2025-korean-asr-benchmark

A personal benchmarking tool for ASR models across multilingual datasets.

---

Features

• Supports Whisper & Qwen2.5-Omni
• Languages: Korean, English, Chinese, Russian, French
• Computes WER & CER
• Caching for faster re-runs

---

Models

• openai/whisper-large-v3
• openai/whisper-large-v3-turbo
• ghost613/whisper-large-v3-turbo-korean
• seongsubae/openai-whisper-large-v3-turbo-ko-TEST (experimental)
• Qwen/Qwen2.5-Omni-7B (requires custom Transformers build)

---

Dataset

• Bingsu/zeroth-korean
  (Add more via load_dataset())

---

Run

python benchmarks/run_benchmark.py --normalize --verbose