My CPU is a neural net processor
Actually a 6502.
Nvidia has designed a couple of new Tesla processors for AI applications – the P4 and the P40 – and is talking up their 8-bit math performance. The 16nm FinFET GPUs use Nv's Pascal architecture and follow on from the P100 launched in June. The P4 fits on a half-height, half-length PCIe card for scale-out servers, while the …
This post has been deleted by its author
I gather that 8 bits sound quality isn't adequate for computer speech recognition (sub-A.I. of course) That's very roughly AM radio quality compared to FM radio or Compact Disc, and the lesser quality is not good enough.
Maybe 8 bits A.I. is good enough to be an AM radio presenter, and understanding what other people say isn't a requirement?
Actually AM radio is rather better than linear 8bit, and the "goodness" of it depends on the bandwdith and interference. It was MUCH better in the 1930s! Even ordinary AM can be better than ISDN phone quality (8bit u or A law).
Also u-Law or A-Law 8 bit audio is perfectly good for speech recognition, that's the method used to improve on linear 8 bit for the majority of the worlds digital fixed line phones. GSM codecs are lower quality. 3G can use similar or better than GSM, but it's rarely better than 8bit ISDN. Similarly, VOIP connections can rarely support Fax or Analogue Modems, though 8 bit ISDN can.
I suspect that if the sound level is ideal, even 8 bit linear audio will work for speech recognition.
Humans can recognise 5 bit encoded speech, done using 1:32 variable PWM on a PC on/off loudspeaker.
Depends if it's carrier grade QOS managed VOIP or VOIP over random public Internet.
My own phone service (real geo numbers) is using VOIP on a fixed wireless link, but only as far as my ISP's voice switch. So Fax and analog modems usually work. Your mileage with a 3rd party VOIP supplier over the ordinary public internet will vary with your connection jitter and packet loss. Unlike gaming, or person to person chat, latency isn't an issue.
I assume each neuron need not only be working on it's own.
Just as a 32bit cpu can do 64bit maths, it just takes twice the time or number of cpus.
PS, and if Deepdream from Google is anything to go by, you don't even do your first pass on such fine grained data at times, it may be a single bit (active area or not) etc.
You're confusing quantisation levels with sample rate. I can play sounds out of the single pin of an Atmel ATTiny85 with little more than an RC filter and amplifier on the output.
Stick a CRO probe on the microcontroller pin, and you'd see a 1-bit signal with a 250kHz sample rate, but I'll bet you'll struggle to tell the difference between it and a cassette tape..
A serial bit stream could be any desired number of bits. So called 1 bit DACs are really 16 bit at least and the 48kHz or 44.1KHz parallel sample rate is unchanged, but the clock frequency becomes much higher.
1 bit on/off with 1:32 variation of mark/space ratio isn't one bit. It's five bits quantisation.
Tape can have a poorer dynamic range than 8 bit quantisation on cheap machines using a magnet to erase rather than AC erasing. Speech recognition was working with cassette tape replay in 1990s.
Biting the hand that feeds IT © 1998–2022