Eric Raum (@ericraum) header
Eric Raum (@ericraum) profile picture

Copper Flute, DTC, creating things, storytelling

Minneapolis, MN
Joined June 2007
Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)Photo by Eric Raum (@ericraum)
I liberated my smart workout mirror from the grips of subscription based software… to turn it into a giant family control center
3
300
Doing his best Sygyt throat singing… added vocal fold modeling / hat
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
1
211
An ancient musical mystery… solved?
1
1
3
435
Accordion and Tsuur and Cello
233
The problem with chasing “feels like talking to a person”… say I present you with two audio clips. One is a cheerful human sounding voice, one is a bored, apathetic, disinterested voice. Which is more like talking to a person? What about a person you would want to talk to?
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. block.xyz/dessa
2
1
432
But the question to get to is: “Does this feel like talking to a person you would WANT to talk to?” Not many models clear that bar
16
Most would say, “I don’t think a robot would choose to sound apathetic and kind of dumb, so this feels more like talking to a person, I guess” Do that enough times and you have clear data showing how human like it is to sound low energy and disinterested. OpenAI did the same.
185
I remember the wife scream-crying as we loaded up her furniture. “We built this house as our forever home, and he is tearing our lives apart and moving us away from everyone we love just to make more money”
chat is 1M TC at OpenAI worth moving to SF for?
1
1
240
I was working for a long distance moving company, a flexible job for a young guy in a band. We were packing up the house as she was breaking down. An experience that changed my perspective on money forever.
1
1
158
Transformed Minecraft Dungeons into an educational reading game for my son. This being the only video game he has access to, his reading has never been better…
2
192
Amazing how fast the magic normalizes. Had Codex open the Fusion360 file of a machine I’ve been building, look for any holes & use their size to build a fastener purchase list… Then build a LEGO style assembly guide.
1
1
2
206
It correctly assumed a lot about the machine that wasn’t made explicitly clear. Very impressive combination of technical ability and reasoning.
116
Here’s David, my Operations Manager. A Mac Mini and the device I built, combined with a voice pipeline and operations tracking. I’m on track to triple my last holiday season output.
1
1
140
From the bottom… originally an orange color, like the button… but it just didn’t look right.
123
The biggest problem with AI movies: BAD ACTING You get “okay” performances even with the best prompt. Uncanny valley stuff With this workflow, we are one step closer to COMPLETE physical and vocal acting transfer to a character. No blender required. Kling 3, Seedance 2.5
2
1
7
1,321
You can find the workflow in my last post, BUT… One thing that is easy to miss: set your phone / camera to film at the same frame rate as your final movie (probably 24) or you will end up with some frame issues. You can see that here before she counts down. There’s a weird frame jump from all the frame rate conversions through the process
1
487
Step 3: use Kling 3.0 motion control to swap the motion. Select use the background from the image. Step 4: use Seedance 2.5 edit to change the audio of her voice. This is the weakest link in this workflow and there is probably a better way. The goal, though, is to keep the IDENTICAL acting delivery, emotion, inflection, timing, etc
1
65
Where do we go from here? This workflow “works” but it isn’t ideal for the main reason… one dependable, repeatable voice isn’t available, and that needs to be locked down per character. It is a reason you don’t see much dialogue in AI videos. When this gets solved, things accelerate
35