How we build and measure the voice behind Nehorin: the method, the plan and the targets. It is in development.
There are no results yet. This page states the method, the plan and the targets we are aiming for. When we have measurements, we will post them here with dates. We do not publish audio yet. A demo is available on request.
A Hebrew-first voice for phone agents. It is not a wrapper around someone else's voice service: we fully fine-tune VoxCPM2, an open-source speech model released under Apache-2.0, on Hebrew speech.
Our training data includes public Hebrew speech from ivrit.ai, used under the ivrit.ai licence. We start at about 800 hours and expand toward about 20,000 hours. We also mix in English speech from LibriTTS-R (CC BY 4.0), so the model keeps its English.
We add data only while the score improves. No results yet.
We do not choose a model by its training loss. We choose it by measuring the speech it makes.
| Measure | How it is taken, and where it stands |
|---|---|
| Hebrew character error rate (CER) | Two independent speech recognizers, from different model families, transcribe the generated speech. We compare each transcript with the text the model was asked to say. Planned. No results yet. |
| Stop failures | Cases where the voice fails to stop, cuts off the first or last word, or repeats or skips words. Planned. No results yet. |
| Listening | People listen to samples and judge them. Planned. No results yet. |
| First sound | How soon the agent starts to speak after the caller stops. Target: within 1.0 second. This is a target, not a result. |
| Training data | Plan: about 800 hours, growing toward about 20,000 hours, only while the score keeps improving. |
Targets and plans are goals, not results.
Normal Hebrew text leaves out the vowel marks (niqqud). A reader uses context to pick the right word. A speech model has to do the same before it speaks.
The spelling ספר can be read sefer (book), sapar (barber) or safar (he counted).
This is an illustration of the problem, not a measurement.
In text, a date, a time and a count are written as digits. The voice has to say them as words, and in Hebrew some of those words change with gender.
In these two examples the digit 3 is the same, but Hebrew says it differently: שלושה before the masculine ילדים, שלוש before the feminine בנות. A digit does not show which form to use, so the voice has to work it out from the noun. These examples show the task, not results from our model.
Studio speech is recorded with a wide range of frequencies. A phone call is sampled at 8 kHz, so the sound it carries reaches only up to about 3.4 kHz. Our voice has to stay clear and natural inside that narrow slice.
We have not yet measured our voice on a real phone line. How well a synthetic voice stays clear there is an open question for us.