KLM v7s Batch II (PRE-TRAINED MODEL)

Created: April 22, 2024
KLM v7s Batch II (PRE-TRAINED MODEL)

This is Batch 2 of the KLM7 share version, a pre-trained Korean language model. It has been tested on Applio and comes with pre-trained models at RVC v2 / RVMPE / 32k, 40k, and 48k.

It is compatible with all tools that can utilize pre-trained models for discriminators and generators, though it has been specifically tested with Applio. To use it, decompress the data in the folder where Applio is installed, under rvc\pretraineds\pretraineds_custom.

While KLM can be used for languages other than Korean, it contains the most information for Korean, particularly based on vocal-related data.

P.S.:
Do not treat KLM as a magical tool. Always monitor the training process through TensorBoard to ensure it is sufficient. It helps to prepare the cleanest and most ideal source of data by organizing as much of the dataset as possible beforehand. The more detailed and clean the source you use, the more detail you can preserve.

If you have a 10-minute dataset, and you can prepare a high-quality 5-minute subset, use that 5-minute data, but it will require considerable effort to make the most of those 5 minutes. I recommend conducting various tests based on discussions among many AI hubbers in channels like model maker

ν•œκ΅­μ–΄ μ‚¬μ „ν•™μŠ΅λͺ¨λΈμΈ KLM7 μ‰μ–΄λ²„μ „μ˜ 배치2 μž…λ‹ˆλ‹€.
Applio μ—μ„œ ν…ŒμŠ€νŠΈ λ˜μ—ˆμœΌλ©° RVC v2 / RVMPE / 32k, 40k, 48k 사전λͺ¨λΈμ΄ μ€€λΉ„λ˜μ–΄ μžˆμŠ΅λ‹ˆλ‹€.

μ°¨λ³„μžμ™€ μƒμ„±μžλ₯Ό μ‚¬μ „ν•™μŠ΅ λͺ¨λΈλ‘œ μ΄μš©ν•  수 μžˆλŠ” λͺ¨λ“  νˆ΄μ—μ„œ 이용이 κ°€λŠ₯ν•˜μ§€λ§Œ Applioμ—μ„œ ν…ŒμŠ€νŠΈ λ˜μ—ˆμŠ΅λ‹ˆλ‹€. μ••μΆ•λœ 데이터λ₯Ό Applioκ°€ μ„€μΉ˜λ˜μ–΄ μžˆλŠ” 폴더 μ•ˆ rvc\pretraineds\pretraineds_custom 에 압좕을 ν’€μ–΄μ€€ ν›„ μ΄μš©ν•˜μ‹œλ©΄ λ©λ‹ˆλ‹€.

KLM은 μ˜μ–΄ μ™Έ λ‹€λ₯Έ κ΅­κ°€μ˜ μ–Έμ–΄μ˜ μ΄μš©λ„ κ°€λŠ₯ν•˜μ§€λ§Œ ν•œκ΅­μ–΄μ— κ°€μž₯ λ§Žμ€ 정보λ₯Ό λ‹΄κ³  있고 특히 보컬과 κ΄€λ ¨λœ 데이터λ₯Ό κΈ°λ°˜μ— 두고 μžˆμŠ΅λ‹ˆλ‹€.

μΆ”μ‹  :
KLM을 λ§ˆλ²•μ˜ λ„κ΅¬λ‘œ μ‚¬μš©ν•˜μ§€ λ§ˆμ‹­μ‹œμ˜€. ν…μ„œλ³΄λ“œμ™€ λͺ¨λ‹ˆν„°λ§μ„ 톡해 μΆ©λΆ„ν•œ ν•™μŠ΅μ΄ μ΄λ€„μ§€λŠ”μ§€ ν•­μ‹œ μ§€μΌœλ³΄μ•„μ•Ό ν•˜λ©° κ°€λŠ₯ν•œ ν•  수 μžˆλŠ” 만큼의 μ΅œλŒ€ν•œμ˜ 데이터셋 정리λ₯Ό 사전에 ν•˜μ—¬ κ°€μž₯ ν΄λ¦°ν•˜κ³  이상적인 μ†ŒμŠ€λ₯Ό μ΄μš©ν•˜λ©΄ ν• μˆ˜λ‘ 더 λ§Žμ€ λ””ν…ŒμΌμ„ 살릴 수 있게 λ„μšΈ κ²ƒμž…λ‹ˆλ‹€.

(λ˜₯을 아무리 이쁘게 꾸며도 이쁜 λ˜₯이 λ˜λŠ” 근본적인 법칙을 λ¬΄μ‹œν•  수 μ—†μŠ΅λ‹ˆλ‹€.)

10λΆ„μ§œλ¦¬ 데이터셋이 μžˆλ‹€κ³  κ°€μ •ν• λ•Œ 졜고의 ν€„λ¦¬ν‹°λ‘œ 5λΆ„μ§œλ¦¬ 데이터셋을 μ€€λΉ„ν•  수 μžˆλ‹€λ©΄ κ·Έ 5λΆ„μ˜ 데이터λ₯Ό μ΄μš©ν•˜μ‹œλ˜ κ·Έ 5뢄을 살리기 μœ„ν•΄ 적지 μ•Šμ€ λ…Έλ ₯을 ν•˜μ…”μ•Ό ν•©λ‹ˆλ‹€. model maker 채널등을 톡해 λ§Žμ€ AI huber듀이 ν† λ‘ ν•˜λŠ” λ‚΄μš©μ„ ν† λŒ€λ‘œ λ‹€μ–‘ν•œ ν…ŒμŠ€νŠΈλ₯Ό ν•΄λ³΄μ‹œκΈΈ ꢌμž₯λ“œλ¦½λ‹ˆλ‹€.

Batch 1 :
10 Actors and Vocalist's Voice, Vocal Data

Batch 2 :
22 Actors and Vocalist's Voice, Vocal Data

Model Link - https://huggingface.co/SeoulStreamingStation/KLMv7s/resolve/main/KLMv7s_Batch2.zip?download=true

Additional Details:

Tags: No tags available

Download Link: https://huggingface.co/SeoulStreamingStation/KLMv7s/resolve/main/KLMv7s_Batch2.zip?download=true Hosted on Hugging Face