USF Study will be virtual - ideas welcome

Part 1 (2020)
25

He is an alternative but bargin basement.

Noted from this site;

Not sure if available outside UK. But may help somebody
Just realised the link from lifework to amazon does not produce the correct item. I brought the wired version shown above. HeyHo

26

Well if we have to be isolated at least we have somewhere to channel our energies.

1 Like
27

Late to the game, but my 2 cents…

Use zoom for video meetings. It’s the best in my experience.

And this idea is a bit more fringe…

Since this is entirely online, have you considered knocking off 50% and offering the course certificate to everyone as a piad option? I think a lot of the folks paying for the in person class are paying for that and for the ability to work with you and other classmates in person. Perhaps they get that experience virtually as much as possible whereas new folks can get the certificate and work in other virtual groups. I think this may be a way to make the folks signed up for the on person class not feel jaded and also recoup on $ that may be lost should in person folks decided to opt out. We could even have group competitions against imagewoof and it’s partners and/or group based virtual poster presentations. Just an idea.

28

The absolute most important thing is that the mic is close to your mouth. That means a long boom. Even an omni mic on a boom can be tolerable. The best mic in the world, if stuck up next to your ear, wont perform well and will be reverberant and will even echo back to the rest of the group. Of course, a noise cancelling mic on a boom is even better.

A “mic on a wire” solution like iPhone stock headset is mediocre at best because it’s not close enough to your mouth unless you tape it there (in which case it’ll work okay, but you’ll look like you’ve gone crazy).

(BTW did I mention… my day job is making microphone arrays and writing echo cancellation software :slight_smile: )

8 Likes
29

Ah - I was wondering why you had such excellent advice. As someone who has done a lot of remote work for decades, I’ve experienced the “long boom” issue many times, but few people discuss it!

30

I don’t have the ability to change the price (that’s USF’s decision), but I have opened up applications to folks outside SF to pay to join the virtual course.

5 Likes
31

I’m pretty sure you’ve researched on this topic a lot already. If I were to add to this, here’s some links from my bookmarks that I plan to use when setting up my remote/home work desk. Perhaps useful, and given that @ccrome is an expert here, maybe could comment on the content.

https://marco.org/podcasting-microphones

1 Like
32

Do you have a cutoff date for applications? I’d assume it’s like the 12th?

33

Nice Article. I totally agree with the article – everyone forgets about the microphone, and don’t pay attention to it. We run into this all the time, where if the speaker sounds good, that’s all anybody cares about – they assume if the speaker is good, people make the implicit assumption that microphone is good too. What nearly everyone fails to recognize is that making a good speaker is relatively easy. Making a good microphone, with good echo cancellation and echo management is very hard.

Comments on the article… (I know it’s slightly off topic, so if there’s more discussion around audio setups, maybe we should take it to another thread)

There is definitely a bit of misleading/incorrect information in the article, but overall it’s pretty good.

Some comments on what they got right:

  • Microphones and their usage make a huge difference in audio quality
  • You don’t need to spend a lot of get very good performance, perhaps even excellent performance
  • placement and usage of the mic is paramount (i.e. close to you, out of the air stream)

And a couple inaccuracies:

  • “Large microphone sensors (capsules) similarly produce noticeably better audio quality”. Well, yes and no. Large capsules produce lower self noise, but at the expense of a high quality polar pattern at higher frequencies. The self noise of microphones is a real problem when recording from either far away, or in extremely quiet environments. Physically large sensors essentially interfere with themselves when wavelength gets to be roughly the size of the sensor. So, for example, a large diaphragm microphone might have a diameter of, say, 25mm, or 1 inch. It will have extremely low self noise because the electrical signal produced will be very large compared to, say, a 5mm diaphragm (about 15dB better all else being equal), but a 25mm diaphragm will start not behaving very well when the dimension is about 1/2 wavelength. So, if 1/2 wavelength = 25mm, 1 wavelength=50mm, which is a frequency of 343m/s / 0.050m = 6800 Hz. Not the end of the world if recording voice. But doesn’t sound like a great idea for recording music. A 5mm cardioid mic, will exhibit the same amount of distortion at 34kHz, well outside audible range.
  • “Cardioid mics have a very narrow pickup pattern” Well, not really, it’s pretty broad – roughly speaking, a cardioid picks up most stuff from the front, and rejects most from the back. Even at 90 degrees off-axis it’s still -6 dB from the forward looking direction, which means it still picks up from 90 degrees off axis quite well. Humans perceive 10 dB to be ‘half as loud’ or ‘twice as loud’, so at 90 degrees off axis, a cardioid isn’t even cutting out half of the perceived loudness. So, noise at 90 degrees will still get picked up great – like if you hover the microphone above your keyboard, it’ll get picked up quite well.

