOSS-TTS + OpenMOSS-Team/MOSS-Audio-Tokenizer + OpenMOSS-Team/MOSS-TTS-Local-Transformer + OpenMOSS-Team/MOSS-TTSD-v1.0 | |
| | `mossnano` | comfyui | Apache-2.0 | 'Apache 2.0 code AND weights' — mossnano-clone-v1/mossnano-preset-v1 manifest notes + parrot-tts-nodes VENDORED-README.md line 23 | OpenMOSS-Team/MOSS-TTS-Nano-100M-ONNX + OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano-ONNX | |
| | `neutts` | hub | Apache-2.0 | SDK + Air weights cardData (v25.53.0 addendum); Nano sibling EXCLUDED (revenue-capped licence, never shipped) | neuphonic/neutts-air-q8-gguf + neuphonic/neucodec | |
| | `pocket-preset` | hub | CC-BY-4.0 | cardData both repos (v25.42.0 addendum); ATTRIBUTION required in README | kyutai/pocket-tts-without-voice-cloning | |
| | `qwen3` | comfyui | Apache-2.0 | HF API cardData, all five present-on-disk repos (1.7B-Base, 1.7B-CustomVoice, 1.7B-VoiceDesign, 0.6B-Base, 0.6B-CustomVoice) plus Qwen3-TTS-Tokenizer-12Hz, fetched 2026-08-18 | Qwen/Qwen3-TTS-12Hz-1.7B-Base (+ -CustomVoice, -VoiceDesign, -0.6B variants, Qwen3-TTS-Tokenizer-12Hz) | |
| | `soprano` | comfyui | Apache-2.0 | wheel LICENSE + HF cardData (v25.52.0 addendum) | ekwek/Soprano-1.1-80M | |
| | `vibevoice` | comfyui | MIT | HF API cardData, all three repos (microsoft/VibeVoice-1.5B, aoi-ot/VibeVoice-Large, FranckyB/VibeVoice-Large), fetched 2026-08-18 | microsoft/VibeVoice-1.5B (+ aoi-ot/VibeVoice-Large, FranckyB/VibeVoice-Large) | |
| | `voxcpm` | comfyui | Apache-2.0 | voxcpm2-clone-v1 manifest notes: 'Weights Apache-2.0, ungated' | OpenBMB/VoxCPM2 | |
| | `voxtream` | hub | CC-BY-4.0 | HF API cardData herimor/voxtream2, fetched 2026-08-18 | herimor/voxtream2 | |
| | `zonos2` | comfyui | Apache-2.0 | cardData both repos; zonos2-clone-v1 manifest README section Licence | Zyphra/Zonos-v0.1-hybrid + Zyphra/Zonos-v0.1-speaker-embedding | |
|
|
| > **Attribution required.** |
| > `pocket-preset` is licensed **CC-BY-4.0** by kyutai/pocket-tts-without-voice-cloning (https://huggingface.co/kyutai/pocket-tts-without-voice-cloning). If you redistribute or build on these files, credit the original authors and link back to the source repository. |
| > `voxtream` is licensed **CC-BY-4.0** by herimor/voxtream2 (https://huggingface.co/herimor/voxtream2). If you redistribute or build on these files, credit the original authors and link back to the source repository. |
|
|
| ## 2. Not mirrored — origin and gated engines |
|
|
| ### 2a. Gated upstream — never re-hosted (policy) |
|
|
| - **`orpheus`** — gate: `canopylabs/orpheus-3b-0.1-ft` (HF-gated (fine-grained token, 'read gated repos' permission); gate-state discrepancy RESOLVED 2026-08-18, see notes). Licence: Apache-2.0 — gate canopylabs/orpheus-3b-0.1-ft; gate-state discrepancy resolved (Task 6) — CLAUDE.md was accurate, the workflow manifest note was stale. Accept the gate at https://huggingface.co/canopylabs/orpheus-3b-0.1-ft before use. See `engines/orpheus/README.md`. |
| - **`pocket-clone`** — gate: `kyutai/pocket-tts` (HF-gated repo — the clone half of PocketTTS; POLICY: never re-hosted, the owner's own accepted gate is required per machine). Licence: CC-BY-4.0 — gate kyutai/pocket-tts — POLICY: never re-hosted. Accept the gate at https://huggingface.co/kyutai/pocket-tts before use. See `engines/pocket-clone/README.md`. |
