Korean Female Voice Model Pro (Epoch 1000 / RVC2)

[regal forge]

This model is newly developed, incorporating a novel recording method and having undergone dataset cleaning. It has been refined to handle expressions in Korean, which are challenging to implement in previous models. This model is designed for use in Korean speech-to-speech, voice changers, music covers, and more.

위의 모델은 새로운 녹음 / 리빌트 방식과 데이터셋 클린 작업을 거친 모델 입니다. 해당 모델은 기존 모델들에 비해 구현이 어려운 한국어의 표현이 가능할 수 있도록 교정된 모델로 한국어 스피치 2 스피치, 보이스 체인저, 음악 커버등에 사용할 수 있도록 제작하였습니다.

Having fun! 🙂

Information :

Model :
Model Name : Korean Female Voice Pro

Train info :
RVC 2
1k Epochs
Batch size per GPU : 40

Dataset - source Info :
22 mins of speech
12 mins of Sing

Dataset - source correction :
KLM V7 Pretrained

Model Link-
https://huggingface.co/SeoulStreamingStation/IU-Voice-MultiLanguage-V1/resolve/main/SSSFemaleVoicePro.zip?download=true

[ripe osprey]

Rate the model

To rate this model, either use /rate or click the buttons below.

[hollow hemlock]

Where do I get KLM V7 pretrain

I'm thinking of making a GI Kirara Korean model

CV: Kang Eun-ae

[serene canyon]

Hey, mind to ask you, whats this KLM V7 Pretrain? where you got it from ?

[blissful lion]

its just a korean pretrain they made

simply put they took a bunch of high quality audio, cleaned it up and trained a pretrain with it

the reason why this sounds clean is because

1. it's a language u dont understand
2. the audio which was fed to RVC was high quality
3. high batch size which usually makes learning more stable but ofc limits the model a bit more

so to answer ur questions, no this isnt some ultimate pretrain they made

[serene canyon]

oh okay thought it was a public avaible pretrain lol

[blissful lion]

maybe it is, who knows

[reef elm]

Maybe with this pretrain now it will be possible to make some proper BTS and Blackpink members models.

[regal forge]

I woke up to find that a lot of people had left comments. Thanks for all the kind explanations. As Lisa mentioned, KLM is based on materials published by the Korean Language Research Institute and is composed of pretrained data. It specifically aims to generate sentences that can produce various sounds connecting words, focusing on learning and implementing various irregular sounds. This is not an open model yet, so it can't be downloaded. Separately, as many experts here have mentioned, regardless of the pretrained model, the quality of audio or its quality ultimately depends much more on the cleanup work of the dataset to be trained. We're trying various methods for restoration and studying hard to create a model of the highest quality possible.

[river vapor]

Ayo? @regal forge level 6 !!!

[regal forge]

Thanks again for your interest!

[hollow hemlock]

OK

[blissful lion]

There ya go ppl. :3