Turns out that it's actually 2-bit and not 1-bit. It uses a combination of the EAR and MIC jacks, and when you output through the MIC jack, it slightly boosts the EAR jack. That gives you 4 possible levels, but they are not evenly balanced.
I gotta try this on real hardware, reading the original article it seems it's barely audible without an amplifier, unlike on an emulator.
Will post an update after I get back to my parent's home, where I have these beauties that launched me into computer science: https://imgur.com/a/GKrA0vX
I created a program like that back in the day. In my case, I recorded my voice through the "MIC" input and played it back through the TV. A few seconds of sampled audio required the entire available memory. While I fully understood what was going on, I thought it was magical that the Spectrum could "speak with my voice".
On ZX-Spectrum? There was no "MIC" input on ZX-Spectrum, not that I know of. I was in high school at the time with only a rough understanding of the hardware but it was kinda obvious to me that whatever I plugged into the casette input port was only read as "0" or "1" by the IN <port> (forgot which) instruction. So not really possible to record voice at the time (didn't even know of PCM anyway), but I tried to decode Morse transmissions from radio. Never got anywhere with the software but my ear got good enough at some point that I could decode them myself by hearing :)
There was Z80 code to record from the MIC socket in one of the small books (Melbourne House maybe?) that came out in the early 1980s. I remember borrowing it from the library and it worked! Heavily quantized of course by the snapping to 2 or 0.
Also some games had sound samples, Activision’s Ghostbusters had a couple.
With proper volume control, even one bit could probably sound voice-like enough to a person in the 1980s.
It may be feasible to retrieve a pcm for a very short clip by playing the same recording at different levels which then could be converted on the same machine to a sequence of opcodes playing a more realistic voice.
Nope. I did the same thing myself with my Spectrum. I guess I was 14. I didn't have a microphone so I tried it with my dad's stereo headphones. I got the idea to try it after looking at the Z80 machine code of cassette saving and loading routines. Maybe it was also after ETX came out, first game I remember with speech samples. (Terrible game). I forget the details now but it's all about the timing between turning the mic/speaker on or off.
I recall, One bit sound was detailed extensively in Byte Magazine starting in mid 70s. A square wave that has its frequency varied.
A google of the words " Generating speech with a square wave wave , Byte magazine " picks up lots links about this technique.
Fourier transform show a square wave consists of odd sine wave harmonics.
[A more advanced technique (more than one bit) would use microprocessor interrups to generate several sqare waves, output them with more than one IO pin and combine them with a resister network]
This mention of one bit sound made me remember the Lis’ner 1000 project of Steve Ciarca , cira 1984 , WHICH WAS speech recognition system of about 30 words. It used a GI SP1000 chip
Perhaps. I think these words have somewhat flexible definitions, and that thought process leads me to some rhetorical questions.
Questions like: How many channels can a mod file[1] have? Does that number of channels this file has change depending on whether it is being played back with monophonic or stereophonic (or some other arrangement of) speakers? What happens to the count when one or more individual samples within that mod file contain polyphonic data (chords!), themselves?
It also leads to rhetorical statements, such as: When I play a film that has a soundtrack with 8 (ie 7.1) discrete audio channels on a stereophonic pair of speakers with a sane playback chain, I'm definitely still hearing all 8 of those channels. The same happens if I extend that stereophonic system to greater number of output channels, by adding surround speakers (quadraphonic!) or whatever.
Or maybe I take one earbud out and put it back in its charging case, and through the magic of active electronics, good software, and wireless comms, the output deliberately collapses to being monophonic.
However it happens, I've still got an 8 channel of film soundtrack in my ear(s).
And of course, it also works the other way: If the electronic musical keyboard in front of me is said to support 10-voice polyphony, then that's a hard limit on the number of notes it can produce at once regardless of the number of audio output channels.
Multichannel and polyphonic are pretty bendy terms, I think. Maybe that flexibility was wrong at some point, but they've been flexible in this way for so many decades that it definitely seems right to accept the overlapping uses as being correct.
> My previous work with PWM relied on being able to set a hardware timer to control when the speaker turns on and off.
