Two Shores

GoIPpro works from two places. Platform development happens in Ukraine. Portugal is the second line: this is where we train our own speech model.

Why our own model, when cloud models can already hold a conversation? The cloud is a stage. It is a convenient place to run experiments, test ideas and learn from live conversations. But a platform that promises customers a private perimeter has to speak, in the end, with its own voice — on its own model, on its own hardware. That is where we are heading.

We started with Portuguese. European, not Brazilian — the Portuguese you hear in Porto, not in São Paulo. The work has been going on for more than a month, and we will say plainly where it stalls: compute. We have few GPUs, so training is slow — every epoch takes longer than we would like.

To speed things up, we have applied to several large organisations that help young technology companies with compute — at reduced rates or free of charge. No answers yet. When they come, we will tell you here, whatever they turn out to be.

One more thing. This month has left us with a lot worth sharing. Where the model went wrong, and why. What looks simple in textbooks and breaks on live speech. What we tried and what did not work. We have not been doing this for long — and that is exactly why we notice what seasoned people have stopped seeing.

So starting with this dispatch, the journal gets one more line — a training diary. No gloss: the model's mistakes, our mistakes, numbers when we have them. Like everything else behind the glass.