| - **`sesame`** — gate: `sesame/csm-1b` (HF-gated (fine-grained token, 'read gated repos' permission)). Licence: Apache-2.0 — Apache-2.0 per HF card. Accept the gate at https://huggingface.co/sesame/csm-1b before use. See `engines/sesame/README.md`. |
|
|
| ### 2b. Un-hostable — known restrictive licence |
|
|
| - **`dramabox`** — LTX-2 Community (conditional) — dramabox-clone-v1 manifest notes — check terms before any commercial use. See `engines/dramabox/README.md`. |
| - **`echo`** — CC-BY-NC-SA-4.0 — HF API cardData, both repos (jordand/echo-tts-base, jordand/fish-s1-dac-min), fetched 2026-08-18. See `engines/echo/README.md`. |
| - **`f5tts`** — CC-BY-NC-4.0 — HF API cardData SWivid/F5-TTS, fetched 2026-08-18. See `engines/f5tts/README.md`. |
| - **`fish2`** — fish-audio-research-license (custom, non-commercial-only; dependency fish-s1-dac-min is CC-BY-NC-SA-4.0) — HF API cardData all three repos, fetched 2026-08-18. See `engines/fish2/README.md`. |
| - **`higgs-v2`** — BOSON HIGGS AUDIO 2 COMMUNITY LICENSE AGREEMENT (custom, Llama-3-derivative, commercial use above 100k MAU requires a separate agreement) — raw LICENSE file at huggingface.co/bosonai/higgs-audio-v2-generation-3B-base/raw/main/LICENSE, fetched 2026-08-18 (cardData license field itself just says "other"). See `engines/higgs-v2/README.md`. |
| - **`higgs-v3`** — BOSON HIGGS TTS 3 RESEARCH AND NON-COMMERCIAL LICENSE AGREEMENT — local LICENSE file read directly, 2026-08-18, at models\comfyui\TTS\higgs_audio_v3\higgs-audio-v3-tts-4b\LICENSE (ships on disk with the weights). See `engines/higgs-v3/README.md`. |
| - **`indextts2`** — bilibili Model Use License Agreement (custom, conditional — revenue/user thresholds require a separate licence, high-risk-deployment prohibitions, indemnification terms) — GitHub raw LICENSE at raw.githubusercontent.com/index-tts/index-tts/main/LICENSE, fetched 2026-08-18 (the HF model card itself carries NO license field or file — only the linked GitHub repo, whose own api.github.com license probe reports SPDX NOASSERTION, has the actual terms). See `engines/indextts2/README.md`. |
| - **`minimax`** — MiniMax-H3 Community License (TERRITORY-RESTRICTED) — the audio-cpp/audio.cpp-gguf card row for MiniMax-H3-Q4-GGUF, fetched 2026-08-26; OWNER RULING 2026-08-26 ('minimax_h3 - yes, install it'): accepted for LOCAL use only. See `engines/minimax/README.md`. |
| - **`omnivoice`** — CC-BY-NC (code is separately Apache-2.0, but the pre-trained weights are not) — k2-fsa/OmniVoice model card prose, fetched 2026-08-18: 'Our code is released under the Apache 2.0 License. The pre-trained model is licensed under the CC-BY-NC due to constraints from its training data (e.g., Emilia).' — no cardData.license field exists on the HF repo, this is prose-only. See `engines/omnivoice/README.md`. |
| - **`outetts`** — CC-BY-NC-SA-4.0 + Llama 3.2 Community License - NON-COMMERCIAL — HF model cards OuteAI/Llama-OuteTTS-1.0-1B and OuteAI/Llama-OuteTTS-1.0-1B-GGUF license:cc-by-nc-sa-4.0 (2026-08-23); the base model is Llama 3.2 (its Community License applies too); the mirror repo's own card says apache-2.0 - that is the MIRROR's claim, the original cards govern. See `engines/outetts/README.md`. |
| - **`piper`** — PER-VOICE, VARIES — piper voice param description + TTS ENGINE REPORT.md L980-982; runtime piper-tts 1.6.1 is GPL-3.0. See `engines/piper/README.md`. |
| - **`seedvc`** — GPL-3.0 — HF model cards Plachta/Seed-VC and mlx-community/SeedVC-MLX license:gpl-3.0 + the audio-cpp card row (2026-08-23). See `engines/seedvc/README.md`. |