I've got an Arduino project where I did that. The output is on a single digital pin, audio encoded as PWM. It turns the output off at the beginning of an audio frame, and sets a timer to trigger an interrupt when it's time to turn the output on.
I've got a memory block of event structs, each event representing a write to NES sound hardware. I emulate enough of the features to support everything that the Legend of Zelda title theme uses. 16MHz is more than enough to decide on the sample for 2 square channels, 1 triangle, and 1 noise, outputting a 6-bit output sample at a 31250Hz sample rate. The output is only slightly scratchy; the polyphony is actually surprisingly smooth.
I did a PWM demo on the Apple II using a format similar to the DPCM samples on the NES. Those samples have a pretty simple format: a 1 bit raises the amplitude, a 0 bit lowers it. So you're kinda approximating the waveform with a bunch of triangles.
The NES has a 7-bit DAC, but since we're bit-banging with the CPU we only have enough time between samples for about 13 levels. Each level has its own toggle-wait-toggle-wait routine, and then we shift another bit and decide whether to branch to the next (level+1) routine or the previous (level-1) routine.
Reminds me of an experiment that we did with TI calculators in the late 90s. Someone found that you can put an AM radio next to the calculator and it would beep based on what instructions were run on the Z80. There were some demos that played songs via 1 bit beeps like this, but since games actually have to run instructions, you couldnt reliably play background music during gameplay.
I was one of the people who added sound into a TI83 game. It ended up in my IceClimb game. Not particularly well done or anything, it's just there. I also didn't properly adjust the timing as the notes got higher, so it plays faster when you get higher in the level.
A friend of mine built his entire career as a composer around 1-bit music (famously in the form of an album that is physically a circuit inside a jewel case) https://www.youtube.com/watch?v=dBUVKNqdsEo
Delta-sigma modulation generates an inefficient encoding of the audio signal, with much more bits than necessary.
Converting a delta-sigma encoded bit stream to a PCM (pulse-code modulation) stream (e.g. with 24-bit or 16-bit samples at a sampling frequency of 48 kHz or 44.1 kHz) is a method of data compression.
On the other hand, delta-sigma modulation is more efficient for both analog-to-digital and digital-to-analog conversions, in the sense that for a given quality of the conversion it is much easier and much cheaper to make ADCs and DACs with delta-sigma modulation than with pulse-code modulation.
Because of this, for audio signals, normally delta-sigma ADCs and DACs are used at the analog inputs and outputs, but the delta-sigma bit stream is converted by digital filtering to PCM for data storage or for audio processing.
Delta-sigma is not a compression technique. A good delta-sigma encoder will use roughly the same number of bits as a PCM encoder. Delta-sigma uses more samples per second but fewer bits per sample to capture the same information. The reason delta-sigma is desirable is noise shaping, not compression.
An ADC that merely samples much faster than Nyquist but with the same number of bits per sample as PCM is not a delta-sigma converter; it's an oversampler. Delta-sigma converters always sample faster and use fewer bits per sample (usually just one bit per sample).
No. For equivalent quality, even a high-order delta-sigma encoder will need more bits than PCM.
The ratio between the sampling frequencies is greater than the ratio between the number of bits per sample. How much greater it is depends on the order of the sigma-delta modulator and higher-order modulators need much less extra bits, but they have other problems, which prevent the increase of the order too much.
Unlike PCM, delta-sigma modulation does not provide any guarantee about the maximum error at a given point in time, i.e. about the maximum instantaneous difference between the original signal and the encoded signal. Delta-sigma modulation only guarantees that certain statistical properties of the original signal and of the encoded signal are the same. For example in the case of the first-order delta-sigma modulator it is guaranteed only that the difference between the integrals over an interval of time of the original and of the encoded signals is smaller than a superior limit that depends on the length of the time interval.
This is why even for high-order delta-sigma modulators the ratio between their sampling frequencies and the sampling frequency of an equivalent PCM signal must remain great, because the sampling period of the PCM signal must be long enough so that the statistics of the delta-sigma signal accumulated during it make sense.
