Four pairs · about five minutes
Can you hear a meaningful difference?
Compare the qualified LuxTTS control and ERAIS K7 speech route with their roles hidden. Rate both clips, choose A, B, or a tie, then send one response.
Why this set is open
Machine-screened, then heard by a human.
All eight clips passed the retained waveform and intelligibility gates. Both routes transcribed with zero word errors on these four prompts. The first complete operator review called the result good overall, with slight electronic distortion still worth measuring.
K7 sparsifies two learned projections in one vocoder block. It does not represent 87.5% of LuxTTS or prove whole-model efficiency.
- Prompts
- 4
- Required decisions
- 8 ratings + 4 preferences
- Speech recognition
- 0% WER · both routes
- Mean pace
- 177.4 words/min
- ERAIS K7 STOI
- 0.9966
- Collection limit
- 500 complete responses
Before you begin
A small product-feedback task.
Taking part is voluntary, unpaid, and for adults 18 or older. Your choices help prioritise speech-quality work. This convenience sample is not a formal research study and cannot establish population preference, scientific MOS, model superiority, or production readiness.
Nothing is sent until you press Submit. You can stop by closing the page beforehand. Once submitted, the pseudonymous response cannot be linked back to you, so individual withdrawal is not possible.
What we store
One accepted record contains four A/B/tie choices, eight 1–5 ratings, the exact study identity, a random HMAC-protected pseudonym, and acceptance/expiry times. It is deleted after 30 days.
NeuroForge does not ask for or retain your name, email, account, demographics, free text, IP address, playback history, listening time, or per-clip interaction telemetry with the response. Cloudflare still processes ordinary request and anti-abuse signals to deliver and protect the page.
Collection closes automatically at 500 complete responses. Incomplete browser choices are never collected.
Questions: privacy@neuroforge.io · Privacy notice
Exact release identity
The audio and collection contract are hash-bound.
Every WAV filename is its SHA-256 identity. The release also binds the machine qualification, operator review, source/data rights, and bounded collection authority.
- Study SHA-256
bc0841031cecc1abbecdb5a0b0b9f49a5e6c9647df7c32e072f3a024c4dcdd27- Qualification
d4e018132b5b86fac5f7a3f75a05e3bcf76e9b87e8a39bf4dc519389e6733686- Operator review
f68f9142e5535d6656e57358fad960ffd754daff4b6beacacbc4be0574c3ba23- Rights receipt
428e7016d8eba3b9a918b3a7f40c460dde2159aa56f30c8c90139d73fa7a5b21- Collection authority
5eb294255b4cb3dca03c1eb119cef20878f113a264ad8d09deea599c7342baaa- Published audio
- 8 WAVs · 5.0 MB
Sources and rights
Source attribution and release terms.
The reference voice comes from LibriSpeech (OpenSLR SLR12), prepared by Vassil Panayotov with assistance from Daniel Povey and licensed under CC BY 4.0. No endorsement is implied.
LuxTTS and upstream ZipVoice are identified as Apache-2.0 projects in the retained rights receipt. The generated clips are published only for this bounded feedback task.
The A/B roles stay concealed while collection is open. Model weights, private traces, checkpoints, and training material are not distributed by this page.