PWM versus codec

I looked at the same problem. I wanted to use an AVR micro to build a "phone line to RS232" converter. The problem is that none of the cheap micros have a good DAC on board.

You want the PWM frequency to be ten or so times greater than the sample rate, or about 80 KHz for telephony applications. For my design, this ruled out the 16 bit PWM since it could not count up fast enough to give even 14 bits at 80 KHz.

The best I could come up with was a hybrid approach. I used both PWMs on the AVR, with both working at 6 bits of resolution. The idea was to combine the two outputs with an "R-64R" resistor network to do a weighted sum of the PWM outputs. Luckily I found another way to solve the problem and never had to get this crazy idea to work. I look forward to seeing the real solutions posted in this thread.

My ultimate goal was just an answering machine that would work under Linux. I found a winmodem approach which, if you are interested, is documented at:

Bob

audio

quality"

Ultimately what you are suggesting is certainly possible, and the basic concept has been successfully implemented many times before. How high the maximum achievable quality you can attain will depend upon your microcontroller's PWM capabilities, as well as your cleverness as a programmer and engineer.

I've got a couple of brothers who are fans of Microchip microcontrollers who have done just this before. Both have produced surprisingly decent quality music (even in stereo) from the built in 10 bit PWM module on the higher speed PIC (40MHz) devices. Of course, I'm no audiophile, but I found the music quite acceptable to listen to.

As the sample output rate increases , there are limitations on how many bits of PWM resolution that are available, I'm not sure if they were using the full 10 bits. I suspect they were using something more like 8, but I don't know that for sure.

See also:

Even at 8 bits, the quality attainable can be significantly better than that of "telephone quality." I imagine most people's ears aren't really sensitive to 16-bits of audio resolution (a mere 15 microvolts for a 1Vp-p signal range), the real benefit of using 16-bits is you get to implement various tricks like digital audio amplification and scaling without running into serious limitations. Assuming your audio source is properly scaled and normalized to begin with, 8-bits is enough to deliver quite high quality audio.

Obviously you want to run your PWM frequency as high as possible, but there are still practical limitations in terms of the maximum audio sample rate achievable, due to the low pass filter on the output (which filters the square wave output to an analog voltage). An "R-2R ladder" (hint, google search terms) poor man's DAC approach could likely significantly exceed the performance of the PWM module in this case in terms of maximum audio frequency renderable, but R-2R ladders take a bunch of I/O lines. At telephone quality, you shouldn't need to bother, although that does depend of course on what microcontroller you are using and it's capabilities (as well as your capabilities).

codec.

(digitally

the

audible

It appears you may have the wrong idea about how the PWM module would typically be used. You don't need a 500MHz microcontroller or any kind of fancy digital or analog filters.

Depending upon what your audio format is in to begin with, all you likely actually need is a single RC low pass analog filter. You hook the resistor up to the output of the PWM module. On the other side of the resistor you hook up a capacitor to ground. You get your analog output signal off the top of that capacitor.

The PWM module is typically quite easy to use, you simply configure a couple of registers and it starts oscillating at some fixed frequency at some set duty cycle. You get to set the frequency and duty cycle by playing with the registers, but normally you just set it and it goes by itself. Normally, for DAC use, what you would do is simply configure it to output the maximum practical frequency available (which ideally should be at least 10X the

*sample rate* (not sample bits) of the audio you are trying to output, the higher the frequency the better). So for instance, if you wanted to output 8-kilosamples of audio per second, you would ideally configure your PWM module to start oscillating at least at 80kHz. This should be no sweat, just set up the proper registers and it will be good to go.

Once you have the PWM module configured, you are ready to start outputting the audio samples. Suppose you have a 10-bit PWM module available (meaning that you can adjust the duty cycle of the output frequency anywhere from

100% to 0% in 1024 discrete steps). This would allow you to output at maximum 10-bit audio samples. For each audio sample, you adjust the PWM duty cycle (but leave the frequency the same) corresponding to what the audio sample actually is. When you low pass filter the digital square wave coming out of the PWM output, you get an analog DC voltage. For example if you had say a 5V CMOS microcontroller, presumably the output of the PWM module would go from 0V to 5V. Suppose you output a "512" audio sample, which would correspond to 50% duty cycle on the PWM output. This would cause the low pass filter to produce an average DC voltage of 2.5V (half on at 5V, half off at 0V, average 2.5V when low pass filtered). If you outputted a "256" audio sample, or a 25% duty cycle output, the DC voltage on the low pass filter capacitor would drop to 1.25V. Etc. This is how you get your analog output voltage. Typically you would then AC couple this analog DC voltage through a capacitor to an op-amp buffer (although you can simply AC couple it to earphones and listen to it albeit with very weak volume).

Since the output of the PWM module is a digital square wave instead of a true analog voltage, you want to very heavily low pass filter it, lest the square lumps of the PWM main frequency will appear in your analog output. This means you want to place the low pass filter breakpoint at very low frequency, as low as possible without unreasonably attenuating the higher frequencies of your real audio signal. If you were outputting 8-kilosamples of audio per second, then it would seem reasonable to place the low pass filter breakpoint somewhere very vaguely around 8 kilohertz. Of course, if your main PWM frequency is a mere 80kHz, that is only one decade higher, so you will only get 20dBs of attenuation of the 80kHz content... That is not that great. The higher your PWM frequency, the better the attenuation will be, so set it for the maximum practical that your microcontroller will produce (subject to the constraint that the duty cycle resolution will probably decrease at high PWM frequencies, so there may be a compromise here as to what maximum frequency is optimal).

The speed of your microcontroller really only needs to be as fast as you can handle the audio samples to make them appear in the PWM duty cycle control register(s). Only a modest microcontroller would be adequate for modest quality audio, although that depends on how clever and efficient a programmer you are.

Hmm... Actually since 8 kilosamples per second only produces a maximum audio frequency of 4kHz, maybe putting the breakpoint closer to 4kHz would make more sense. I guess you get to play with it, see what sounds best and looks the nicest on the o-scope.

What I dont get is this: For a couple bucks I get a standard 13-bit codec.

13 bits is *sort of* standard resolution for telephony when using a linear codec. So to equal this with PWM i need to run at 2^13 * 8000 = 65.5MHz. Then I hear that as a rule of thumb you want to be 10x faster! This would mean i require a clock of 1/2 a gigahertz! Is this 10x really necessary?

And say you do bump the frequency up 10x, do you need to upsample (digitally filter) the audio signal you are putting out? Or just keep putting out the same value for 10 samples...which seems pointless?

I am worried about how hardcore the filtering needs to be after the PWM output. I dont have weeks to design the greatest filter ever....Are audible artifacts of all this switching easily removed?

I hadn't thought of doing volume control, but that is a clever idea. I do need two distince volume levels and was toying with the idea of modifying the level of the square wave before the low pass filter.

Thanks for your comments.