Whisper Medium English

Whisper medium.en is an automatic speech recognition (ASR) model that transcribes English audio into text. It is a Transformer-based encoder-decoder (sequence-to-sequence) model trained on 680,000 hours of weakly supervised, English-only labelled data. The model is designed for zero-shot generalization, requiring no fine-tuning to perform well across diverse domains, accents, and recording conditions.

Architecture and Capabilities

Whisper medium.en contains 769 million parameters and is optimized for English speech recognition. It processes audio in 30-second segments and supports long-form transcription via chunking, enabling arbitrary-length audio processing. The model can also predict word-level timestamps.

Benchmark Performance

On LibriSpeech test-clean, the model achieves a Word Error Rate (WER) of 4.12% (official result) and 3.02% under alternative inference settings. Additional benchmark results include:

• LibriSpeech test-other: 7.43% WER
• AMI: 16.68% WER
• Earnings22: 12.63% WER
• Gigaspeech: 11.03% WER
• Open ASR Leaderboard mean WER: 8.09

Inference Characteristics

The model requires approximately 5 GB of VRAM and runs at roughly 2x the inference speed of the large-v2 variant. It supports batched inference and chunked processing for audio of arbitrary length.

Training Data and Robustness

Trained on 680,000 hours of internet-sourced audio, the model demonstrates improved robustness to accents, background noise, and technical language compared to many existing ASR systems. It achieves near-state-of-the-art accuracy in a zero-shot transfer setting without fine-tuning.