Datasets:
Precise Model Checkpoints
#5
by morshedm - opened
Hello! First of all, thank you for your wonderful work on this dataset. MLAAD is very rich in the variety of deepfake models and also in how up-to-date it is.
Would it be possible to have a more precise description of each model used to generate MLAAD v11?
A majority of the models are quite clear: for example facebook/mms-tts-eng.
But several of them are a little ambiguous, like griffin_lim (this is just an algorithm, is there a specific model used?), or maybe GPT-SoVITS (there are several versions, which one is used?).
Hi,
- griffin lim is just an algorithm for vocoding: https://librosa.org/doc/0.11.0/generated/librosa.griffinlim.html
- GPT SoVITS: https://github.com/RVC-Boss/GPT-SoVITS
Let me know if there are others you'd need more info about.
Hi, thanks for the response!
GPT-SoVITSFor https://github.com/RVC-Boss/GPT-SoVITS, there are 5 or more versions (https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90features(Latest%E2%80%90Including%E2%80%90v5)). I was wondering which version is used. Is it just v1?OuteTTS: There are many models in https://huggingface.co/collections/OuteAI/outetts, and there also seems to be a v1.0 (https://huggingface.co/collections/OuteAI/outetts-10). Could you point to the specific hf model?RVC: Is it from https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI ? There seems to be a v1 and v2 model.- For
tts_models/fa/custom/glow-tts,tts_models/it/mai_female/glow-ttsandtts_models/it/mai_male/glow-ttsis it possible to know the vocoder? I believe these are all coqui models. Are they using griffin lim or some other specific neural vocoder?
Thanks again.
morshedm changed discussion status to closed
morshedm changed discussion status to open
Hi again,
- OuteTTS: OuteAI/Llama-OuteTTS-1.0-1B (https://huggingface.co/OuteAI/Llama-OuteTTS-1.0-1B).
- Coqui glow-tts (fa/custom, it/mai_female, it/mai_male): these models ship without a default vocoder in Coqui's registry, and we ran them with the stock Coqui TTS CLI, so the waveform was produced with Griffin-Lim. The uk/mai and en/ljspeech glow-tts models use Multi-band MelGAN and tr/common-voice uses HiFi-GAN.
- RVC: I think it was v2 but I cannot confirm this right now
- GPT-SoVITS: I think it was v2 but again, cannot confirm this right now either.
Thank you for the help!
morshedm changed discussion status to closed