OpenAI Whisper Small

A 244-million parameter speech recognition model balancing high-speed transcription with robust accuracy across 99 languages.

Whisper Small hits the efficiency sweet spot for production-grade speech-to-text. It utilizes a Transformer sequence-to-sequence architecture trained on 680,000 hours of supervised data. While the Large-v3 model targets maximum precision, the Small variant processes audio roughly 6x faster on standard GPUs (like the NVIDIA T4) without a significant drop in reliability. It is the primary choice for developers building low-latency tools: live captioning, rapid indexing, and real-time translation pipelines.

https://github.com/openai/whisper