from faster_whisper import WhisperModel import os # 1. Choose model size (best Korean accuracy: large-v3 or large-v3-turbo) # Options: tiny, base, small, medium, large-v3, large-v3-turbo model_size = "large-v3-turbo" # fastest good accuracy (~6× faster than large-v3, almost same quality for Korean) # model_size = "large-v3" # highest accuracy if you have good hardware # Load model (downloads automatically first time; ~1.5–3 GB) model = WhisperModel( model_size, device="cuda", # "cuda" if NVIDIA GPU, else "cpu" compute_type="float16" # "float16" or "int8" for faster/lower memory on GPU # compute_type="int8" # good for CPU or low VRAM ) # Your Korean MP3 file audio_file = "your_korean_audio.mp3" # or .wav, .m4a, etc. (supports most formats) # Transcribe segments, info = model.transcribe( audio_file, language="ko", # Force Korean (improves accuracy & speed) beam_size=5, # Higher = slightly better accuracy, slower vad_filter=True, # Remove silence/noise (recommended) vad_parameters=dict(min_silence_duration_ms=500), ) # Print detected language & probability (should be ko ~0.95+) print(f"Detected language: {info.language} (probability: {info.language_probability:.2f})") # Output text + timestamps print("\nTranscription:\n") for segment in segments: print(f"[{segment.start:.2f}s → {segment.end:.2f}s] {segment.text}") # Optional: Save to file with open("transcript_korean.txt", "w", encoding="utf-8") as f: for segment in segments: f.write(f"[{segment.start:.2f} → {segment.end:.2f}] {segment.text}\n")