Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

They are going to collect so much voice training data from this.


telephony* voice data

[*] limited bandwidth (8 kHz), providing a valuable opportunity to enhance and specialize models for telephony applications, ensuring better performance and user experience even with low-fidelity audio inputs.


I mean, nothing prevents them from running their existing data through a "noisy POTS" filter in A/B tests to see how that impacts customer satisfaction.

But being able to blame the user's phone line probably goes a long way to avoiding unhappiness due to testing :)


While true... real world wins?


How does that work, without a transcript of what the voices are supposed to be saying?


Humans can label the data




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: