Mati Staniszewski
speaker
2,430 appearances
3 recordings
1 series
first heard Sep 2026
last heard 6 Sep
Mati Staniszewski’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 3 in all, peaking in Sep 2026 with 3.
Appearances
Then when you move from English to Mandarin or English to Hindi, then similarly you need good examples of how that works.
There's not that many examples that exist.
So you need to, and what we do internally, we have thousand people, contractors, that annotate and create data for us.
Then we refine this every time.
So we spoke about, you know, how we started and that voice was Canadian and then we optimized this, the translation wasn't perfect.
Uh
We learn on on on that to improve and improve and make it better and and effectively reinforce the model to to get smarter over time.
But for that we need data and we need high quality data.
So that's where the source piece comes in.
If you have a bad source, then the target doesn't matter.
You cannot like produce, you know, it's you need a good thing to start with too.
I know, I know, that's true.
And that's but that's where the kind of and that's where that second layer comes in, which is you need uh human in the loop for high quality content where translation is so essential, where the conversation carries the meaning.
That's what we
Over time, of course, the gap will start closing.
So you can teach based on, you know, today human the loop is like full.
We cross check how AI did it, make corrections, and then um validate that it's great, and then bring it to production.
So we actually have people internally that do that services work to make content great.
Um, but of course, the model will increasingly learn on what we did in the in the in the past.
But what I
Showing 441–460 of 2,430 · page 23 of 122
← Previous
Next →