V-GEN: An Autonomous TV News Generator Built on Google Gemini and FFmpeg

A regional TV channel ran into an acute staffing problem: its editor and camera operator both quit. To avoid losing its broadcast license, the station had to automate its airtime and content production. In the first year, the tech-support administrator combined live news segments with video presentations and slideshows pulled from local sources. But even gathering information, processing it, selecting visuals, and doing the final edit took 20–30 minutes per segment. A solution was needed.

Architecture: seven stages of automation

V-GEN combines several modern technologies into a single production pipeline.

1. Intelligent news search Classic RSS parsers break quickly whenever a site's layout changes. V-GEN uses Google Gemini (the gemini-3.8-flash model) with built-in Grounding: direct, real-time access to Google's search index. The system scans publications from the past 7–10 days for a given geography, cross-references facts from regional media, Emergency Ministry press releases, and administration posts, and extracts confirmed events.

2. Editorial rewriting and prompt engineering V-GEN's prompt enforces strict television-format requirements:

Length: strictly 800–1,000 characters including spaces (~45–60 seconds of reading time). Structure: a lead hook in the first sentence, followed by facts and details, ending with a status update or the date of the next event. Language: spoken, conversational style with no bureaucratic jargon (phrases like "for the purposes of implementation" are banned), no greetings ("Hello" is not allowed), and numerals written out as words ("the seventh of October," "thirty million rubles").

Geography and filters cover events in the city and district, regional decisions by the governor, Emergency Ministry inspections, and utility tariffs. A hard ban applies to: lost items, garages, and private classified ads.

Topic categories include incidents and safety (Emergency Ministry, traffic police, courts), utilities and the urban environment (heating, roads, beautification), social affairs (hospitals, schools), culture, sports, and event listings.

3. Speech synthesis: Google Gemini TTS Narration is voiced by Callirrhoe — a model professional TV anchor: energetic, warm, with no mechanical pauses. Alternative voices are available: Puck, Charon, Kore, Fenrir, Aoede, Leda, Orpheus, and Zephyr. The system is protected against prompt injection: only the cleaned news text, with no instructions, is fed to the TTS engine. The audio stream is normalized to broadcast standards: 48 kHz, 16-bit stereo.

4. Visuals: Smart UNCROP and Ken Burns The Smart UNCROP algorithm automatically fits any image to Full HD 1920×1080. If the aspect ratio is close to 16:9, the image is scaled and centered. If the photo is vertical or square, a blurred background is generated (the source image is stretched, blurred with a Gaussian radius of 28 pixels, and darkened by 30%), with the original sharp photo overlaid in the center.

The Ken Burns effect is implemented via sub-pixel zoom: coordinates are passed as floating-point values, and scaling is performed with bilinear interpolation. Zoom ranges from 1.0 (100%) to 1.15 (115%), frames alternate between zoom-in and zoom-out, and the final slide always shows a zoom-out.

5. Precise timing math The anchor's narration is the absolute master of the timeline. The formula: Lead-In (1 second) + Voice Duration (measured to sub-second precision) + Lead-Out (3 seconds). If the video contains 7 frames, each lasts (total time – 4 seconds) / 7. B-roll clips are looped, cropped to 1920×1080, and normalized (SAR 1:1, 30 fps, yuv420p).

6. Final edit in FFmpeg The system sets the background music volume (−5.4 dB), applies a synchronized fade-out (1.5 seconds for both video and audio), splices in the channel's on-air stinger, and overlays the logo with transparency.

The result is a finished MP4 file with a Russian-language filename ("Heating Network Repairs in Tutaev.mp4") ready for live broadcast.