| - **`silero`** — CC BY-NC-SA 4.0 — silero-tts-v1 manifest notes + parrot-tts-nodes README L21 (bold NON-COMMERCIAL). See `engines/silero/README.md`. |
| - **`spark`** — CC-BY-NC-SA 4.0 — relicensed at HF commit b63203d4; pack header (1038lab/Comfyui-Spark-TTS) STALE/WRONG, still claims Apache-2.0. See `engines/spark/README.md`. |
| - **`supertonic`** — BigScience Open RAIL-M License (dated 2022-08-18) — HF API cardData Supertone/supertonic-3 (license: "openrail") + the repo's own raw LICENSE file at huggingface.co/Supertone/supertonic-3/raw/main/LICENSE, both fetched 2026-08-18. See `engines/supertonic/README.md`. |
| - **`vevo2`** — CC-BY-NC-ND-4.0 - NON-COMMERCIAL, NO DERIVATIVES — HF model card RMSnow/Vevo2 license:cc-by-nc-nd-4.0 + the audio-cpp card row (2026-08-23). See `engines/vevo2/README.md`. |
| |
| ### 2c. Permissive licence — not yet mirrored |
| |
| The licence itself is permissive enough to mirror (Apache-2.0/MIT/BSD/CC-BY), but |
| these rows have not graduated to `mirror` yet — usually because the files are not |
| currently staged locally in a clean, mirror-ready form. See each row's reason below. |
| |
| - **`bsroformer`** — Apache-2.0 — the audio-cpp/audio.cpp-gguf card row for BS-RoFormer-ep368-GGUF (2026-08-24). SEPARATION MODEL, NOT A RENDER ROUTE (v25.90.0, the Separation lane): bs-roformer-cpp in the separation registry root audiocpp-adapter/sep/ - 2 stems (vocals + instrumental), the Clean Vocals / Refiner 🧹 audio.cpp choice's 'pass' model. Apache-2.0 - kept source origin like the ASR rows (audio-cpp's own repo is the ungated home). See `engines/bsroformer/README.md`. |
| - **`confucius4`** — Apache-2.0 — HF model card netease-youdao/Confucius4-TTS license:apache-2.0 + audio-cpp/audio.cpp-gguf card row 'Apache-2.0' (2026-08-23). LOCAL USE via the audio.cpp route (v25.83.0, wave 2 of the exclusives lane): confucius4-clone-cpp, a 14-language TRANSCRIPT-FREE cloner (--voice-ref required; --reference-text ignored); en/zh are upstream's validated text frontends. UPSTREAM STATUS: EXPERIMENTAL (the binary's own model_specs). Original-dtype GGUF (the only packaged precision). NOT mirrored by us. See `engines/confucius4/README.md`. |
| - **`glmtts`** — MIT — HF model card zai-org/GLM-TTS license:mit (2026-08-23); the mirror repo's own card says apache-2.0 - that is the MIRROR's claim, the original card governs; the GGUF's provenance is community (pinned by sha). LOCAL USE via the audio.cpp route (v25.84.0, wave 3 of the exclusives lane): glmtts-clone-cpp, a zh/en zero-shot cloner that REQUIRES the reference's exact transcript (the first consumer of the sidecar-or-omitted transcript rule). Mixed Q8_0/F16 GGUF, upstream status community. NOT mirrored by us. See `engines/glmtts/README.md`. |
| - **`htdemucs`** — MIT — the audio-cpp/audio.cpp-gguf card row for HTDemucs-GGUF (2026-08-24); facebookresearch/demucs is MIT. SEPARATION MODEL, NOT A RENDER ROUTE (v25.90.0): htdemucs-cpp - 4 stems (drums/bass/other/vocals), the 🥁 Split Stems tab's model; needs 44.1 kHz input (the app decodes). MIT - kept source origin. See `engines/htdemucs/README.md`. |
