2 ms·
Never is a long time. What if some day the technology exists to interpolate dropped packets (e.g., using ML) fast and cheaply enough to fill in the distortion?
by Xophmeister 10y ago
Never is a long time. What if some day the technology exists to interpolate dropped packets (e.g., using ML) fast and cheaply enough to fill in the distortion? That doesn't seem too outlandish in the next 20, or even 10 years.
- scandox 10y agoThat sounds fine, except for the timeline. What I'd like is an affordable, ubiquitous solution next week :)
- AstralStorm 10y agoUh, the modern codecs are already using linear prediction and correction. (essentially a rephrased more advanced perceptron) This is why SILK (Skype), Speex and Opus (SILK-derived part) can work even mildly well on bad links. They predict syllables and also long term speech characteristics. Going to full sentences or words is much harder and not likely to bring gain other than free subtitles. And would require really good voice recognition comparable to Siri at least. Good luck running that on an embedded device reliably. Remember, the codec has to handle less than 20 ms latency. Running a sizable ANN (fast) or HMM in that time is not easy or even possible. A small one will be no good for various speakers. Instead the codecs use various specialised error hiding tricks and adaptive strategies. Heck, Codec2, one of the strongest voice codecs around could be used if all else fails. It is beginning to be used in SDR and amateur radio - optimised for availability and intelligibility not quality. It would be nice to have a variant integrated with CELT for higher quality.