openai-whisper mit Homebrew, Nix, MacPorts installieren

Projektlinks

• Homepage
• Upstream-Dokumentation
• Paketmanager-Seite

Befehle und Aliase

• whisper

Projektgeschichte

The public repository was created on September 16, 2022, around OpenAI's release of Whisper as code plus model weights under the MIT license. The accompanying paper, submitted to arXiv on December 6, 2022, framed Whisper as a robustness-first speech-recognition system trained at web scale rather than a narrowly benchmark-tuned ASR model.

Whisper's design used a single multitask token interface for speech recognition, speech translation, spoken-language identification, and voice activity detection. That made the package unusually self-contained for an ASR release: users could install the Python package, ensure ffmpeg was available, choose a model size, and transcribe local audio without training a model or calling a hosted API.

Adoptionsgeschichte

The project became a major reference point for local and open speech transcription because OpenAI released both inference code and model weights. The model family also fed a wider ecosystem of ports, front ends, batch transcribers, and integrations, including downstream implementations optimized for smaller devices or different runtimes.

Homebrew, MacPorts, and Nix packaging made the command-line workflow convenient for Unix-like systems. In package-nerd terms, `openai-whisper` sits at the intersection of Python packaging, system multimedia dependencies through ffmpeg, and model artifact distribution.

Wie es verwendet wird

Developers use the `whisper` command to transcribe audio files, specify model sizes, set input languages, and request translation into English. Python users load a model with `whisper.load_model()` and call `transcribe()` for scripts, pipelines, notebooks, and media-processing jobs.

The README documents six model-size families plus English-only variants for some sizes, with memory and speed tradeoffs. That packaging shape matters because installing the package is only one part of operating it; users also choose model weights, hardware, ffmpeg availability, and task settings.

Warum Paket-Nerds sich dafür interessieren

Whisper is a rare package-manager entry that installs a small CLI front end for very large model artifacts. It demonstrates how ML tools blur the usual package boundary: the executable is ordinary Python software, while most practical value comes from downloaded weights and GPU/CPU runtime behavior.

It also made speech recognition feel like a normal developer dependency. For many users, `brew install openai-whisper` or `pip install openai-whisper` turned multilingual ASR from a cloud service integration into a local command-line primitive.

Zeitleiste

• September 16, 2022: the GitHub repository was created. December 6, 2022: the Whisper paper was submitted to arXiv. Later releases added model updates such as large-v3 and turbo, while keeping the package centered on the same CLI and Python API.

Related projects

• Whisper depends on PyTorch and ffmpeg in normal use and uses OpenAI's tiktoken tokenizer. Its adoption also encouraged alternate runtimes and ports, most famously C/C++ implementations that target smaller machines and offline workflows.

Quellen

• https://api.github.com/repos/openai/whisper
• https://arxiv.org/abs/2212.04356
• https://cdn.openai.com/papers/whisper.pdf
• https://formulae.brew.sh/formula/openai-whisper
• https://github.com/openai/whisper