| - **`inflect`** — Apache-2.0 — HF model card owensong/Inflect-Micro-v2 license:apache-2.0 + the audio-cpp card row (2026-08-23). RUNTIME DEPENDENCY: eSpeak-ng (GPL-3.0-or-later) - the engine phonemises through an external espeak-ng.dll + espeak-ng-data that the app places beside its models, copied from the NeuTTS 1.4.1 wheel already vendored in parrot-tts-nodes; a separate process dependency of the audio.cpp CLI, never linked into Parrot. LOCAL USE via the audio.cpp route (v25.87.0, the rider): inflect-tts-cpp, ONE synthetic English voice (9.4M VITS, 72 MB). Apache-2.0 weights, community GGUF from audio-cpp's own repo; NOT mirrored (the espeak-ng GPL dependency would ride along with any redistribution - a deliberate exclusion). See `engines/inflect/README.md`. |
| - **`irodori`** — MIT — HF model card Aratako/Irodori-TTS-v4.1-Small license:mit + audio-cpp/audio.cpp-gguf card row 'MIT' (2026-08-23); codec Aratako/Semantic-DACVAE-Japanese-32dim MIT; model card adds non-binding ethical-use notes (no impersonation without consent). LOCAL USE via the audio.cpp route (v25.82.0, wave 1 of the exclusives lane): irodori-tts-cpp / irodori-clone-cpp / irodori-design-cpp all on ONE f16 v4-Small GGUF from audio-cpp's own HF repo. NOT mirrored by us (owner: 'if it's available do it' - local use; the licence law governs redistribution only). See `engines/irodori/README.md`. |
| - **`melroformer`** — MIT — the audio-cpp/audio.cpp-gguf card row for Mel-Band-RoFormer-GGUF (2026-08-24). SEPARATION MODEL, NOT A RENDER ROUTE (v25.90.0): mel-band-roformer-cpp - 2 stems; audio.cpp flags this family's GGUFs as 'drift' (f16 chosen as the closer one); the owner compares it with the venv Mel-Band by ear. MIT - kept source origin. See `engines/melroformer/README.md`. |
| - **`miotts`** — Apache-2.0 (MioTTS-1.7B) + MIT (MioCodec) — HF model cards Aratako/MioTTS-1.7B license:apache-2.0 and Aratako/MioCodec-25Hz-44.1kHz-v2 license:mit + the audio-cpp card rows (2026-08-23). LOCAL USE via the audio.cpp route (v25.82.0, wave 1): miotts-clone-cpp, a pure cloner (--voice-ref required, no presets, no transcript), en/ja, orig precision (the only 'pass' file). NOT mirrored by us. WAVE E (v25.104.0, 2026-08-26): miocodec-vc-cpp - an STS voice changer whose MAIN model is the MioCodec file wave 1 fetched as the miotts COMPANION (MIT per the audio-cpp card row for MioCodec-25Hz-44.1kHz-v2-GGUF). Non-deterministic by measurement (no seed declared); the family's s2s task lands in wave F. No new files. WAVE F (v25.105.0, 2026-08-26): miocodec-s2s-cpp - the s2s sibling of the wave-E voice changer on the same wave-1 file (two conversion algorithms on one codec, A/B by ear). Non-deterministic like its sibling. No new files. See `engines/miotts/README.md`. |
| - **`mosslocal`** — Apache-2.0 — audio-cpp/audio.cpp-gguf card row 'Apache-2.0' for MOSS-TTS-Local-v1.5-GGUF + the originals list naming OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5 (2026-08-23). LOCAL USE via the audio.cpp route (v25.83.0, wave 2): mosslocal-clone-cpp (31 languages + code-switching; an OPTIONAL reference transcript rides a sidecar or is omitted) and, if the probe shipped it, mosslocal-tts-cpp (text-only, model-chosen voice). A DIFFERENT model family from the ComfyUI `moss` (v1 8B delay) and `mossnano` kinds. NOT mirrored by us. See `engines/mosslocal/README.md`. |
| - **`parakeet`** — CC-BY-4.0 — HF model card nvidia/parakeet-tdt-0.6b-v3 license:cc-by-4.0 (2026-08-23). ASR MODEL, NOT A RENDER ROUTE (v25.89.0, the ASR lane): parakeet-tdt-cpp in the ASR registry root - the CPU transcription model (11x realtime on the CPU build), 25 European languages, native word timestamps. CC-BY-4.0 (attribution) - mirror-eligible in principle; kept source origin like qwen3asr. See `engines/parakeet/README.md`. |
