tamarand/whisper-small-ko-new-onnx

Whisper Small KO New (ONNX)

Korean automatic speech recognition model exported to ONNX for fast inference.

Model

• Base model: kimthegarden/whisper-small-ko-new
• Task: automatic speech recognition (Korean)
• License: MIT

Files

• encoder_model.onnx, decoder_model.onnx
• config.json, generation_config.json
• Tokenizer and processor files (tokenizer.json, vocab.json, etc.)

Quickstart (Python)

Install dependencies:

pip install transformers optimum[onnxruntime] soundfile

Run inference:

import numpy as np
import soundfile as sf
from optimum.onnxruntime import ORTModelForSpeechSeq2Seq
from transformers import WhisperProcessor

model_id = "YOUR_USERNAME/whisper-small-ko-new-onnx"
processor = WhisperProcessor.from_pretrained(model_id)
model = ORTModelForSpeechSeq2Seq.from_pretrained(model_id)

# Load 16kHz mono audio
audio, sr = sf.read("sample.wav")
if sr != 16000:
    raise ValueError("Please provide 16kHz audio.")
if audio.ndim > 1:
    audio = np.mean(audio, axis=1)

inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
predicted_ids = model.generate(inputs.input_features)
text = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(text)

Notes

• Input audio must be 16 kHz mono.
• For long audio, split into chunks (e.g., 30 seconds) before decoding.