Model reference · open weights

ko-spelling-wav2vec2-conformer-del-1s

Audio 42MARU Speech→text 1 build Open weights 1k dl/mo

ko-spelling-wav2vec2-conformer-del-1s is an open-weight audio or speech model from 42MARU. ko-spelling-wav2vec2-conformer-del-1s (FP32) weighs 360 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.

  • ko-spelling-wav2vec2-conformer-del-1s is a 180M parameter automatic speech recognition model developed by 42MARU for Korean language processing.
  • It is based on the wav2vec2-conformer architecture and was fine-tuned on the KsponSpeech dataset using spelling transcription data.
  • The model is released under the apache-2.0 license.

Summary of the 42MARU/ko-spelling-wav2vec2-conformer-del-1s model card, 2026-10-01

What it is

Released by42MARU
Released2022-10-31
Parameters180M
VRAM360 MB for the weights

What it runs on

Memory and cards for ko-spelling-wav2vec2-conformer-del-1s (FP32)

360 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 42MARU says about ko-spelling-wav2vec2-conformer-del-1s

Read the model card

Table of Contents

Model Details

  • Model Description: 해당 모델은 wav2vec2-conformer base architecture에 scratch pre-training 되었습니다. Wav2Vec2ConformerForCTC를 이용하여 KsponSpeech에 대한 Fine-Tuning 모델입니다.

  • Dataset use AIHub KsponSpeech Datasets는 해당 Data를 전처리하여 임의로 만들어 사용하였습니다. del-1s의 의미는 1초 이하의 데이터 필터링을 의미합니다. 해당 모델은 철자전사 기준의 데이터로 학습된 모델입니다. (숫자와 영어는 각 표기법을 따름)

  • Developed by: TADev (@lIlBrother, @ddobokki, @jp42maru)

  • Language(s): Korean

  • License: apache-2.0

  • Parent Model: See the wav2vec2-conformer for more information about the pre-trained base model. (해당 모델은 wav2vec2-conformer base architecture에 scratch pre-training 되었습니다.)

Evaluation

Just using load_metric("wer") and load_metric("wer") in huggingface datasets library

How to Get Started With the Model

KenLM과 혼용된 Wav2Vec2ProcessorWithLM 예제를 보시려면 42maru-kenlm 예제를 참고하세요

import librosa
from pyctcdecode import build_ctcdecoder
from transformers import (
    AutoConfig,
    AutoFeatureExtractor,
    AutoModelForCTC,
    AutoTokenizer,
    Wav2Vec2ProcessorWithLM,
)
from transformers.pipelines import AutomaticSpeechRecognitionPipeline

audio_path = ""

# 모델과 토크나이저, 예측을 위한 각 모듈들을 불러옵니다.
model = AutoModelForCTC.from_pretrained("42MARU/ko-spelling-wav2vec2-conformer-del-1s")
feature_extractor = AutoFeatureExtractor.from_pretrained("42MARU/ko-spelling-wav2vec2-conformer-del-1s")
tokenizer = AutoTokenizer.from_pretrained("42MARU/ko-spelling-wav2vec2-conformer-del-1s")
beamsearch_decoder = build_ctcdecoder(
    labels=list(tokenizer.encoder.keys()),
    kenlm_model_path=None,
)
processor = Wav2Vec2ProcessorWithLM(
    feature_extractor=feature_extractor, tokenizer=tokenizer, decoder=beamsearch_decoder
)

# 실제 예측을 위한 파이프라인에 정의된 모듈들을 삽입.
asr_pipeline = AutomaticSpeechRecognitionPipeline(
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    decoder=processor.decoder,
    device=-1,
)

# 음성파일을 불러오고 beamsearch 파라미터를 특정하여 예측을 수행합니다.
raw_data, _ = librosa.load(audio_path, sr=16000)
kwargs = {"decoder_kwargs": {"beam_width": 100}}
pred = asr_pipeline(inputs=raw_data, **kwargs)["text"]
# 모델이 자소 분리 유니코드 텍스트로 나오므로, 일반 String으로 변환해줄 필요가 있습니다.
result = unicodedata.normalize("NFC", pred)
print(result)
# 안녕하세요 123 테스트입니다.

Beam-100 Result (WER):

"clean""other"
22.0127.34

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