LT Player
v3 showcase
Media Processing Engine
MPE smoke run 21 September 2026
First complete run of the inherited pipeline on the LT GPU server: three recorded chapters through transcription, fraud check, metadata, silence trimming, zoom, eye-level framing, background replacement, stitching and the audio pass, 27 minutes end to end. Both players stream an HLS ladder (1080p, 720p, 480p) with an MP4 fallback. The context panel shows the title, summary, key concepts and research field the pipeline suggested. Compare the audio: the left player is the video before the noise reduction step, the right one is the final video as the pipeline delivers it today.
Processing time
Steps in pipeline order, one job of three chapters
| # | Step | Agent image | Runs | Per run | Total |
|---|---|---|---|---|---|
| 1 | Audio extraction per chapter | ffmpeg | 3 | 5 s | 14 s |
| 2 | Language detection | whisper | 3 | 9 to 10 s | 29 s |
| 3 | Transcription | whisper | 3 | 13 s | 40 s |
| 4 | Fraud detection (Claude Opus 5) | rubyllm | 1 | 21 s | 21 s |
| 5 | Suggest metadata (Claude Opus 5) | rubyllm | 1 | 20 s | 20 s |
| 6 | Silence trimming | silero | 3 | 14 to 16 s | 43 s |
| 7 | Zoom timeline (Claude Sonnet 5) | rubyllm | 3 | 24 to 30 s | 78 s |
| 8 | Eye-level framing | dlib | 3 | 24 to 26 s | 73 s |
| 9 | Background mask | video-matting | 3 | 223 to 251 s | 700 s |
| 10 | Background adjust | video-matting | 3 | 87 to 97 s | 270 s |
| 11 | Stitching with intro, wipes and outro | ffmpeg | 1 | 41 s | 41 s |
| 12 | Audio extraction of the stitched video | ffmpeg | 1 | 5 s | 5 s |
| 13 | Noise reduction | noisereduce | 1 | 11 s | 11 s |
| 14 | Audio replace | ffmpeg | 1 | 5 s | 5 s |
| Downloads from the bucket, volume setup | minio-client | 12 | 5 s | 60 s | |
| Uploads of every intermediate to the bucket | minio-client | 39 | 5 s | 210 s | |
| Total, wall clock, 81 steps | 27 min 3 s | ||||
Measured on 21 September 2026 on the LT GPU server (RTX 4000 SFF Ada). Background mask and adjust take 60 percent of the run. Every step also carries about 4.5 seconds of orchestration overhead, and the transfers add 17 percent; both go away once the job runner on the server replaces the laptop-driven controller.