Anyway, here’s a summary of what we in a conference-call situation are fighting, in roughly the order of importance:

  1. Echo. By far, the worst possible noise source in a conference call is echo. Echo is when the sound you’re hearing from the remote side is played our your loudspeakers, and captured by your microphone, then sent back out to the rest of the world. If people on a call are complaining of echo, and you don’t hear the echo… you are the source of the echo! Different platforms (Hangouts, Zoom, Skype, etc) have different echo management quality. The best way to reduce pain for everybody that’s trying to hear you, is to reduce echo picked up by your microphone. True echo cancellation is a difficult problem (and happens to be what I do for a living). There are a couple salient points to think about: If your loudspeaker and mic are on separate sound-cards, they very likely don’t have their clocks synchronized, which makes echo cancellation difficult to impossible (There are pages and pages of background to this problem that I’ll just skip over). So, systems like Zoom that run on a general purpose PC, pretty much have to assume the loudspeakers and mics are not synchronized, and also have to assume that the loudspeakers are not very linear. (nonlinearities can’t be cancelled), and so they resort to echo suppression, which is very different from echo cancellation. Echo suppression suppresses your speech as well as the echo. So, your voice will cut-in and cut-out whenever somebody on the far side speaks. With echo cancellation, it’s literally removing just the echo, leaving your speech. You can easily get echo on a poorly designed headset, and you will have a ton of echo in a standard set up with a microphone hovering over the table and open loudspeakers, even if you’re using a good directional microphone, but far away (say 0.5m).
  2. External Noise: If you’re in a noisy environment, you want the most directional microphone possible, and to get it as close as possible to your mouth. Generally speaking this is a so-called ‘noise-cancelling’ microphone, also known as a ‘bidir’ microphone, placed on the end of a boom close to your mouth. It has a bidirectional response, picking up from the back exactly as much as it picks up from the front. “How can that be,” you ask yourself? The reason these work really well is that they only pick up equally from front and back, assuming your interfering sources are at the same distance. They use what’s called the proximity effect, which naturally boosts anything close to the front of the microphone much more than things far away. So, even though in the far field, this type of microphone looks like it points backwards as much as forward, if used properly, it’s dramatically better from the front, rejecting everything far away. Anyway, the best way to reject unwanted noise is to be in a place that is quiet :-). If that’s not possible, use a noise cancelling mic on a boom.
  3. Reverberation: The next thing you’re fighting is room reverberation. This is the effect that we in this class will probably be fighting the most: It’s the effect that we’ve all heard when on conference calls. It’s that everybody sounds like they’re inside a tin can. If we all had good anechoic chambers to take the course in, then an omni microphone placed just about any distance would sound great because it would capture the direct sound from your mouth, and all the rest of the sound would get absorbed by the walls. This is not realistic of course – normal rooms reflect much of the sound back. If your microphone is at about 1 meter, then 50% of the sound the microphone picks up will be reflected sound and 50% will be direct sound. So, an omni mic at 1 meter sounds terrible. Which is why when you throw your iPhone down on the table and have a conference call with it, it’s not too great (iPhones have many mics and can do some fancy signal processing… but I’ll ignore that for now…). By moving the mic closer to your mouth, the ratio of direct to reverberent energy increases, and it sounds better. Or equivalently, you can use a directional microphone, which sort of ‘brings it closer’ by rejecting the reverberation. If you need to record at a bit of a distance, say 0.5m or farther, the Hypercardioid pattern is the best at rejecting all reverberation. The Cardioid pattern has the best front/back ratio: that is if you integrate all the energy in the front hemisphere and compare it to all the energy collected in the rear hemisphere of the microphone, the cardoid is the best at that. So, for example, talking at a podium, where you clearly want to reject from behind (the audience), a cardoid is a good solution. In fact, cardioids are the most common directional mic for good reason, they strike a good balance.
  4. Table Bounce. Table bounce (and similarly monitor bounce), occurs when you’re talking to a microphone, say, 0.25 or farther away, and the mic is hovering over a table. This can be catastrophic to sound quality. The reason is because the microphone gets the direct sound (i.e. draw a straight line from your mouth to the mic – that’s the direct sound), and reflected sound (draw a line, using the table as a ‘mirror’ for sound), and the mic picks up both at roughly equal levels. This creates a wonderful comb-filter that can notch out the most important part of the speech spectrum, around 1kHz. Search google for comb-filter or table-bounce. Here’s one article, but there are a zillion. BTW, DPA microphones definitely know what they are doing.

So, what’s the solution? all of these problems are helped by increasing your signal level, with respect to the noise (interferer) level. What’s the best way to do that? Simple, a headset with a noise-cancelling microphone on a boom :slight_smile: And it finally comes full-circle. Every halving of the distance to the microphone, increases your SNR by 6 dB. If you’re starting with a mic at 50 cm, then 25cm will be 6 dB better, 12.5 cm will be 12 dB better, 6.25 cm will be 18 dB better, and 3cm will be 24 dB better. Note that switching from an omni to a cardioid, at the same 50cm distance, only helps your SNR by about 5 dB. So you can see, the microphone’s directinoality is way less important than placement. You can get 24 dB by moving it from 50cm to 3cm, but only 5 dB by going to a fancier microphone type. This is quite independent of price you paid. A $0.10 omnidirectional microphone placed at 3 cm from your mouth will destroy a $10,000 microphone placed at 50cm in all these performance metrics, except for self-noise, and who cares about self noise in a conference call, especially if the mic is next to your mouth? It’s never an issue.