I do not like the term "noise shaping" and it is somewhat misleading, because it is not an intrinsic property of the delta-sigma modulator, but in order for the noise to be "shaped", it must actually exist. That means that when the input signal does not include any noise, the output of the delta-sigma modulator will oscillate periodically and deterministically around a value and there will be no "noise shaping". However, real electronic circuits always have noise, and when that noise is insufficient, in delta-sigma ADCs additional noise is injected at input, to prevent the periodic oscillations at output.
A simple oversampler can be regarded in some sense as a zeroth order delta-sigma modulator, because some of the formulae for delta-sigma modulators apply to it when the order is set to 0.
Just oversampling cannot produce any improvement in the resolution of a PCM signal obtained by filtering the oversampled signal. But if white noise with an amplitude equal to the quantization interval is added at the input of the oversampler, then by averaging the output, i.e. by low-pass filtering it, additional bits per sample can be obtained depending on how long is the averaging interval. This is the well known statistics property that the average of N random samples will have a lower variance than the original samples, which decreases for increasing N.
The same happens with delta-sigma modulators, the difference is only that the number of additional bits of resolution per PCM sample grows much faster with the length of the averaging interval than in the case of the simple oversampler, the higher the order of the delta-sigma modulator, the more additional bits at a given averaging interval.
Nonetheless, while the simple formula that expresses the achievable signal-to-noise ratio, i.e. the number of bits per PCM sample, as a function of the averaging interval, looks like one could get infinite resolution by just increasing the order of the delta-sigma modulator, that does not work in reality, where the order is limited to rather low values and the formula for the additional bits of resolution is not valid for very short averaging times, when the higher-order statistics are meaningless.
Super cool. Anyone recommends an oscilloscope for this day and age? I loved the Tektronix ones, but haven’t touched one in 16 years. I think it would be nice to have one to look into these kinds of signals - I know they are relatively simple enough to imagine, but this looks like something cool to show to kids.
The USB and handheld ones are great and can be had for less than $100. The Hanteks can be interfaced with open source software which is great since you get a ton of features that normally wouldn't be included in budget models (like wire protocol decoding). As for benchtop models, Siglent and Rigol are going to be the most affordable. There's a way to generate license keys from certain models if you care. Obviously Tektronix and Keysight are top end but are like 10x the price of Siglent/Rigol.
It would be interesting if someone could implement or adapt Roman Black's BTC algorithm to model the Speccy beeper.
https://www.romanblack.com/BTc_alg.htm
I hope everybody knows Tim Follin - or will now come to know his genius.
> Agent X ( https://www.youtube.com/watch?v=T42WuUpBuHE ) or Agent X II ( https://www.youtube.com/watch?v=gNc_xczyGLc ) or Chronos ( https://www.youtube.com/watch?v=u-D24A_N4d4 ) or Raw Recruit ( https://www.youtube.com/watch?v=kl8dAVybwq8 ) or Future Games ( https://www.youtube.com/watch?v=orEXKOBIv_8 )... // A modern attempt, the ON and OFF album by Rich 'Tufty' Hollins ( https://www.youtube.com/watch?v=4nEfO4Yu7Mg )
(Repost originally for submission Furnace - the biggest multi-system chiptune tracker ever made https://news.ycombinator.com/item?id=41609254)
It's just Pictionary Tim! He got even crazier when they gave him more than a single system beeper.
Silver Surfer (https://m.youtube.com/results?sp=mAEA&search_query=silver+su...) Pictionary (https://m.youtube.com/results?sp=mAEA&search_query=pictionar...)
Wow, that Chronos track is like early My Bloody Valentine!
The article ends to a link to a video that shows off the technique and it blows my mind the Spectrum beeper could produce this: https://www.youtube.com/watch?v=7cOu32m3wvQ
Turns out that it's actually 2-bit and not 1-bit. It uses a combination of the EAR and MIC jacks, and when you output through the MIC jack, it slightly boosts the EAR jack. That gives you 4 possible levels, but they are not evenly balanced.
I gotta try this on real hardware, reading the original article it seems it's barely audible without an amplifier, unlike on an emulator.
