The thing you could do is train a localizer on the separated audio. Phase is estimated by the source separation process, so you can actually train an ML model provided you have some ground truth (e.g. estimated human locations from camera detections)
I haven't seen it provided as a commercial service or free model yet, but there is open source code for Mixit that lets you train using the open source / canned FUSS dataset.
https://github.com/google-research/sound-separation/blob/mas...
Re: #2. I’m assuming the varying depth is manually-controlled? Or is it automated by some method? If it’s manual, can the adjustment be made while transcription is active? In other words, can I change the focal distance to match the speaker without interrupting the speaker?
All in all, cool stuff! Best of luck with the work.
Indeed, as Dan said, one can change the depth on the fly. In fact, one goal of the development team is to make as much functionality as possible changeable on the fly. For example, you can currently change subtitle depth, pinning, and size on the fly, spoken language, subtitle language, microphones, and audio settings on the fly, etc.
I’d love for every setting and feature to support on the fly changes. That said, some things are currently fixed for a session, such as recording audio, and some third party software we utilize is less dynamic and forgiving of changes on the fly. For better or worse, in our software world of today, the sage advice of The IT Crowd “Have you tried turning on and off again?” still seems to hold with pragmatic force. And it still holds with XRAI ... sometimes ;-)
If you had more microphones placed in multiple spots on the glasses, such as up to 5 microphones - 1 in the center, 2 on the frames/end pieces, and perhaps a final 2 on the arms/temples. Then that would be able to catch conversation coming at a person from behind them, the sides, or directly in front, etc.
This would be such a cool use case for the latest ChatGPT tech when it gets faster in the near future.
Could you have an AI model that extracts some characteristics of the speaker's voice for each individual word, then translates that to color and font?
If the model was not confident about a word it could show slightly blurred, if it was loud it could be bold, perhaps(Although there's some stereotype issues) you could use different fonts for different pitches, whispers could be grey, quiet could be transparent.
Maybe there's a language model that can pick up overlapping words if you don't have the constraint of needing to sort them out into who said it, just show all the possibile words that could have been said by anyone stacked together, in a "not sure" color, and maybe the wearer would eventually learn to figure it out without much effort?
You could also try to stay consistent so the same speaker gets the same colors I'd possible, and also not reuse colors for new speakers that have been recently used, to best make use of the limited bits of data in font and color.
Maybe just by showing all the words from every speaker all together like that, the wearer would be able to figure it out even if it made mistakes in the speaker identification?
As for the rest, thank you for wonderful ideas! Everything you propose is technically possible. The difficulties arise first in assessing the increased benefit to users versus the increased complexity of user experience, and second in prioritizing the work versus other features. Over time we do hope to add additional selectable “skins”, which is to say different UI designs, that allow users to choose UI’s from a wide range. Everything from the simple to complex layouts, from accessible to exotic color palettes, from professional to playful themes, etc. I could definitely see more advanced visual representations of transcription uncertainty showing up in such optional skins.
1. For multiple simultaneous speakers of comparable volume, it’s only as good as the underlying speech-to-text engines we’ve implemented/integrated, which is currently not very good. It’s active area of research and engineering for us and we believe we’ll make strides to improve things; but, as you rightly point out, solving the crosstalk problem is very difficult. For the more general so-called "cocktail party problem", we can do a good job of filtering out more distant/lower volume voices and other environmental noise. Choosing the right microphone can improve things further, for example by pairing a noise canceling Bluetooth lapel mic.
2. We allow one to project the subtitles at varying depth, within the capabilities of glasses. We're seeing an effective focal depth range for fixed apparent size of about 0.5m to 3m. If one also allows change in apparent size, to simulate perspective scaling, the range is higher.