Qwen in production: auto-generating 79 baseball highlight clips for a Korean broadcaster at the 2026 Asian Games

#228
by abafffff - opened

Qwen in production: auto-generating 79 baseball highlight clips for a Korean broadcaster at the 2026 Asian Games

We used Qwen models end to end to build an automatic highlight-clip system for SBS, a Korean broadcaster, and ran it on live operations during the 20th Asian Games (Aichi-Nagoya 2026). Sharing what worked, what broke, and two questions we'd love input on.

Results (baseball, 5 Team Korea games)

Metric Value
Clips generated automatically 79, published on SBS play "AI PICK"
Operator edits needed (first game, KOR vs TPE, Sep 21) ~10% (missing segments, boundary fine-tuning)
Manual baseline 18 min per clip (24 h for 79 clips, measured)

The system does not replace editors: AI drafts the clips, operators review and publish. Full quantitative evaluation (tIoU-based precision/recall/F1) over all 79 clips is in progress.

Service page: https://programs.sbs.co.kr/sports/ag2026/clips/90399

Pipeline

  1. Preprocess - FFmpeg (NVENC/NVDEC) splits video and audio; pyannote handles diarization.
  2. Speech - Qwen3-ASR-1.7B transcribes, Qwen3-ForcedAligner-0.6B adds word-level timestamps.
  3. Transcript cleanup - a Qwen LLM [Qwen3.8-27B-FP8] corrects context errors and hallucinations per dialogue segment.
  4. Scenes - PySceneDetect finds cuts; one representative frame per scene plus STT/OCR/subtitles go to a Qwen VLM for a scene description.
  5. Index - scene summaries embedded with Qwen3-Embedding-4B into Milvus (1024-d, HNSW, cosine); broadcast metadata (game, inning, matchup) kept as filter fields.
  6. Clip - RAG over the top-k chunks lets the LLM propose titles and IN/OUT points; only segments with 0 <= IN < OUT <= source duration go to FFmpeg; operators review.

Lessons learned

  • Skip ASR when subtitles exist. Where broadcast subtitles are available we bypass speech recognition entirely: cheaper and fewer errors.
  • Don't trust first-pass ASR on sports audio. Crowd noise and commentary overlap, so every segment goes through hallucination gates and LLM correction before it counts as dialogue.
  • Fewer frames, same understanding. In our PoC on 6-second Korean broadcast clips, sampling at 0.5 fps instead of 30 fps was 2.4x faster and used 6.89x fewer tokens. In production we send one representative frame per scene.
  • Visual cuts are not event boundaries. We merge adjacent scenes by dialogue continuity and game context (scoreboard OCR: inning, outs, count, score) before retrieval.

Question for the community

  • Frames for scene description: for Qwen3.8 as a VLM, is there guidance on how many frames per scene (or which resolution) gives the best accuracy/cost trade-off for short broadcast scenes?

Any pointers, or experiences from similar media workloads, are very welcome.

Links

Built by Solbox and SBSi under the 2026 NIPA Open-Source AI/SW Development and Utilization Support Program (PJT-26-040089). Thanks to the Qwen team for the open weights.

Sign up or log in to comment