Gradium's New Default Reaches the Top Group
Gradium repointed the alias behind their default text-to-speech model on 31 August. We re-measured it on the same eight English sentences and the same rig: naturalness ~1585, into the group at the top of our English board — the group our own human panel cannot separate.
On 31 August, Gradium repointed the alias default to a new text-to-speech model.
Nothing changed on our side. Same model string, same request, same 200 — so our row kept publishing a figure fitted on software that had stopped serving. We found out because Gradium sent us a message. An alias is a silent expiry date on every number you have published about it.
So we re-measured, and the new default is a clear step up.
Into the top group
Same eight English sentences, same rig, same day: ~1585, up from 1448. That puts Gradium in the group at the top of our English board, a group our own human panel cannot separate.
It scores better on seven of the eight sentences. Three control systems re-synthesised the same minute moved by −0.05, −0.15 and +0.24 on the same scale, so nothing in the pipeline is doing the work.
The synthetic edge is gone, which is the change we had asked for. In July we told Gradium their voice read a little robotic, and backed it with the cleanest signal in our set: lowest jitter, highest harmonics-to-noise ratio, which is usually what “robotic” measures as. On the same analyser today they sit mid-pack, with more pitch variation across a sentence than ElevenLabs v3.
Hear it
Same sentence, previous default and new default.
Previous default
New default
Published copies are volume-matched; the systems got the originals.
What we did not measure
Gradium’s notes say the new model reads phone numbers, email addresses and IBANs with no pre-processing. Emails hold, six of six draws across two independent transcribers. US phone numbers hold, fourteen of fifteen. IBANs we are not reporting: long digit strings are where a speech recogniser inserts and drops digits of its own, so a wrong transcript cannot be pinned on the speech engine rather than the listener.
Ship a version, not a repoint
A version string next to default costs a vendor nothing, and it is the difference between a benchmark that expires loudly and one that expires silently. Every board that reads a bare alias, ours included, is carrying an assumption nobody is checking.
Gradium told us. That is why this post has numbers in it.