Will post an update after I get back to my parent's home, where I have these beauties that launched me into computer science: https://imgur.com/a/GKrA0vX
I created a program like that back in the day. In my case, I recorded my voice through the "MIC" input and played it back through the TV. A few seconds of sampled audio required the entire available memory. While I fully understood what was going on, I thought it was magical that the Spectrum could "speak with my voice".
On ZX-Spectrum? There was no "MIC" input on ZX-Spectrum, not that I know of. I was in high school at the time with only a rough understanding of the hardware but it was kinda obvious to me that whatever I plugged into the casette input port was only read as "0" or "1" by the IN <port> (forgot which) instruction. So not really possible to record voice at the time (didn't even know of PCM anyway), but I tried to decode Morse transmissions from radio. Never got anywhere with the software but my ear got good enough at some point that I could decode them myself by hearing :)
There was Z80 code to record from the MIC socket in one of the small books (Melbourne House maybe?) that came out in the early 1980s. I remember borrowing it from the library and it worked! Heavily quantized of course by the snapping to 2 or 0.
Also some games had sound samples, Activision’s Ghostbusters had a couple.
With proper volume control, even one bit could probably sound voice-like enough to a person in the 1980s.
It may be feasible to retrieve a pcm for a very short clip by playing the same recording at different levels which then could be converted on the same machine to a sequence of opcodes playing a more realistic voice.
On the Speccy at the time we had no volume control, only frequency... or more accurately you could but the frequency and the volume were tied somehow: https://retrocomputing.stackexchange.com/questions/14655
You would adjust the volume on your tape player to do what I proposed.
> There was no "MIC" input on ZX-Spectrum, not that I know of.
Don't be daft. Of course there was. It could save programs onto cassette as well as load them from cassette.
The ports are labelled MIC and EAR because that's what you connected them to on the cassette recorder.
Nope. I did the same thing myself with my Spectrum. I guess I was 14. I didn't have a microphone so I tried it with my dad's stereo headphones. I got the idea to try it after looking at the Z80 machine code of cassette saving and loading routines. Maybe it was also after ETX came out, first game I remember with speech samples. (Terrible game). I forget the details now but it's all about the timing between turning the mic/speaker on or off.
More mind blowing, that entry took _second_ place! First place would probably break my brain for the day so I won't look it up.
Release party: DiHalt 2026 summer Compo: Demo Platform: zx spectrum Ranked: 2nd
Sounds pretty good. Reminds me of the Demo Scene and the types of music they were able to create.
I recall, One bit sound was detailed extensively in Byte Magazine starting in mid 70s. A square wave that has its frequency varied. A google of the words " Generating speech with a square wave wave , Byte magazine " picks up lots links about this technique. Fourier transform show a square wave consists of odd sine wave harmonics. [A more advanced technique (more than one bit) would use microprocessor interrups to generate several sqare waves, output them with more than one IO pin and combine them with a resister network]
This mention of one bit sound made me remember the Lis’ner 1000 project of Steve Ciarca , cira 1984 , WHICH WAS speech recognition system of about 30 words. It used a GI SP1000 chip
> multichannel sound through the 1-bit speaker—basically, playing chords.
Ahem, that's polyphonic sound, not multichannel sound. Multichannel sound is playback through multiple speakers.
Perhaps. I think these words have somewhat flexible definitions, and that thought process leads me to some rhetorical questions.
Questions like: How many channels can a mod file[1] have? Does that number of channels this file has change depending on whether it is being played back with monophonic or stereophonic (or some other arrangement of) speakers? What happens to the count when one or more individual samples within that mod file contain polyphonic data (chords!), themselves?
It also leads to rhetorical statements, such as: When I play a film that has a soundtrack with 8 (ie 7.1) discrete audio channels on a stereophonic pair of speakers with a sane playback chain, I'm definitely still hearing all 8 of those channels. The same happens if I extend that stereophonic system to greater number of output channels, by adding surround speakers (quadraphonic!) or whatever.