| - **`qwen3asr`** — Apache-2.0 — HF model cards Qwen/Qwen3-ASR-1.7B and Qwen/Qwen3-ForcedAligner-0.6B license:apache-2.0 (2026-08-23); the Silero VAD companion file is MIT (snakers4/silero-vad, redistributed inside the audio.cpp repo). ASR MODEL, NOT A RENDER ROUTE (v25.89.0, the ASR lane): qwen3-asr-cpp in the ASR registry root audiocpp-adapter/asr/ - the Transcribe tab's + the Refiner sidecars' audio.cpp backend, 30 languages incl. ja/zh/ko. Word timestamps need the aligner + VAD companions (the 0.6.1 binary refuses --words-out without them). Apache-2.0 weights - mirror-eligible in principle; kept source origin because audio-cpp's own repo is the ungated home and nothing in this row is needed by the render installer. See `engines/qwen3asr/README.md`. |
| - **`rvc`** — MIT — the audio-cpp/audio.cpp-gguf card row for RVC-GGUF (2026-08-24); RVC upstream is MIT. LOCAL USE via the audio.cpp route (v25.91.0, STS part 2): rvc-vc-cpp - voice conversion into a VOICE MODEL (the owner's own .pth/.index files under models\voices\rvc\<Voice>\ or a character's _refined\.rvc-model\ - never part of any mirror), server mode. The kind `rvc` existed as the Applio/RVC TRAINING lane's kind; this is the first engine that RUNS it. MIT base - kept source origin (audio-cpp's own repo is the ungated home; user voices are the owner's files). See `engines/rvc/README.md`. |
| - **`zonos`** — Apache-2.0 — HF API cardData, all three repos (Zyphra/Zonos-v0.1-hybrid, -transformer, -speaker-embedding), fetched 2026-08-18. Task 6 (2026-08-18): licence resolved permissive (Apache-2.0, same org and same verdict as the already-mirrored zonos2 row) — but NOT graduated this pass. The local copy at models\zonos\Zonos-v0.1-hybrid is a live git-lfs CHECKOUT, not a plain snapshot: it carries a .git\ directory measured at 3.1GB (56 objects) sitting inside the same folder as the actual weights. Pointing localPath at it as-is would sweep 3.1GB of git internals into the mirror lock/upload, which build-lock.js's walker cannot exclude (it hashes every file under localPath). Per the graduation rule ('licence-verified-permissive but files NOT locally stageable → stays origin'): needs a clean re-stage (e.g. a plain hf download or an rsync excluding .git) before this can graduate. Zonos-v0.1-transformer, by contrast, IS a clean snapshot on disk with no .git — only the hybrid variant has this problem. Do not conflate with the zonos2 row — different pack, different model generation. See `engines/zonos/README.md`. |
|
|
| ### 2d. Pending upstream licence verification |
|
|
| These engines' weights live on their own upstream repositories; this project has |
| not yet confirmed a redistributable licence for them, so nothing is mirrored here |
| until that verification lands (`licence.weights` reads "pending upstream |
| verification" until then). |
|
|
| - **`stepaudio`** — origin stepfun-ai/Step-Audio-EditX + stepfun-ai/Step-Audio-Tokenizer, licence: UNVERIFIED (pending upstream verification). See `engines/stepaudio/README.md`. |
|
|
| ## 3. Not applicable |
|
|
| - **`elevenlabs`** — cloud API — nothing to host. Direct engine descriptor lives in main.js (DIRECT_DESCRIPTORS.elevenlabs), not a workflow folder. |
| - **`windows`** — OS voices via SAPI5 (parrot-sapi.ps1) — nothing to host. Direct engine descriptor lives in main.js (DIRECT_DESCRIPTORS.windows). |
|
|
