New Open ASR Models Challenge Whisper as Accuracy Gap Narrows

OpenAI’s Whisper has been the default open automatic speech recognition model since its 2022 release, but new systems from IBM, Cohere and Alibaba’s Qwen have narrowed the accuracy gap among leading contenders. With top benchmark word error rates now clustered near 5%, deployment choices increasingly hinge on language coverage, licensing, streaming or batch latency, and the cost of running models on private infrastructure rather than on accuracy alone.

IBM released Granite Speech 4.1 2B on April 29, 2026, posting a 5.33% average WER on the OpenASR Leaderboard. That edged Cohere’s 2-billion-parameter Transcribe, launched March 26 at 5.42%, by just 0.09 percentage point. Both carry Apache 2.0 licenses, making license fees $0. Granite covers six languages and offers a lower-latency non-autoregressive variant, while Cohere supports 14 languages and free rate-limited API trials; production Model Vault service is billed per instance-hour, with pricing undisclosed.