Or maybe I take one earbud out and put it back in its charging case, and through the magic of active electronics, good software, and wireless comms, the output deliberately collapses to being monophonic.
However it happens, I've still got an 8 channel of film soundtrack in my ear(s).
And of course, it also works the other way: If the electronic musical keyboard in front of me is said to support 10-voice polyphony, then that's a hard limit on the number of notes it can produce at once regardless of the number of audio output channels.
Multichannel and polyphonic are pretty bendy terms, I think. Maybe that flexibility was wrong at some point, but they've been flexible in this way for so many decades that it definitely seems right to accept the overlapping uses as being correct.
[1]: https://en.wikipedia.org/wiki/MOD_(file_format)
> My previous work with PWM relied on being able to set a hardware timer to control when the speaker turns on and off.
I've got an Arduino project where I did that. The output is on a single digital pin, audio encoded as PWM. It turns the output off at the beginning of an audio frame, and sets a timer to trigger an interrupt when it's time to turn the output on.
I've got a memory block of event structs, each event representing a write to NES sound hardware. I emulate enough of the features to support everything that the Legend of Zelda title theme uses. 16MHz is more than enough to decide on the sample for 2 square channels, 1 triangle, and 1 noise, outputting a 6-bit output sample at a 31250Hz sample rate. The output is only slightly scratchy; the polyphony is actually surprisingly smooth.
I did a PWM demo on the Apple II using a format similar to the DPCM samples on the NES. Those samples have a pretty simple format: a 1 bit raises the amplitude, a 0 bit lowers it. So you're kinda approximating the waveform with a bunch of triangles.
The NES has a 7-bit DAC, but since we're bit-banging with the CPU we only have enough time between samples for about 13 levels. Each level has its own toggle-wait-toggle-wait routine, and then we shift another bit and decide whether to branch to the next (level+1) routine or the previous (level-1) routine.
https://8bitworkshop.com/v3.12.1/?platform=apple2&file=delta...
Reminds me of an experiment that we did with TI calculators in the late 90s. Someone found that you can put an AM radio next to the calculator and it would beep based on what instructions were run on the Z80. There were some demos that played songs via 1 bit beeps like this, but since games actually have to run instructions, you couldnt reliably play background music during gameplay.
I was one of the people who added sound into a TI83 game. It ended up in my IceClimb game. Not particularly well done or anything, it's just there. I also didn't properly adjust the timing as the notes got higher, so it plays faster when you get higher in the level.
People may enjoy this too https://www.youtube.com/watch?v=_6f8PURcPEE
End result of it playing 1-bit music: https://www.youtube.com/watch?v=IjTihhFG03o
A friend of mine built his entire career as a composer around 1-bit music (famously in the form of an album that is physically a circuit inside a jewel case) https://www.youtube.com/watch?v=dBUVKNqdsEo
Turns out 1-Bit encoding is the most efficient solution, even for the highest end of audio quality: https://en.wikipedia.org/wiki/Delta-sigma_modulation
Depends on what you mean by "efficient".
Delta-sigma modulation generates an inefficient encoding of the audio signal, with much more bits than necessary.
Converting a delta-sigma encoded bit stream to a PCM (pulse-code modulation) stream (e.g. with 24-bit or 16-bit samples at a sampling frequency of 48 kHz or 44.1 kHz) is a method of data compression.
On the other hand, delta-sigma modulation is more efficient for both analog-to-digital and digital-to-analog conversions, in the sense that for a given quality of the conversion it is much easier and much cheaper to make ADCs and DACs with delta-sigma modulation than with pulse-code modulation.
Because of this, for audio signals, normally delta-sigma ADCs and DACs are used at the analog inputs and outputs, but the delta-sigma bit stream is converted by digital filtering to PCM for data storage or for audio processing.
Delta-sigma is not a compression technique. A good delta-sigma encoder will use roughly the same number of bits as a PCM encoder. Delta-sigma uses more samples per second but fewer bits per sample to capture the same information. The reason delta-sigma is desirable is noise shaping, not compression.
