Whisper Inferencev5 1

Whisper Inferencev5 1

MLCommons released the Whisper Inference v5.1 benchmark in September 2025, the latest standardized evaluation of inference performance for OpenAI Whisper speech-to-text models. The benchmark aims to provide fair, comparable performance data for the industry, helping developers optimize deployment.

Key Updates and Model Support

The v5.1 version introduces the Whisper large-v3 model, while retaining variants such as tiny, base, small, medium, large-v1, and large-v2. The dataset is expanded to Multilingual LibriSpeech, containing English and multilingual audio with a total duration exceeding 1000 hours.

Key Metrics

  • RTF (Real-Time Factor): Ratio of inference time to audio duration, lower is better.
  • WER (Word Error Rate): Measures accuracy.
  • Primary score: RTF @ target WER (e.g., RTF@WER20%), balancing speed and accuracy.
  • Energy efficiency metric: RTF × Power (Watts).

Test Hardware and Frameworks

Support for CPUs (x86, ARM, e.g., Intel Xeon, AMD EPYC, Apple M series), GPUs (NVIDIA A100/H100, AMD MI300X, Intel Gaudi), and edge devices. Frameworks include ONNX Runtime, TensorRT, SGLang, and others.

Top Performance Results

Server-side top submissions:

HardwareModelRTF@WER20%Submitter
NVIDIA H100 SXMlarge-v30.008NVIDIA
AMD MI300Xlarge-v30.012AMD
Intel Xeon 6medium0.15Intel

Edge highlight: Apple M3 Max achieves 0.05 for RTF@WER20%, with best energy efficiency.

Leaderboard and Submission Guidelines

Full results available at MLCommons official website. Developers can submit via the MLPerf Inference v5.1 rules, emphasizing end-to-end performance, including pre-processing and post-processing.

This benchmark highlights the advantages of GPUs in high-accuracy STT tasks while driving low-power optimization, accelerating the deployment of AI voice applications.

This article is from MLC blog, translated in full by Winzheng (winzheng.com). Click here to view the original When republishing the translation, please credit the source. Thank you!