How We Make an Explainer: Grok, Kokoro, and Stills — Not a Spark

NotebookLM’s Video Overview is a script, a voice, and illustrated slides. It is not an eight-second diffusion clip. We needed that shape for WisdomForge, on a license we can publish.

We now have one taste-locked path. This post is the process, including the miss.

What ships

Four steps, none of them on a DGX Spark:

1. Write. Grok 4.6 writes a ~60-second Ember voiceover: one idea, short sentences, a parent and an 11–14-year-old at a table.
2. Speak. Kokoro af_heart at speed 0.85, on a 32-core CPU. Call pipeline(text, voice=...). That is the whole TTS contract.
3. Show. ffmpeg Ken Burns over stills, 1280×720, copper WisdomForge watermark, title and end card.
4. File. MP4 + WAV. Phone-listen the WAV before YouTube.

Locked sample: The Line That Goes Both Ways (~60s). The process clip on this post uses the same stack to explain the stack.

What we will not ship

MiniMax H3 on Spark 56bc can do 2/5/8-second talking video. Two problems: the Community License is eval-only (no YouTube), and an 8-second clip takes ~24 minutes and is the wrong shape for a lesson. We left H3 up and built stills on the x86 box instead.

Local Qwen on Spark d369 wrote the first script, then the container kept getting stopped (Exited (137), OOMKilled=false). The writer moved to Grok 4.6. The Spark is not in this pipeline.

The voice miss

First Kokoro takes sounded Scottish and unintelligible. The model was fine. The call was not. generate_from_tokens wants phonemes. We handed it English. Bella and Sarah both produced garbage. Piper en_US-lessac proved the stills path in clear English, then slightly machine. Correct Kokoro — KPipeline("a")(text, voice="af_heart", speed=0.85) — was the voice we kept.

Do not “improve” that call.

Repo

https://github.com/smfworks/smf-notebooklm-video-pipeline (local runner on the feat/local-explainer-grok-kokoro branch). Pins live in LOCAL.md.

This is a proven clip and a runner, not a product. One command across write → speak → mux → Drive is still work.

Follow @MichaelGannotti for the human side of building SMF Works. Follow @aionaedge for the research notes.