Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[dragdeler]

-

[methylatedspirit]

A peek through the manual suggests that you can directly connect that camcorder straight into a TV via an AV/SCART cable. I suppose if the capture card supports it? But wait, no, that's only for playback? Hell kinda camera (of its era) can't just stream what it sees to a VCR/capture card in real time?

[dragdeler]

-

[methylatedspirit]

I know that the quality of audio recordings on magnetic tape is proportional to the tape speed.

My question is, why? I'm somewhat aware that tape has this thin layer of magnetic particles, so I'm guessing it's some kind of stochastic sampling, where increasing the tape speed increases the "sampling probability", the likeliness that any given point in a continuous signal gets sampled by the magnetic particles. I don't want to guess, though.

And then there's dynamic range. Why do some tape formulations have better dynamic range than others? Without noise reduction, Type I Compact Cassettes have ~50 dB of dynamic range, while Type IV can do ~65 dB, so there's something different between them. (This pales in comparison to even CD audio, which does 96 dB (84 dB at worst if you have an early/crappy CD player), and pulls it off predictably and reliably. That's just the technology's potential. See "loudness war" for why this usually wasn't achieved in practice.)

[Iduno]

I would assume it's inversely proportional, like with VHS tapes. You'd either get higher quality recordings that lasted like 2 hours, medium quality at 4 hours, or record 8 hours at nearly-usable quality.

[Rose]

So if you're familiar with how didital audio works, it's essentially the sample rate. Magnetic tapes have little grains in them, each grain has a single magnetic charge over the entire thing. This is basically a single bit. The faster the tape is going, the more of these bits are going past the reading head per second, which directly increases the sample rate of the tape. Double the tape speed is double the frequency response.

Doing some quick math, based on numbers I can find, your standard cassette tape has about 8,000 effective audio samples per inch on the tape. You can multiply that by the tape speed to get the maximum reproducable frequency.

[Iduno]

The answer was bitrate. I typed that as a guess, then got on the VCR thing because that made more sense. S-M-R-T.

[methylatedspirit]

The answer was bitrate. I typed that as a guess, then got on the VCR thing because that made more sense. S-M-R-T.

I'm miffed by the imprecision of that summary. Allow me to get pedantic.

In the analog realm, yes, bitrate and sampling rate are related. The maximum data rate of a POTS line (assuming no data compression) is 64 kbps. That is as high as you can ever go, assuming perfect conditions. That's because these POTS lines can only transmit 8-bit (non-linear) PCM at 8000 Hz. 8 bits * 8000 Hz = 64000 bits a second.

Indeed, a VHS tape recording at a third of its normal tape speed (known as EP/SLP) only has a bandwidth of ~1 MHz total (compared to ~3 MHz normally, ~6 MHz for broadcast NTSC signals) to encode the entire video signal. It's sampling at a lower rate overall, so it encodes less information on its magnetic tape. That's why the video looks blurrier than one running at normal speed.

However, once you go digital, they're no longer as strongly related. I'll give this caveat: in uncompressed formats (raw anything, especially audio), they are related quite strongly. A raw stereo 192 KHz, 24-bit PCM file has 12x the bitrate of a raw mono 48 KHz, 16-bit PCM file. Bitrate = (no. of channels) * (sampling rate) * (bit depth). Indeed, if you just mess with the sampling rate, you're going to see that the bitrate is closely-tied to sampling rate.

(As a footnote, I'll point you to 24/192 Music Downloads ...and why they make no sense for why going above 48 KHz is totally unnecessary for an end-user. Or, if you want a video demonstation, I'll point you to Digital Show and Tell.)

For lossy formats, however, bitrate and sampling rate are no longer intrinsically tied to each other. Now, the lines do get somewhat blurred. LAME, the de facto standard MP3 encoder, does not let you go all the way down to 8 kbps unless if you also decrease the sampling rate down to at most 24000 Hz, which is really only good enough for speech. I can demonstrate right now with a copy of Audacity and a copy of Tom's Diner by Suzanne Vega. Import the song into Audacity, and attempt to save it as MP3 at a constant bitrate of 8 kbps. This happens:



(I got pranked by FFmpeg by trying to produce an 8 kbps MP3. -r and -ar are entirely different arguments. I'm too lazy to demonstrate with it again.)

This also occurs in more advanced formats like Opus. Opus is built on the music-oriented CELT algorithm and the speech-oriented SILK algorithm. CELT is good at music, stretching across the entire human hearing frequency range (maxing out at 48 KHz sampling rate/24 KHz max. frequency). However, it needs more bandwidth to work. That's where SILK comes in. It's good at encoding speech at low bitrates, but its frequency range is limited, only going up to 16 KHz sampling rate (8 KHz max. frequency).

The magic of Opus is that at lower bitrates, it's able to switch between (or even run both) the algorithms as needed. I'll use Money Game Part 2 by Ren. Here's a spectrogram, showing the original (Opus @ 160 kbps), Opus @ 24 kbps, Opus @ 16 kbps, Opus @ 12 kbps, and Opus @ 8 kbps, in that order. They're evenly-spaced in 45-second intervals.



