Models¶
VOXPrint MDL downloads weights into the shared Voxprint model folder. Movie Dubber does not download them.
The app manifest is mdl.json at the program root. Each entry below is a model that manifest pins.
The licence in dubber/models.py is the licence on that model's card. A download is pinned to the revision
below, and MDL checks every file (size, then SHA-256) before the folder is used. Diarization and the speaker
embedding are optional: one voice does not need them. Qwen3-TTS is written out in full, with the same
repository and revision Audiobook Builder uses, because voxprint-common 0.2.4 cannot resolve a shared-component
id into files. See the TODO in mdl.json.
The pyannote diarization pipeline and the speaker embedding are ungated mirrors. File hashes are the SHA256SUMS
published with those revisions. The pipeline config loads embedding, segmentation and PLDA from the mirror folder.
Registry¶
| Key | Repository | Revision | Licence | Checksums |
|---|---|---|---|---|
tts_1_7b |
Qwen/Qwen3-TTS-12Hz-1.7B-Base |
fd4b254389122332181a7c3db7f27e918eec64e3 |
Apache-2.0 | pinned in mdl.json (24 GB) |
tts_0_6b |
Qwen/Qwen3-TTS-12Hz-0.6B-Base |
5d83992436eae1d760afd27aff78a71d676296fc |
Apache-2.0 | pinned in mdl.json |
asr |
deepdml/faster-whisper-large-v3-turbo-ct2 |
4df90f75321148c3a29a9e2351b7ddf8f5b115a8 |
MIT | pinned in mdl.json |
sep |
JusperLee/TIGER-DnR |
b7a59560bbca10febbcd46fb01600f868e587f57 |
Apache-2.0 (weights), MIT (code) | pinned in mdl.json |
diar |
Mitroshenkov87/voxprint-pyannote-speaker-diarization-community-1 |
9c851af727f1085a361777d13e32734f7f5ebb30 |
CC-BY-4.0 | pinned in mdl.json |
embed |
Mitroshenkov87/voxprint-pyannote-embedding |
4655ca656e1d7b1d8fe8100b39b4fdbf3015c7b8 |
MIT | pinned in mdl.json |
roformer |
KimberleyJSN/melbandroformer |
ac9b0614ab3cd7f77219e18ba494dfd93956c348 |
MIT | pinned in mdl.json |
mt_en_ru |
Helsinki-NLP/opus-mt-tc-big-en-zle |
708be1d372fe4c358a352f404e6dc9ca0126ba48 |
CC-BY-4.0 | pinned in mdl.json |
mt_ru_en |
Helsinki-NLP/opus-mt-tc-big-zle-en |
09a40f722d6d8b76aaad6fe51a06c914622a13d1 |
CC-BY-4.0 | pinned in mdl.json |
mt_ru_de |
Helsinki-NLP/opus-mt-tc-big-zle-de |
b2e247f0c413ca6aa51a32f2f2be8666cf72405e |
CC-BY-4.0 | pinned in mdl.json |
mt_de_ru |
Helsinki-NLP/opus-mt-tc-big-de-zle |
d4db2a2cbaa6c2f1ea57d0ed40924d35767b05f9 |
CC-BY-4.0 | pinned in mdl.json |
mt_en_de |
Helsinki-NLP/opus-mt-tc-bible-big-deu_eng_fra_por_spa-gmw |
0a10a154b1d057720d03d8227bc59ea9633c590b |
Apache-2.0 | pinned in mdl.json |
mt_de_en |
Helsinki-NLP/opus-mt-tc-bible-big-gmw-deu_eng_fra_por_spa |
a6933ed14d06d91d797809258bdb2d52f5487a36 |
Apache-2.0 | pinned in mdl.json |
The 2020 Helsinki-NLP opus-mt-en-ru and opus-mt-de-en cards are Apache-2.0. opus-mt-ru-en and opus-mt-en-de are
CC-BY-4.0. Those four checkpoints are not used. en↔ru and ru↔de use tc-big. en↔de uses tc-bible-big.
Roles¶
Pipeline stages ask dubber/model_resolver.py for a role. The resolver finds the matching mdl.json entry
(model id, plus the VRAM tier for Qwen3-TTS), then the complete local folder. Stage code does not name a
repository, a revision or a path, and nothing is downloaded. A missing model raises ModelUnavailable: the CLI
exits with code 4 (models) and the window shows its Manage models notice.
| Role | mdl.json entry |
Used by | Placeholder |
|---|---|---|---|
separation |
sep |
separation stage (TIGER-DnR) | yes |
separation.best |
roformer |
separation stage, roformer option |
yes |
asr |
asr |
recognition, reference-clip check | yes |
diarization |
diar |
speaker stage (optional) | yes |
embedding |
embed |
voice matching (optional) | yes |
translation.<src>-<tgt> |
mt-<src>-<tgt> |
translation stage | no |
tts.1_7b |
qwen3-tts, tier 24gb |
speech stage | yes |
tts.0_6b |
qwen3-tts, tier 16gb |
speech stage | yes |
Entries marked "placeholder": true in mdl.json point at the current upstream weights until Model Forge publishes
the suite model for that role. Replacing one is a mdl.json change only.
