whisper

Whisper turns speech in audio files into text through a command-line tool or Python code. It supports multilingual transcription, spoken language identification, and translation into English, with several model sizes offering speed and accuracy tradeoffs.

Key features

• Multilingual speech recognition
• Translation of non-English speech into English using multilingual models
• Spoken language identification
• Six model sizes with different memory requirements
• English-only variants for tiny, base, small, and medium models
• Command-line transcription of FLAC, MP3, and WAV files
• Python transcription API with lower-level detection and decoding functions
• Optimized turbo model for faster transcription

Why it helps in Pakistan

• Turn recorded speech into text for review and reuse.
• Choose a model that fits your available GPU memory and accuracy needs.
• Add speech recognition to Python applications.
• Translate supported non-English recordings into English.
• Use and adapt both the code and model weights under the MIT licence.

• 01

  Pakistani education teams can transcribe recorded English lectures into text for study material.

• 02

  Local media teams can turn recorded interviews in supported languages into draft transcripts for editorial review.

• 03

  Pakistani software developers can add audio-file transcription to Python business applications.

• 04

  Export businesses can translate recorded non-English speech in supported languages into English for staff review.

Hardware

Approximate model VRAM: tiny/base ~1 GB, small ~2 GB, medium ~5 GB, large ~10 GB, turbo ~6 GB.System RAM and other hardware requirements: Check the official docs.

Operating system

WindowsmacOSLinux: Ubuntu, Debian, and Arch installation examples are provided.

Skills

Command-line package installationBasic Python for application integration

Accounts needed

Check the official docs

Commands are copied from the project's README and official docs.

1. 1

   Prepare Python

   The README expects compatibility with Python 3.8–3.11 and recent PyTorch versions. Check the official docs for installing this environment.

   Not fully covered in the README. Check the official docs before running this step.

2. 2

   Install ffmpeg

   On macOS with Homebrew, run brew install ffmpeg. On Windows, use Chocolatey with choco install ffmpeg or Scoop with scoop install ffmpeg.
3. 3

   Install Whisper

   Install or update the release with pip install -U openai-whisper. Alternatively, install the latest repository commit with pip install git+https://github.com/openai/whisper.git.
4. 4

   Resolve optional build dependencies

   If tiktoken has no pre-built wheel for your platform, follow the linked Rust getting-started guide. For a No module named 'setuptools_rust' error, run pip install setuptools-rust.
5. 5

   Transcribe audio

   The documented example is whisper audio.flac audio.mp3 audio.wav --model turbo. Substitute your own audio filenames. The turbo model requires approximately 6 GB VRAM.
6. 6

   Review available options

   Run whisper --help to view available options. For speech translation into English, choose a multilingual model other than turbo.

• Python applications

  Import whisper, load a model with whisper.load_model(), and call model.transcribe() to obtain text.

• PyTorch

  Whisper uses PyTorch; the README expects compatibility with recent versions.

• ffmpeg

  Required command-line dependency installed through your system package manager.

• tiktoken

  Provides the tokenizer implementation used by Whisper.

• GitHub
• Model card
• Supported language list
• Colab example
• Community examples
• Rust installation guide
• MIT licence