kimthegarden/whisper-small-ko-new

Model Card for whisper-small-ko-finetuned

This is a fine-tuned version of the SungBeom/whisper-small-ko model on a custom Korean speech recognition dataset.
It performs automatic speech recognition (ASR) for Korean audio data and achieves strong performance on the validation set.

---

Model Details

Model Description

This model is based on the Whisper-small architecture and fine-tuned on 62,327 Korean audio-transcript pairs using Hugging Face Transformers and PyTorch.
It is designed for general-domain Korean speech recognition (conversational, broadcast, news, etc.).

• Developed by: [Jeongwon Kim]
• Shared by: [kimthegarden]
• Model type: Encoder-decoder Transformer (WhisperForConditionalGeneration)
• Language(s): Korean (ko)
• License: MIT
• Fine-tuned from model: SungBeom/whisper-small-ko

---

Model Sources

• Repository: [https://huggingface.co/kimthegarden/whisper-small-ko-low-qual-voice]
• Notebook: Fine-tuned using a custom whisper_finetuning.ipynb
• Demo [optional]: [Gradio or Streamlit demo link if available]

---

Uses

Direct Use

• Korean automatic speech recognition (ASR)
• Offline or batch transcription of Korean speech data
• Integration into Korean-language voice assistant systems

Downstream Use

• Further fine-tuning on domain-specific datasets (e.g. legal, medical, education)
• Research into Korean ASR model robustness or multilingual Whisper models

Out-of-Scope Use

• Transcription of non-Korean speech (this model is Korean-only)
• Real-time streaming ASR (not latency-optimized)
• Zero-shot or few-shot adaptation to other languages

---

Bias, Risks, and Limitations

• The model may show reduced accuracy on:
  • Regional dialects or accents not represented in the training data
  • Very noisy environments
  • Children’s speech or non-native pronunciation
• The model has not been tested for fairness across different speakers (gender, age, etc.)

Recommendations

We recommend testing the model on your specific data domain before deployment.
Additional fine-tuning or data filtering may be required for sensitive use cases (e.g. education, healthcare).

---

How to Get Started with the Model

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

model = WhisperForConditionalGeneration.from_pretrained("your-username/whisper-small-ko-finetuned")
processor = WhisperProcessor.from_pretrained("your-username/whisper-small-ko-finetuned")

# Input: 16kHz waveform (float32 numpy or tensor)
inputs = processor(audio_waveform, sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    predicted_ids = model.generate(inputs.input_features)

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(transcription[0])

---

Contact

jwk20001007@gmail.com