FFmpeg 9.0 ‘Lei’: Honoring a Chinese Developer Ten Years After His Passing

Overview

FFmpeg is a multimedia framework that can decode, encode, transcode, mux, demux, stream, filter, and play audio‑video, as well as handle subtitles and metadata.

Command‑line tools and libraries

Executable tools: ffmpeg, ffprobe, ffplay. Development libraries: libavcodec, libavformat, libavfilter, etc., which are embedded in browsers, players, streaming software, editing suites, and cloud‑based batch transcoding services.

Historical milestones

2000 – Fabrice Bellard initiates the FFmpeg project.

2011 – Governance dispute leads to the Libav fork.

2012 – FFmpeg 1.0 released, establishing a stable release cadence.

2016 – 3.0 “Einstein” adds hardware acceleration and modern format support.

2022 – 5.0 “Lorentz” performs large‑scale API cleanup.

2025 – 8.0 “Huffman” introduces a Whisper speech‑recognition filter.

2026 – 9.0 “Lei” adds ONNX Runtime DNN backend and GPU Execution Provider support.

Version codenames

3.0 – Einstein (Albert Einstein)

4.0 – Wu (Wu Jianxiong)

5.0 – Lorentz (Hendrik Lorentz)

6.0 – Von Neumann (John von Neumann)

7.0 – Dijkstra (Edsger Dijkstra)

8.0 – Huffman (David Huffman)

9.0 – Lei (Lei Xiaohua)

AI preprocessing role

Video‑model training – decode videos, unify resolution/frame‑rate/pixel format, extract frames, strip audio.

Speech‑model training – extract audio, convert to mono, resample to 16 kHz or other rates.

Video understanding – batch‑generate keyframes, short clips, and metadata for visual or multimodal models.

AI video generation – assemble generated frames or clips into video, add frame‑rate, encoding, audio, subtitles, and cover art.

Digital dubbing – align audio‑video length, mix audio, adjust sample rate, mux subtitles, produce final output.

General AI users – compress uploads, trim clips, extract audio, generate GIFs, burn subtitles, convert formats.

Integration with Whisper and ONNX Runtime

OpenAI’s Whisper calls load_audio, which internally invokes FFmpeg to decode various audio inputs, mix to mono, and resample to 16 kHz. FFmpeg 8.0 added a Whisper audio filter via whisper.cpp. FFmpeg 9.0 further added an ONNX Runtime DNN backend and GPU Execution Provider support, enabling image‑processing models to run directly in FFmpeg filter chains.

Case study: ringtone conversion client (2011)

A WinForms client built in C# used FFmpeg to address two user needs:

Convert audio/video formats.

Extract an arbitrary time segment.

The UI (illustrated below) leveraged FFmpeg’s command‑line interface to perform format conversion and segment trimming without requiring a separate library.

Community contribution: Lei Xiaohua

In March 2026, community member Liu Qi proposed naming the 9.0 release “Lei” to honor Lei Xiaohua, a Chinese media‑technology Ph.D. student who, between 2013‑2016, authored extensive FFmpeg API documentation, codec and format analyses, and practical examples that helped Chinese developers enter the audio‑video field. The proposal was accepted and the codename applied to the 9.0 release.

Conclusion

As large‑model AI pipelines require decoded, resized, and sampled tensors, FFmpeg functions as the essential preprocessing layer, handling decoding, scaling, frame extraction, and audio preparation. New filters such as Whisper and ONNX Runtime integration extend FFmpeg’s role from a background utility to a direct component of AI model pipelines, underscoring its continued relevance in the fast‑evolving AI era.