Whisper Model Options Guide | Generated by AI
Here’s the full rundown of Whisper options:
== MODELS ==
tiny 39M params ~1 GB VRAM fastest, lowest quality base 74M params ~1 GB VRAM good for prototyping small 244M params ~2 GB VRAM solid balance medium 769M params ~5 GB VRAM high quality large 1550M params ~10 GB VRAM best quality, multilingual only turbo 809M params ~6 GB VRAM best speed/quality tradeoff
Recommendation: turbo for general use, base for quick tests.
== LANGUAGES (99 total) ==
Top-tier (WER < 10%): en English zh Chinese ja Japanese es Spanish ko Korean fr French de German it Italian pt Portuguese nl Dutch pl Polish ru Russian
Good (WER 10-20%): ar Arabic tr Turkish vi Vietnamese sv Swedish fi Finnish cs Czech ro Romanian hu Hungarian da Danish no Norwegian th Thai he Hebrew el Greek id Indonesian ms Malay
Also supports: Afrikaans, Albanian, Amharic, Armenian, Bengali, Bosnian, Bulgarian, Burmese, Cantonese, Catalan, Croatian, Estonian, Finnish, Georgian, Gujarati, Haitian Creole, Hindi, Icelandic, Javanese, Kannada, Kazakh, Khmer, Lao, Latin, Latvian, Lithuanian, Macedonian, Mongolian, Nepali, Persian, Punjabi, Sanskrit, Serbian, Slovak, Slovenian, Somali, Sundanese, Swahili, Tagalog, Tamil, Telugu, Tibetan, Turkish, Ukrainian, Urdu, Uzbek, Welsh, Yiddish, Yoruba, and more.
Language codes are ISO 639-1 (2-letter).
== KEY OPTIONS ==
language=”zh” # Specify language (faster than auto-detect) task=”transcribe” # Transcribe in original language (default) task=”translate” # Translate any language -> English initial_prompt=”…” # Context hint (improves accuracy for tech terms) word_timestamps=True # Word-level timing temperature=(0.0, 0.2, 0.4, 0.6, 0.8, 1.0) # Fallback temps device=”cuda” # Force GPU (auto-detects if not set) output_format=”srt” # txt, srt, vtt, json
== CLI USAGE ==
whisper audio.mp3 –model turbo –language Chinese whisper audio.mp3 –task translate –output_format srt
== CHINESE-SPECIFIC TIPS ==
- Use medium or large model (small has ~15% WER, large drops to ~5%)
- Specify language=”zh” explicitly
- Add initial_prompt for context: initial_prompt=”这是一段关于技术的讨论”
- Cantonese has its own code: language=”yue”