| So I've been recording voice notes for a number of months now, and there's a feature in the client I use called voice notes, which allows you to download the audio—in other words, the original audio recording that you submitted that then got transcribed. I don't really like this feature because I never want, I never use those audio recordings. I don't like the idea that I have them and the rest of the team. So I have a bunch of these audio recordings out there in the cloud, but I am interested to know if there's any use I can make of them. | |
| So I have something I've really thought about doing for a while, which I'd love to do, is creating a fine-tuned speech-to-text model optimized for my own voice recognition performance and the words that I use. The issue is this: the transcripts from voice notes are going to be imperfect. And I know that for training speech-to-text models, you typically have a source of truth file and a transcript. And that's how they're trained. I don't have the source of truth because I haven't manually written any transcripts. Without that, is there anything that I can still do? | |
| Let's say that I have several hundred hours aggregated of these recordings. It's quite a substantial bank of data of me speaking in different audio environments, with different levels of background noise, different times of the day, that I feel like could be useful. So I can download the MP3s, put them into a repository pretty easily. But I'd be curious if you can think of any actual uses I can make in the speech-to-text thing that I've mentioned. |