Whisper

Whisper is OpenAI's open source automatic speech recognition (ASR) model, released in September 2022, and it quickly established itself as one of the most capable and widely deployed speech recognition systems available. Its combination of strong accuracy, broad language support, and open source availability has made it a foundational component in many AI applications, transcription services, and developer projects.

The model was trained on 680,000 hours of multilingual audio from the internet, which produced exceptional breadth across languages, accents, and audio conditions. Whisper handles accented speech, background noise, and technical vocabulary significantly better than many specialised commercial transcription services, particularly for languages outside the major Western European group.

Whisper comes in multiple model sizes: tiny, base, small, medium, large, and the more recent turbo variants. Smaller models run faster on consumer hardware with lower accuracy; larger models produce better transcription at higher computational cost. The large-v3 model is the most accurate and is used in most production applications where quality matters.

For developers, Whisper represents a free, high-quality alternative to commercial speech APIs like Google Speech-to-Text or AWS Transcribe. Self-hosted deployment keeps audio data on the developer's own infrastructure without paying per-minute API fees, which is a significant cost advantage for high-volume transcription workloads.

Whisper is accessible through the OpenAI API at $0.006 per minute for users who do not want to manage their own deployment. Many third-party platforms including Groq (which runs Whisper at very high speed) and Replicate also provide hosted Whisper access.

Limitations include real-time performance, which is challenging with larger models, and word-level timestamps that require additional tooling on some model versions. For very low-latency real-time transcription, purpose-built streaming systems may perform better.