Finnish transcription

[rotund jungle]

Hey, can you tell when the Finnish transcription was good?

We can check if it's because of a model

[agile sun]

January 10 the transcriptions were still pretty much perfect

I can give you a conversation ID as well

[rotund jungle]

would be useful

[agile sun]

I think 21 November was even better actually.

But I'll give you both test IDs

[rotund jungle]

Yeah there was no change since 10th Jan

21st November many changes let me see

We are also shipping a major update now, wondering if you will see a difference

[agile sun]

When will the major update be out? Would love to test it immediately

[rotund jungle]

it's rolling out, give it 1 hour or so

[agile sun]

It could also be that we've just been lucky when testing previously and the model has been responsive, but just seen more bad results recently

Anyway, here's an ID from November that seemed to work 95%: jbG1fUYkMIolTYnjtNJc

[rotund jungle]

OK, an example which were bad?

[agile sun]

here you go: txQ2fZ3mkVQycXCnMqLc

[rotund jungle]

at what time the transcription is incorrect?

I see you switched to ulaw format for telephony?

[agile sun]

Yeah, most of these tests are phone calls

[rotund jungle]

We shipped an update that should impact quality of ulaw conversations few hours ago too, so try to observe a bit more from now on!

[agile sun]

The very first thing that the user says is incorrect, then at 1:06, 2:02, 2:17 and then I couldn't hear anything and asked the model to repeat twice and then the final message I could hear again.