Whisper Small English

Whisper-small.en is an automatic speech recognition (ASR) model that transcribes English speech into text. It is a Transformer sequence-to-sequence model trained on 680,000 hours of weakly supervised audio data, giving it a strong ability to generalize to new domains without fine-tuning. The model achieves a word error rate (WER) of 3.05% on the LibriSpeech test-clean benchmark, demonstrating high accuracy on standard English ASR tasks. It is designed for single-language English recognition and tends to outperform its multilingual counterpart for English transcription, though the gap is smaller at the small size. Key features include support for long-form transcription via chunking (up to arbitrary audio lengths) and the ability to predict sequence-level timestamps. The small.en variant has 244 million parameters, requires approximately 2 GB of VRAM, and offers roughly 4x faster inference than the large model on an A100 GPU. The model is robust to accents, background noise, and technical language, but may occasionally produce hallucinations or repetitive text, and its accuracy varies across languages (for multilingual versions) and demographic groups, as noted in the original study. The model is hosted as a managed API on gigarouter, eliminating the need for local installation or GPU management. Users simply call an OpenAI-compatible endpoint to transcribe English audio with production-ready performance.