openai/whisper

Whisper is OpenAI's open-source speech recognition model -- 97,297 GitHub stars as of 2026. Large-scale weak supervision on 680k hours of audio delivers robust results offline.

What Whisper does:

• •Multilingual transcription Converts speech to text in 99 languages, even with accents or background noise.
• •Speech translation Translates audio from any language directly to English text.
• •Auto language detection Identifies spoken language without manual input.
• •Flexible model sizes Pick tiny (39M params, 1GB VRAM) for speed or large-v3 (1.5B params, 10GB VRAM) for precision.
• •Word-level timestamps Provides exact timing for words on large-v2 and later models.
• •Wide audio support Processes MP3, WAV, M4A, and more files directly.
• •Full offline use Keeps data private with zero cloud dependency or API fees.

Whisper ecosystem:

• •Python integration Install via pip and call from scripts or CLI.
• •Hardware optimization Accelerates on NVIDIA, AMD GPUs, or Apple Silicon.
• •Community extensions Ports to mobile, web assembly, and faster inference engines.

Getting started:

Run pip install -U Whisper then whisper your_audio.mp3 --model base. Use --language for non-auto detect. Check docs at github.com/openai/whisper. Pairs with ffmpeg for best results.

Limitations:

Large models demand 10GB VRAM and take minutes per hour of audio on consumer hardware. Lacks real-time streaming -- batch process only. Setup requires Python 3.9+ and ffmpeg install.