Saturday, November 7, 2026 · Practical AI
Videos · Narration · Stories
Making videos.
Finding a voice.
I’ve been making short videos with Codex, trying synthetic narration, and experimenting with a local version of my own voice. Here’s what we used, what needed another pass, and a few places to explore.
What we’re using
Codex + HyperFrames
I describe the video and review the changes. Codex helps build the composition in HyperFrames, where visuals, motion, narration, and captions come together.
Our setup: Codex and an installed HyperFrames workflow on my laptop.
Explore HyperFrames →ElevenLabs
We turn an approved script into audio, download the take, and place it in the video. We’ve used the selected Juniper voice; that isn’t the same as cloning my voice.
Our access: I have an ElevenLabs account and approve the use of its credits.
Explore text to speech →HeyGen
For Mazie’s birthday, we used her photos to generate a short motion clip. HyperFrames then assembled that clip with narration and captions.
Our test: One approved five-second generation. Availability and credits depend on your account or connected access.
Browse HeyGen Academy →One project, several tools
Mazie’s birthday story
Four photos became a short birthday video with movement, narration, captions, and a little smile at the end.
How the pieces fit: HeyGen supplied the motion. ElevenLabs supplied the narration. HyperFrames brought the pieces together into a vertical video.
What needed work: Our first mouth-animation test in HyperFrames wasn’t convincing. The HeyGen clip resembled Mazie, but her mouth movements still did not align with the narration in the final video. More editing would have been needed.
The review is part of making it
In another video, we illustrated a conversation about recurring payments. We timed the phone visuals to the dialogue, then discovered an audio cutoff in an export. We repaired the audio and reviewed another version.
Our short video checklist
- Does the script say what I intended?
- Do the visuals match the narration?
- Are names and words pronounced clearly?
- Are captions readable and correctly timed?
- Does the exported video play all the way through with complete audio?
A separate local experiment
Trying my own voice with Chatterbox Nano
We installed Chatterbox Nano on my laptop and used a recording of my voice to generate a short, approved script. This used an existing speech model; we didn’t train a new model from scratch.
A more expressive reference recording gave us a better result.
Short narration has worked in our setup. Long-form consistency is still untested, and this does not change the voice in a live ChatGPT conversation.
What happened: The first sample sounded close to me, but needed more expression. I supplied another recording. We kept the script and settings the same for comparison, and I felt the third take “nailed it.”
The steps we took
- Installed the software and downloaded model files.
- Checked that the setup could generate audio.
- Used a clean recording of my own voice and an approved short script.
- Generated a WAV file locally, with outgoing connections blocked in the generation process.
- Listened, changed the reference recording, and compared another take.
The improved test made about 16 seconds of audio in roughly a minute overall. That is a result from our laptop, not a speed promise for other setups.
Setup required technical help. Our voice recordings and generated samples stayed on the laptop.
Explore the Chatterbox Nano model →Try one small idea
Choose three photos for a short story, explain one task, or write a brief introduction. Draft a few sentences and decide which visuals belong with them.
Start with your own images and voice, or material you have permission to use. Check account requirements and credits before generating. When sharing, make clear when a voice or scene is synthetic.
Bring your experiment to the group: what worked, what surprised you, or what you’re still figuring out.