OpenAI's robust speech-to-text model trained on 680K hours of multilingual audio data.
Why it matters now: De facto open-source speech-to-text; multilingual, reliable, actively maintained.
Whisper is OpenAI’s speech recognition system trained on 680,000 hours of multilingual and multitask supervised data collected from the web, handling diverse audio in 96 languages with minimal fine-tuning. It’s available both as a free, open-source model under the MIT license (self-hostable) and as a paid API endpoint through OpenAI at $0.003-$0.006 per minute depending on the model variant.
The project has become the de facto standard for local speech-to-text due to its reliability, multilingual support, and reasonable performance across noisy audio. First released in September 2022, it has maintained momentum with regular updates through 2025 and 2026, including newer Python version support and performance optimizations.
While newer competitors like Voxtral (Mistral AI, February 2026) offer superior accuracy on some benchmarks, Whisper maintains an advantage in language coverage (96 languages vs. Voxtral’s 13), community ecosystem, and production maturity. The base model has remained relatively stable since 2024, with most innovation happening in variant models (turbo, distilled versions).
Python 3.8-3.12, ffmpeg for audio processing. CPU: 8 GB+ RAM. GPU optional (~20x faster); VRAM 1-10 GB depending on model size.