Silero VAD is not in mdl.json: its weights ship inside the silero-vad and faster-whisper wheels, which the
runtime lock pins.
Memory between stages¶
The model session unloads every resident model the next stage does not use and empties the CUDA cache before the stage starts. Speech synthesis therefore starts with separation, recognition and diarization off the card. The TTS batch is sized from the free VRAM measured before each batch, leaving 8 % of the card (between 1.5 and 2.5 GB) unused; an out-of-memory error halves the batch. A VRAM tier picked by hand still caps the lines per call.
Mel-Band RoFormer¶
KimberleyJSN/melbandroformer commit ac9b0614ab3cd7f77219e18ba494dfd93956c348 (2026-04-22T12:30:27Z, "Update README.md")
sets the model-card licence to MIT. The previous card (f45f9e3d8570a406c94ac34b29f49ce43fda3bc8, 2025-06-17) was GPL-3.0.
The pinned revision is that MIT commit. MelBandRoformer.ckpt was uploaded earlier (94a0e5de2622a4160198b158e6f8141296da887e,
2024-08-06); the file at the MIT revision is the one that is downloaded and checksummed.
audio-separator looks up vocals_mel_band_roformer.ckpt. The store keeps the published filename and links that expected
name to MelBandRoformer.ckpt, so the package uses its yaml and does not download a second copy of the weights.
Translation pairs¶
tc-big models, the same repositories as Voxprint AI Audiobook Builder:
| Pair | Model | Target token |
|---|---|---|
| en → ru | Helsinki-NLP/opus-mt-tc-big-en-zle |
>>rus<< |
| ru → en | Helsinki-NLP/opus-mt-tc-big-zle-en |
none |
| ru → de | Helsinki-NLP/opus-mt-tc-big-zle-de |
none |
| de → ru | Helsinki-NLP/opus-mt-tc-big-de-zle |
>>rus<< |
| en → de | Helsinki-NLP/opus-mt-tc-bible-big-deu_eng_fra_por_spa-gmw |
>>deu<< |
| de → en | Helsinki-NLP/opus-mt-tc-bible-big-gmw-deu_eng_fra_por_spa |
>>eng<< |
en→de and de→en use the tc-bible-big models. The model cards list >>deu<< and >>eng<< among the valid target-language labels. The 2020 Opus-MT checkpoints are not used.
CC BY 4.0 attributions¶
The Opus-MT tc-big models below are used unmodified under CC BY 4.0.
Opus-MT tc-big (en→ru, ru→en, ru→de, de→ru). Language Technology Research Group at the University of Helsinki (Helsinki-NLP). Source: https://huggingface.co/Helsinki-NLP. Licence: https://creativecommons.org/licenses/by/4.0/. Tiedemann, Aulamo, Bakshandaeva, Boggia, Grönroos, Nieminen, Raganato, Scherrer, Vázquez, Virpioja: "Democratizing neural machine translation with OPUS-MT", Language Resources and Evaluation 58 (2024).
pyannote speaker-diarization-community-1. Hervé Bredin and the pyannote.audio authors.
Original: https://huggingface.co/pyannote/speaker-diarization-community-1.
Mirror: https://huggingface.co/Mitroshenkov87/voxprint-pyannote-speaker-diarization-community-1 at 9c851af727f1085a361777d13e32734f7f5ebb30.
Licence: https://creativecommons.org/licenses/by/4.0/.
No changes besides removing the access gate. Segmentation, the WeSpeaker embedding and PLDA are the subfolders of that mirror.
MIT attribution¶
pyannote/embedding. Copyright (c) 2022 CNRS. pyannote.audio.
Original: https://huggingface.co/pyannote/embedding.
Mirror: https://huggingface.co/Mitroshenkov87/voxprint-pyannote-embedding at 4655ca656e1d7b1d8fe8100b39b4fdbf3015c7b8.
Licence: MIT.
No changes besides removing the access gate.
Apache-2.0 attributions¶
Opus-MT tc-bible-big (en→de, de→en). Language Technology Research Group at the University of Helsinki (Helsinki-NLP).
Source: https://huggingface.co/Helsinki-NLP/opus-mt-tc-bible-big-deu_eng_fra_por_spa-gmw and https://huggingface.co/Helsinki-NLP/opus-mt-tc-bible-big-gmw-deu_eng_fra_por_spa.
Licence: Apache-2.0.
Both model cards list the licence as Apache-2.0. en→de requires the sentence-initial token >>deu<<. de→en requires >>eng<<.