Speech entered a real donor and returned a waveform.
The experiment split one learned second projection into eight equal channel experts. Selecting all eight reproduced the donor waveform with zero maximum and RMS error.
Back to the demo chooser
ERAIS / make, listen and explore
Listen, compare and explore.
Type something and hear it spoken. Compare the voices, explore the experiments, or make a loop with the browser instrument. Start wherever your curiosity takes you.
Inside the collection
Start here / real model output
Give it your words.
A greeting. A strange little story. Something you have been meaning to say. Generate it here, then hear how the two speech routes compare.
Make something new
Your words. Spoken here.
Write a line, choose a route and hear a fresh recording.
Checking the shared ERAIS service…
What runs when you press Generate?
Both routes run through the shared ERAIS service. The Fracture route executes the two converted projections in one speech decoder block using the ERAIS native engine. Other speech components still come from LuxTTS and Vocos. The reference uses their original projections. This is an experimental conversion, with no claim that the whole speech model is native or that every voice is improved.
Hear the recorded voices.
Choose a passage, then play A and B. The model names stay hidden so you can judge the sound.
“A quiet lantern glowed beside the window while rain softened the empty street.”
“Please carry the blue folder upstairs and place it beside the wooden clock.”
“Can a careful machine explain its choices clearly when the answer is uncertain?”
“Fresh oranges, warm bread, and silver spoons were arranged across the table.”
Choose either voice. Playing one pauses the other.
Inside this comparison
These released clips use the ERAIS K7 route, which sparsifies two learned projections in one vocoder block. They demonstrate that specific conversion step. ERAIS also has native neural execution, persistent state and adaptive routing; this recording is not a demonstration of every component together. Explore the speech results.
Unified World
Follow the continuing experiment.
Open the conversation experiment and see which inputs are available. The session shows the model and current service status.
- 01Perceive
Bring observations into shared state.
- 02Adapt
Apply the capability and effort needed.
- 03Continue
Carry outcomes forward. Rest when quiet.
Fracture / learned capability, modular structure
Compare a recorded speech conversion.
Pick a recorded mode to compare the output and processing used by a converted speech component.
Measured30 August 2026CPU · LuxTTS / Vocos
Select an executed mode
Choose a recorded mode
Full bank: 8 / 8 experts selected per acoustic frame
Stored expert bankOne learned LuxTTS Vocos projection
8 / 8 activeFilled blocks show the selected count only. The measured router chose different expert identities for different acoustic frames.
- Selected projection
- 100%
- Reconstruction
- Exact component reconstruction
- Waveform correlation
- 1.000
- Spectral convergence
- 0.0000
- Whisper proxy word edits
- 0 / 12
- Human listening score
- Not measured
Methods and scope of this recorded experiment
Inspect the measured results through this interface. Donor weights, activation traces and the proprietary engine remain private.
K4 and K2 preserved every proxy-transcribed word.
K4 used 50% and K2 used 25% of that projection. Whisper reported zero word edits on the same 12-word target while the quality sweep exposed the reconstruction trade-off.
Fracture is moving through the full speech stack.
The next vertical moves routing ahead of dense expansion, converts additional learned blocks, and measures end-to-end speech quality and speed.
Next physical path · QwenSelected expert work can now be grouped before timing.
A grouped physical-dispatch path is committed and equivalence-tested for the Qwen evaluator. Full-tower quality and speed measurements remain in progress.
- LuxTTS source
8c893869c4bb- Measurement digest
8af672e6be05…- Qwen executor source
184c9afc3f5f
Bring a model workload. Explore where converted capability could be useful in your system. Talk about Fracture integration.
Bonus / browser sound playground
Interactive 48 kHz stereo loop
A fresh three-stem loop synthesized entirely in your browser.
Interactive · local browser DSP
Build a three-stem instrument loop
Change tempo, palette, or stems. Sound is created in this tab after your click; no recording, model request, or upload occurs.
- Processing mode
- Browser DSP · no network
- Loop duration
- 2.40 s
- Active stems
- 3 / 3
- Start latency
- Not started
Browser instrument: instant local synthesis with no upload or model request. For actual model-generated speech, listen to the Fracture comparison above.
Recorded model output / vision
Fixed sparse classification probes
Three predeclared site-owned cards evaluated offline; no camera or upload.
3 / 3 fixed cards visible
Intended card
Airplane
Top fixed output 74.76%
Intended card
Ship
Top fixed output 91.20%
Intended card
Truck
Top fixed output 87.34%
Measured foundation: the retained 5.22M-parameter run recorded 60.45% on its exact CIFAR-10 slice. These are fixed, offline examples; broader vision performance remains unverified.
Recorded model output / language
A verbatim CPU-only story sample
A fixed capture from the reviewed native-byte checkpoint on an owned desktop CPU.
530 UTF-8 bytes · 110 words
Once upon a time, there was a little girl named Lily. She loved to play in the garden. One day,
Once upon a time, there was a little girl named Lily. She loved to play in the garden. One day, she saw a big bird with a big smile. She wanted to see what it was. Lily said, "Lily, it's time to go home. It is not for the big boy." So, they went to the park and saw the bird. The bird was so happy and said, "Thank you, Lily. You are a good friend." Lily smiled and said, "You're welcome, it's important to be careful with the bird."
What this demonstrates: an owned desktop CPU ran the native-byte checkpoint and produced this captured story. The inspector stays entirely on-page; Story Lab unlocks live generation after its service check.
Recorded proof / symbolic reasoning
Follow the proof. Then change the facts.
Step through this recorded sensor example, then try rules, conflicting reports and calculations in the live lab.
All 4 receipt steps visible
Fact: the sensor is calibrated. Fact: the sensor is powered. Rule: anything that is calibrated and powered is ready. Question: is the sensor ready?
DecisionReady — proved from both facts.
Every step shown
- 1given
The sensor is calibrated.
- 2given
The sensor is powered.
- 3rule application
Calibrated and powered implies ready.
- 4conclusion
The sensor is ready.
Recorded validation: 30 accepted examples and 6 rejected cases. This trace is recorded; the live lab runs your edited request.