One other thing note: It’s the compounding of all impairments that make conference calls really bad. In person, you might be able to talk with somebody who isn’t a native speaker of your language without any problems, but when you get on a call when them, you have to ask them to repeat things all the time, or you just get fatigued by the crappy quality and give up all together. It’s the compounding of many different impairments that make the communication difficult.

  1. Bad microphone placement
  2. Possibly bad microphone frequency response
  3. Echo. Echo is literaly the worst kind of noise. It sends you into brain lock and prevents you from even being able to talk.
  4. bluetooth encoding. Bluetooth headsets have a low bitrate encoder, which really messes up sound quality. Some headset/computer combinations can use a higher quality encoding, but by default, the sound quality is quite messed up.
  5. network encoding: Your app, skype, zoom, whatever, gets the un-encoded signal (perhaps already degraded by the bluetooth encoding) and re-encodes it to send over the network. I’m not that familiar with what Zoom & Hangouts use. Maybe AAC-LD or something like that. But stacking encoders isn’t great
  6. bad networking/choppy networking. Sound drops out depending on network dropped/late packets.
  7. long latency. Even on a reliable network, there might be hundreds of milliseconds of delay, even up to seconds of delay. This makes natural conversation difficult and you often step on each others speech.

So, long story short: It’s the compounding of all these issues that make conference calls so horrid. By far-the best and most important thing you can do to make conversation on a conference call better is to put the microphone close to your mouth (but outside of the puff-stream). Everything else is secondary.

-Caleb

15 Likes
34

I’m not sure how this’ll go but I’ve started a Zoom call, I’ll be in it for the next 4-5 hours (Starting 10:30 AM PT, 9th Mar), join here

I’m working my way through Zach’s course and the fastai book. Please feel free to join the “Study call/group”

Edit: Closing the link, happy to try it again if anyone wants to join.

1 Like
35

From what I gather from this whole discussion, I guess a good solution will be such that:

  1. Anyone in a group can Livestream their screen to others.
  2. There can be comments. Of course, people can also speak through audio.
  3. Everything works as well as a Youtube Livestream and sessions are automatically recorded.
  4. The number of simultaneous connections is not severely limiting.

As far as I know, Google Hangouts limits the participants to 25 on a video chat. I’m not able to think of other better options right now.

EDIT: Zoom has a limit of 100 participants by default which increases to 500 with a “Large Meeting” add on.

Zoom is probably the best way to go about this I guess.

36

Okay, Duly Noted. Sorry for the incorrect assumptions :slight_smile:

IMO the Zoom way might be a great method to start with, and then we can use the forums to communicate about our project ideas to keep the central communication open to all.

Similar to the course lectures: One central Meta threads with links to other “individual threads/SGs chat”

37

And I of course forgot to mention that particular setup in the article you referenced, i.e. a mic on a boom, relatively close to your mouth. That’s a great setup for recording – it’s easy to use and will sound very good during single talk. However, if you’re using open-audio with loudspeakers as opposed to headphones, you need to keep the loudspeakers as quiet as possible to reduce as much echo as possible, because echo is your biggest enemy. -C

38

https://meet.jit.si/ is another option for web conferencing tool. it is free to use, has not limits on time or no of participants. it has screen sharing, livestream to youtube, recording to dropbox, chat. No login or account needed. recently they saw 500% increase in usage (especially in Italy) and scaling up the service to support that.
*works best with chrome / chromium

4 Likes
39

I think Zoom will work well and I’m sure that students around the world will coordinate their schedules to collaborate on projects given this amazing opportunity! I’m looking forward to this experiment- not just in this course, but throughout the entire higher ed system. This may turn out to be a Black Swan event that disrupts higher ed. At the least, it’s going to be very interesting!

In past courses, you’ve highlighted some of the student projects at the start of class. I’m really interested in seeing what develops from having the open link for a virtual study group this time.

The opportunity to participate in the study group virtually is a bright spot for me in the otherwise dismal experience of having been required by my employer to attend a relatively large in-person conference in Seattle last week. It was a very unnecessary risk and attendees came from multiple states + British Columbia and also felt they had to attend. The risks are real and my spouse and young child are at high risk, so we’re nervously waiting to see if we’re going to become sick from my required exposure.

THANK YOU for moving the course and virtual study group online, Jeremy!!

3 Likes
40

Maybe we could make mic noise cancelling an attractive study group project

2 Likes
42

I urge you not to do a face-to-face at this time. In addition, any study group of 10 or more would be prohibited by current federal guidelines. We don’t want to be part of the problem.

1 Like
43

All meetups have been cancelled… we sent out emails to regular attendees a while back. Thanks for your concern

1 Like
44

I would like to join this study group. I live in SF. While that doesn’t matter now, hopefully we can meet in person in the future.

1 Like
45

@golz feel free to join! Details here: https://forums.fast.ai/t/official-project-group-thread/65817

2 Likes