OpenAI open-sourced Whisper under the MIT license, demonstrating that weak supervision across 680,000 hours of multilingual web audio eliminates the need for domain-specific acoustic fine-tuning and establishing the default open-weight ASR backbone.
Whisper & Speech Recognition (ASR)
Also known as: Whisper · OpenAI Whisper · Whisper large-v3 · Distil-Whisper · ASR
Plain English
OpenAI's open-weight speech recognition model family that turns spoken audio in nearly a hundred languages into accurate text transcripts and English translations.
Technical Overview
Whisper processes 30-second log-Mel spectrogram windows through a convolutional stem and an encoder-decoder Transformer trained with multitask special tokens for transcription, voice activity detection, language identification, word timestamps, and X -> en translation. Speculative Distillation & Synthetic Data variants (Distil-Whisper and large-v3-turbo) prune decoder layers from 32 to 2–4 for 5x+ lower latency.
Timeline (4)
2022
Release2023
ReleaseOpenAI released Whisper large-v3 trained on 5 million hours of weakly and pseudo-labeled audio with 128 Mel frequency bins, while Hugging Face published Distil-Whisper showing that pruning the decoder while keeping the encoder frozen cuts latency by 6x and enables speculative decoding for ASR.
2024
ReleaseOpenAI released Whisper large-v3-turbo (809M parameters), reducing the
large-v3decoder from 32 layers to 4 layers to deliver over 5x faster transcription on whisper.cpp and cloud endpoints with negligible accuracy loss.2025
MilestoneElevenLabs launched Scribe v1 as a commercial challenger to Whisper large-v3, lowering multilingual word-error rates across 99 languages while adding native speaker diarization and character-level timestamp alignment.
See also (9)
Open full graphWhisper & Speech Recognition (ASR)