An ADC that merely samples much faster than Nyquist but with the same number of bits per sample as PCM is not a delta-sigma converter; it's an oversampler. Delta-sigma converters always sample faster and use fewer bits per sample (usually just one bit per sample).
No. For equivalent quality, even a high-order delta-sigma encoder will need more bits than PCM.
The ratio between the sampling frequencies is greater than the ratio between the number of bits per sample. How much greater it is depends on the order of the sigma-delta modulator and higher-order modulators need much less extra bits, but they have other problems, which prevent the increase of the order too much.
Unlike PCM, delta-sigma modulation does not provide any guarantee about the maximum error at a given point in time, i.e. about the maximum instantaneous difference between the original signal and the encoded signal. Delta-sigma modulation only guarantees that certain statistical properties of the original signal and of the encoded signal are the same. For example in the case of the first-order delta-sigma modulator it is guaranteed only that the difference between the integrals over an interval of time of the original and of the encoded signals is smaller than a superior limit that depends on the length of the time interval.
This is why even for high-order delta-sigma modulators the ratio between their sampling frequencies and the sampling frequency of an equivalent PCM signal must remain great, because the sampling period of the PCM signal must be long enough so that the statistics of the delta-sigma signal accumulated during it make sense.
I do not like the term "noise shaping" and it is somewhat misleading, because it is not an intrinsic property of the delta-sigma modulator, but in order for the noise to be "shaped", it must actually exist. That means that when the input signal does not include any noise, the output of the delta-sigma modulator will oscillate periodically and deterministically around a value and there will be no "noise shaping". However, real electronic circuits always have noise, and when that noise is insufficient, in delta-sigma ADCs additional noise is injected at input, to prevent the periodic oscillations at output.
A simple oversampler can be regarded in some sense as a zeroth order delta-sigma modulator, because some of the formulae for delta-sigma modulators apply to it when the order is set to 0.
Just oversampling cannot produce any improvement in the resolution of a PCM signal obtained by filtering the oversampled signal. But if white noise with an amplitude equal to the quantization interval is added at the input of the oversampler, then by averaging the output, i.e. by low-pass filtering it, additional bits per sample can be obtained depending on how long is the averaging interval. This is the well known statistics property that the average of N random samples will have a lower variance than the original samples, which decreases for increasing N.
The same happens with delta-sigma modulators, the difference is only that the number of additional bits of resolution per PCM sample grows much faster with the length of the averaging interval than in the case of the simple oversampler, the higher the order of the delta-sigma modulator, the more additional bits at a given averaging interval.
Nonetheless, while the simple formula that expresses the achievable signal-to-noise ratio, i.e. the number of bits per PCM sample, as a function of the averaging interval, looks like one could get infinite resolution by just increasing the order of the delta-sigma modulator, that does not work in reality, where the order is limited to rather low values and the formula for the additional bits of resolution is not valid for very short averaging times, when the higher-order statistics are meaningless.
Super cool. Anyone recommends an oscilloscope for this day and age? I loved the Tektronix ones, but haven’t touched one in 16 years. I think it would be nice to have one to look into these kinds of signals - I know they are relatively simple enough to imagine, but this looks like something cool to show to kids.
The USB and handheld ones are great and can be had for less than $100. The Hanteks can be interfaced with open source software which is great since you get a ton of features that normally wouldn't be included in budget models (like wire protocol decoding). As for benchtop models, Siglent and Rigol are going to be the most affordable. There's a way to generate license keys from certain models if you care. Obviously Tektronix and Keysight are top end but are like 10x the price of Siglent/Rigol.
It would be interesting if someone could implement or adapt Roman Black's BTC algorithm to model the Speccy beeper. https://www.romanblack.com/BTc_alg.htm
reminds me of: https://github.com/protodomemusic/mmml
Some apparent links don't appear to work at the start of the page, and I couldn't see any links to listen to the output?
Another incredible 1-bit sound demo for ZX Spectrum in just 9KB from Dmitry Milk:
https://www.youtube.com/watch?v=98PJFIIJDiA
https://speccy.pl/archive/prod.php?id=833
youtube channel hacker@CK3GMR
@CK3GMR