Whisper Cheatsheet

Q: How do I install Whisper?

Run `pip install openai-whisper`. For GPU acceleration, also install PyTorch with CUDA: `pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124`. On macOS with Apple Silicon, Whisper uses MPS automatically. Requires `ffmpeg` for audio processing.

Q: Which model size should I choose?

tiny (fastest, least accurate), base (good for clean audio), small (balanced), medium (high quality, ~5GB VRAM), large (best quality, ~10GB VRAM). For most English podcasts, small or medium is sufficient. For noisy recordings or non-English languages, use large. For real-time, use tiny.

Q: How do I transcribe non-English audio?

Use `--task transcribe` to transcribe while keeping the original language. Use `--language zh` (or `ja`, `fr`, `de`, etc.) to improve accuracy. Use `--task translate` to transcribe AND translate to English in one step.

Q: How do I generate subtitles (SRT/VTT)?

Use `--output_format srt` for SRT subtitles or `--output_format vtt` for WebVTT format. These work directly with video editors and streaming platforms. Use `--output_dir` to specify where the files are saved.

Q: Can I use Whisper with GPU?

Yes, add `--device cuda --fp16 True` for NVIDIA GPU acceleration. This speeds up transcription 5-10x. Without GPU, Whisper uses CPU which is still usable but slower — expect 2-5x real-time for medium model.

Whisper is OpenAI's open-source general-purpose speech recognition model. It supports transcription and translation across 100+ languages, offers model sizes from tiny to large, runs on CPU or GPU, and outputs in txt/vtt/srt/tsv/json formats.

Updated: 2026-07-20·20 commands

Quick Start

``bash # Install pip install openai-whisper

# Basic transcription whisper recording.mp3

# Specify model and language whisper meeting.mp3 --model small --language zh

# Translate Chinese audio to English whisper chinese_speech.mp3 --task translate

# Generate SRT subtitles whisper video_audio.mp3 --model base --output_format srt --output_dir ./subtitles

# GPU accelerated transcription whisper long_podcast.mp3 --model medium --device cuda --fp16 True`

`Model Sizes`

Whisper offers five pre-trained models trading speed for accuracy: tiny (39M params, ~1GB VRAM), base (74M, ~1GB), small (244M, ~2GB), medium (769M, ~5GB), large (1.5B, ~10GB). Models are downloaded automatically on first use.

`Transcription & Translation`

Two task modes: transcribe (speech to text in original language) and translate (speech to English text). Use --language to hint the source language for better accuracy. For bilingual audio, Whisper auto-detects.

`Output Formats`

Whisper supports multiple output formats. Use --output_format to choose: txt (plain text, default), vtt (WebVTT subtitles), srt (SubRip subtitles), tsv (tab-separated values with timestamps), json (full metadata including segments and word-level timestamps). Use all to generate all formats at once.

`Advanced Options`

--initial_prompt guides style or domain-specific terminology. --temperature controls sampling randomness (0 = deterministic). --condition_on_previous_text` improves long-form coherence. Batch transcribe multiple files with shell glob patterns.

Startup & Modes(20)

Command	Level	Example
`whisper --help` Show Whisper CLI help	Basic	`whisper --help`
`whisper audio.mp3` Transcribe audio file to text (basic usage)	Basic	`whisper recording.mp3`
`whisper audio.mp3 --model` Specify Whisper model size	Basic	`whisper recording.mp3 --model medium`
`whisper audio.mp3 --model tiny` Use smallest model (fastest, least accurate)	Basic	`whisper recording.mp3 --model tiny`
`whisper audio.mp3 --model large` Use largest model (slowest, most accurate)	Basic	`whisper recording.mp3 --model large`
`whisper audio.mp3 --language` Specify audio language (improves accuracy)	Intermediate	`whisper chinese_speech.mp3 --language zh`
`whisper audio.mp3 --task transcribe` Transcription task (default, keeps original language)	Basic	`whisper french_audio.mp3 --task transcribe`
`whisper audio.mp3 --task translate` Translation task (translates any language to English)	Basic	`whisper japanese_audio.mp3 --task translate`
`whisper audio.mp3 --output_format` Specify output format (txt/vtt/srt/tsv/json/all)	Intermediate	`whisper recording.mp3 --output_format srt`
`whisper audio.mp3 --output_dir` Specify output directory	Basic	`whisper recording.mp3 --output_dir ./transcripts`
`whisper audio.mp3 --device` Specify inference device (cpu/cuda)	Intermediate	`whisper recording.mp3 --device cuda`
`whisper audio.mp3 --fp16` Enable half-precision inference (GPU acceleration)	Intermediate	`whisper recording.mp3 --fp16 True`
`whisper audio.mp3 --verbose` Show verbose output (with timestamps and confidence)	Basic	`whisper recording.mp3 --verbose True`
`whisper audio.mp3 --temperature` Set sampling temperature (0-1, default 0)	Intermediate	`whisper recording.mp3 --temperature 0.2`
`whisper audio.mp3 --initial_prompt` Set initial prompt (guides style or corrects terms)	Expert	`whisper medical.mp3 --initial_prompt 'Use medical terminology.'`
`whisper audio.mp3 --condition_on_previous_text` Use previous text for better coherence	Intermediate	`whisper long_lecture.mp3 --condition_on_previous_text True`
`pip install openai-whisper` Install Whisper	Basic	`pip install openai-whisper`
`pip install openai-whisper[audio]` Install Whisper (with audio dependencies)	Basic	`pip install openai-whisper[audio]`
`whisper --model small audio.mp3` Transcribe with small model (speed-quality balance)	Basic	`whisper --model small meeting.mp3 --output_dir ./notes --output_format txt`
`whisper dir/*.mp3` Batch transcribe all audio files in a directory	Intermediate	`whisper podcasts/*.mp3 --model base --output_dir ./transcripts`

FAQ

How do I install Whisper?

Which model size should I choose?

How do I transcribe non-English audio?

How do I generate subtitles (SRT/VTT)?

Can I use Whisper with GPU?

This cheatsheet is compiled from official tool documentation. Last updated: 2026-07-20.