Here, bitrate is related to sampling rate because at these low bitrates, the Opus encoder needs to conserve what little bitrate it has by using the SILK algorithm (itself frequency-limited) in tandem with the CELT algorithm. At 12 kbps and lower, however, it stops using CELT, using SILK exclusively, and you can see the maximum frequency drop. At that point, Opus is no longer able to run the CELT algorithm due to bitrate constraints. It loses the high frequency parts entirely because of this. As the bitrate drops further, the sampling rate drops all the way down to 8000 Hz, losing even more high-frequency parts.

That being said, you absolutely can do stupid shit like encoding 256 kbps Opus with a sampling rate of 8000 Hz. Try the following in a terminal:

Code: [Select]

ffmpeg -i foo.flac -c:a libopus -b:a 256k -ar 8000 foo-opus-256k-8000.webm

If you then run ffprobe on it, you will see:

Code: [Select]

Input #0, matroska,webm, from 'foo-opus-256k-8000.webm':
  Metadata:
    ENCODER         : Lavf59.0.101
  Duration: 00:03:42.58, start: 0.000000, bitrate: 270 kb/s
  Stream #0:0(eng): Audio: opus, 48000 Hz, stereo, fltp (default)

I mean, Opus forces a nominal sampling rate of 48000 Hz regardless of bitrate, but if you listen to it, it will sound muffled because the audio's been downsampled to 8000 Hz. Here's a spectrogram of Money Game Part 2, original and downsampled to 8000 Hz:



Bitrate and sampling rate are not inherently tied to each other in the digital realm. While most codecs are "smart" and won't let you encode, say, 48 KHz audio at 8 kbps, the reverse is not true. You can easily encode 8000 Hz audio at 256 kbps without much issue. They cannot stop you.

This post has been brought to you by the Department of Why The Hell Do I Know This.

[methylatedspirit]

Something that's been on my mind for a while now is that my perception of music tempo is tied to how fast my brain is running. The relationship can be summed up as follows: "Brain slow, music fast; brain fast, music slow". Listening to music just after waking up yields a faster-than-normal perceived tempo, and consuming stimulants like caffeine yields a slower-than-normal perceived tempo. I once experimented a bit too much with coffee, and I could feel music running at, like, 85% speed.

Why, though? I think it affects all time perception, but it's especially noticeable with music, since beats are consistent on each playback of every song (I'd certainly hope so, considering my equipment is all-digital). I have a consistent reference for what constitutes "normal" speed, so deviations from that are quite noticeable.

[methylatedspirit]

Linear predictive coding (LPC) (as paired with the source-filter model, rather) is used in speech synthesis as a way to model the human voice as essentially a buzzer and hiss-maker at the end of a filtering tube. It's basically the definitive way to encode speech efficiently, being used in a lot of speech codecs for precisely this reason. There's a very distinctive sound to LPC, one that we'd describe as "robotic". To me, LPC evokes "robot speech". Here's a page with audio of a TI Speech+ as an example of an early LPC speech synth. Tell me you're not imagining a robot speaking that.

Had people come to a different model for speech synthesis very early on and that got widespread use instead, would we have considered that our conception of "robot speech"?

[wierd]

Probably.

Regardless, stilted speech is almost certain to be a feature.


I remember my grandmother asked me one day 'What is wrong with that man?', referring to the computerized male voice reading the weather report over her weather radio.
I had to explain to her that he was a text to speech engine, that produces human-like speech from written text. It was not a bleeding edge voice synthesis package that can do vocal lilting and natural tone inflection-- but it was at least two steps up from Microsoft Sam.

I remember her being completely aghast that the weather reporter was a robot.

[King Zultan]

First robots steal the weather reporting jobs, nest they'll be stealing all the jobs!

[methylatedspirit]

To be fair, the worst wierd's grandma would've experienced would be something like GSM Full Rate, which doesn't sound robotic at all (but is LPC-based). I can push full music through it, and it certainly doesn't sound quite as bad as whatever Microsoft Sam +2 would be.

"The quality of the coded speech is quite poor by modern standards", my ass! I can't tell the difference between GSM Full Rate (13.2 kbps) and Speex @ 13 kbps, at least when misused to encode Tom's Diner at 8000 Hz, mono. Opus @ 13 kbps seems to do better than both of those, because of course it does.

Not that it matters; we're at sampling rates low enough that I didn't think people actually cared about quality enough to have an opinion on it. And then of course people apparently give a shit, because the Mean Opinion Score of GSM telephony is apparently between 2.9 to 4.1 out of 5. I'm not even sure who you'd have to ask to get that info. What, do they just interrupt you one day and ask "On a scale of 1 to 5, how satisfied were you with the quality of your last call?"?

[dragdeler]

-

[Maximum Spin]

I can't tell the difference between GSM Full Rate (13.2 kbps) and Speex @ 13 kbps, at least when misused to encode Tom's Diner at 8000 Hz, mono.

The important thing that people talking about audio (and video for that matter) quality never seem to fully appreciate is that some people can distinguish more than others.