Model reference · open weights

hubert-korean

Audio team-lucid Speech→text 1 build Open weights 31k dl/mo

hubert-korean is an open-weight audio or speech model from team-lucid. hubert-base-korean (FP32) weighs 568 MB; the smallest configuration that runs it is RTX 3060 12 GB.

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

  • hubert-korean is a 94M parameter speech representation learning model developed by team-lucid for automatic speech recognition in Korean.
  • It uses a self-supervised approach to learn directly from raw waveforms and was trained on approximately 4,000 hours of Korean audio data.
  • The model is available under the Apache 2.0 license.

Summary of the team-lucid/hubert-base-korean model card, 2026-10-01

What it is

Released byteam-lucid
Released2023-05-29
Parameters94M
VRAM568 MB for the weights

What it runs on

Memory and cards for hubert-base-korean (FP32)

568 MBweights, file size
1.6 GBruntime overhead
CardOne streamMemory
RTX 3060 12 GBfits11.6 GB
RTX 4060 Ti 16 GBfits15.4 GB
RTX 3090 24 GBfits23.4 GB
RTX 4090 24 GBfits23.4 GB
RTX 5090 32 GBfits31.0 GB
L40S 48 GBfits44.0 GB
A100 80 GBfits78.2 GB
H100 80 GBfits78.1 GB
RTX PRO 6000 Blackwell 96 GBfits93.8 GB
DGX Spark (GB10) 128 GB unifiedfits107 GB
H200 141 GBfits138 GB
B200 180 GBfits176 GB

From the model card

What team-lucid says about hubert-korean

Read the model card

Model Details

Hubert(Hidden-Unit BERT)는 Facebook에서 제안한 Speech Representation Learning 모델입니다. Hubert는 기존의 음성 인식 모델과 달리, 음성 신호를 raw waveform에서 바로 학습하는 self-supervised learning 방식을 사용합니다.

이 연구는 구글의 TPU Research Cloud(TRC)를 통해 지원받은 Cloud TPU로 학습되었습니다.

Model Description

How to Get Started with the Model

Pytorch

import torch
from transformers import HubertModel

model = HubertModel.from_pretrained("team-lucid/hubert-base-korean")

wav = torch.ones(1, 16000)
outputs = model(wav)
print(f"Input:   {wav.shape}")  # [1, 16000]
print(f"Output:  {outputs.last_hidden_state.shape}")  # [1, 49, 768]

JAX/Flax

import jax.numpy as jnp
from transformers import FlaxAutoModel

model = FlaxAutoModel.from_pretrained("team-lucid/hubert-base-korean", trust_remote_code=True)

wav = jnp.ones((1, 16000))
outputs = model(wav)
print(f"Input:   {wav.shape}")  # [1, 16000]
print(f"Output:  {outputs.last_hidden_state.shape}")  # [1, 49, 768]

Training Details

Training Data

해당 모델은 과학기술정보통신부의 재원으로 한국지능정보사회진흥원의 지원을 받아 구축된 자유대화 음성(일반남여), 다화자 음성합성 데이터, 방송 콘텐츠 대화체 음성인식 데이터 에서 약 4,000시간을 추출해 학습되었습니다.

Training Procedure

원 논문과 동일하게 MFCC 기반으로 Base 모델을 학습한 다음, 500 cluster로 k-means를 수행해 다시 Base와 Large 모델을 학습했습니다.

Training Hyperparameters
HyperparameterBaseLarge
Warmup Steps32,00032,000
Learning Rates5e-41.5e-3
Batch Size128128
Weight Decay0.010.01
Max Steps400,000400,000
Learning Rate Decay0.10.1
\(Adam\beta_1\)0.90.9
\(Adam\beta_2\)0.990.99

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms