whisper-medium-komixv2 (int8 q8, ONNX)
seastar105/whisper-medium-komixv2λ₯Ό optimum 1.27 + onnxruntime 1.29λ‘ μ΅μ€ν¬νΈνκ³
int8(q8) λμ μμνν λͺ¨λΈ β meeting-ai-pwaμ μ μ¬(STT) κΈ°λ³Έ λͺ¨λΈ μΉκ²©μ©.
νμΌ
| νμΌ | ν¬κΈ° | λΉκ³ |
|---|---|---|
onnx/encoder_model_quantized.onnx |
~314MB | encoder (q8) |
onnx/decoder_model_merged_quantized.onnx |
~873MB | merged decoder (q8) |
| κ·Έ μΈ | β | tokenizer/config (optimum μ΅μ€ν¬νΈ μλ³Έ) |
λ‘λ λ°©λ² (transformers.js)
import { pipeline } from '@huggingface/transformers';
const asr = await pipeline('automatic-speech-recognition', 'ALLUCY-Rodent/whisper-medium-ko', {
dtype: { encoder_model: 'q8', decoder_model_merged: 'q8' },
});
λ³ν μ¬ν
# 1) μ΅μ€ν¬νΈ (--task μλ΅ νμ β merged λμ½λ λ―Έμμ± λ²κ·Έ)
optimum-cli export onnx -m seastar105/whisper-medium-komixv2 .convert/whisper-medium-komixv2-onnx
# 2) q8 μμν
python scripts/quantize_whisper_ko.py <encoder.onnx> <encoder_q8.onnx>
python scripts/quantize_whisper_merged_decoder.py <decoder_merged.onnx> <decoder_merged_q8.onnx>
A/B κ²μ¦: 2026-08-26, AI Hub λ¨λ¬Έ 3 + μ₯λ¬Έ 38.8s β small(+11.2%p μ νλ μ°μ).
- Downloads last month
- 11
Model tree for ALLUCY-Rodent/whisper-medium-ko
Base model
openai/whisper-medium Finetuned
seastar105/whisper-medium-komixv2