at a certain point i lost the interest in making music - somewhere in my early 30's.
Uncategorized
68
Posts
7
Posters
3
Views
-
-
@TomF @lritter @Farbs @aeva Iām not so sure float is actually improving anything. 14-bits is all you need for the final mix. 24-bits lets you record a whisper and a jet engine without having to adjust the gain between takes. 32-bit lets you have all that plus room to cover accumulating quantization error in you DSP chain. Fixed point is fine. I think itās more about hardware performance or developer convenience than sound quality.
-
@TomF @lritter @Farbs @aeva Iām not so sure float is actually improving anything. 14-bits is all you need for the final mix. 24-bits lets you record a whisper and a jet engine without having to adjust the gain between takes. 32-bit lets you have all that plus room to cover accumulating quantization error in you DSP chain. Fixed point is fine. I think itās more about hardware performance or developer convenience than sound quality.
@jkaniarz @Farbs @lritter @aeva @TomF I think your comments around performance and developer convenience are spot on, but writing fixed point code is awful for anything of significant complexity. Cascading IIR filters with low frequency poles leads to complexities in analysing headroom requirements. Mixing large numbers of channels requires level analysis. And if processing is occurring in some filterbank domain, then you have to consider fixed point FFTs.
-
@jkaniarz @Farbs @lritter @aeva @TomF I think your comments around performance and developer convenience are spot on, but writing fixed point code is awful for anything of significant complexity. Cascading IIR filters with low frequency poles leads to complexities in analysing headroom requirements. Mixing large numbers of channels requires level analysis. And if processing is occurring in some filterbank domain, then you have to consider fixed point FFTs.
-
@nickappleton @jkaniarz @lritter @aeva @Farbs Right. I'm normally "Mr Fixed Point" but in this case zero has a very good meaning, and having highly dynamic range around it is very valid and useful and solves a lot of annoying problems.
-
@nickappleton @jkaniarz @lritter @aeva @Farbs Right. I'm normally "Mr Fixed Point" but in this case zero has a very good meaning, and having highly dynamic range around it is very valid and useful and solves a lot of annoying problems.
@TomF @nickappleton @aeva @Farbs what @jkaniarz said is part of what i was thinking of.
use 32 bit integers. for n channels being added, keep log2(n) msb free. scaling down is lossy, so pre-aggregate factors in the entire processing pipeline: normalise all volume levels in a partition of unity, so in combination they use maximum range. apply 1 post scale.
but tom and nick are right in that floats are just convenient. the by far nicest option on x64 is to just set the CPU to zero denormals
-
@TomF @nickappleton @aeva @Farbs what @jkaniarz said is part of what i was thinking of.
use 32 bit integers. for n channels being added, keep log2(n) msb free. scaling down is lossy, so pre-aggregate factors in the entire processing pipeline: normalise all volume levels in a partition of unity, so in combination they use maximum range. apply 1 post scale.
but tom and nick are right in that floats are just convenient. the by far nicest option on x64 is to just set the CPU to zero denormals
@lritter @aeva @jkaniarz @TomF @Farbs 32-bits is usually enough for most things. Iām still scarred from working on 24-bit DSPs which are utter pricks for complex (I.e. involve FFTs) signal paths and the quality can noticeably drop. Regarding saturation, many DSPs have saturating ops and ARM NEON has some basic ones for additive operations.
-
@lritter @aeva @jkaniarz @TomF @Farbs 32-bits is usually enough for most things. Iām still scarred from working on 24-bit DSPs which are utter pricks for complex (I.e. involve FFTs) signal paths and the quality can noticeably drop. Regarding saturation, many DSPs have saturating ops and ARM NEON has some basic ones for additive operations.
@nickappleton @aeva @jkaniarz @TomF @Farbs could also escape to 64 bit int ops and saturate there but this limits cpu targets. there's also the tiny teensy matter of emulating transcendentals.
anyway. all this is low level hell. we were talking about high level heaven - the reason why we do all of this: blessed emergence.
-
@TomF @nickappleton @aeva @Farbs what @jkaniarz said is part of what i was thinking of.
use 32 bit integers. for n channels being added, keep log2(n) msb free. scaling down is lossy, so pre-aggregate factors in the entire processing pipeline: normalise all volume levels in a partition of unity, so in combination they use maximum range. apply 1 post scale.
but tom and nick are right in that floats are just convenient. the by far nicest option on x64 is to just set the CPU to zero denormals
@lritter @TomF @nickappleton @Farbs @jkaniarz i just put all the sound numbers in the float32 hole š¤
-
@lritter @TomF @nickappleton @Farbs @jkaniarz i just put all the sound numbers in the float32 hole š¤
@lritter @TomF @nickappleton @Farbs @jkaniarz oh wait no I forgot. I put all the sound numbers in the float64 hole, because it gives a richer, warmer sound
-
@nickappleton @aeva @jkaniarz @TomF @Farbs could also escape to 64 bit int ops and saturate there but this limits cpu targets. there's also the tiny teensy matter of emulating transcendentals.
anyway. all this is low level hell. we were talking about high level heaven - the reason why we do all of this: blessed emergence.
@lritter @nickappleton @jkaniarz @TomF @Farbs just toss them into doubles and get fancy later if you actually need it
-
@lritter @TomF @nickappleton @Farbs @jkaniarz oh wait no I forgot. I put all the sound numbers in the float64 hole, because it gives a richer, warmer sound
-
@lritter @TomF @nickappleton @Farbs @jkaniarz oh wait no I forgot. I put all the sound numbers in the float64 hole, because it gives a richer, warmer sound
@aeva @TomF @nickappleton @Farbs @jkaniarz i wonder if there is a way to make float32 degradation audible. maybe this: add very loud signal A to very quiet signal B, then subtract signal A, normalise signal B. it should now have clicks where it quantizes to zero... or rather, the interference pattern should be more complex. it only pops around where A had max amplitude
-
@aeva okay. so https://en.wikipedia.org/wiki/Datalog#Example
read only the example section. this is the entire language. this is not an intro, a sneak peek (sorry for the typo, a sneak peak is something else), *this is it*. when you have understood the example, you have understood datalog.
@lritter I'm not any less tired but I took another stab at trying to see it. As I understand it, datalog patches describe the explicit and implicit connections between things, and high level audio synthesis is often described as a set of connections, between modules, so I could see that angle working since the math is pulled out of focus, but the implicit connections part is still fuzzy to me. Is this more or less along the lines of what you have in mind?
-
@lritter I'm not any less tired but I took another stab at trying to see it. As I understand it, datalog patches describe the explicit and implicit connections between things, and high level audio synthesis is often described as a set of connections, between modules, so I could see that angle working since the math is pulled out of focus, but the implicit connections part is still fuzzy to me. Is this more or less along the lines of what you have in mind?
@lritter the other angle that comes to mind at the moment is to use this as a sequencer, which isn't really obvious how that would work except that datalog feels markovian, and markov chains etc make sense to me as a procedural sequence generator and I've had some good results using them to generate interesting midi streams
-
@lritter I'm not any less tired but I took another stab at trying to see it. As I understand it, datalog patches describe the explicit and implicit connections between things, and high level audio synthesis is often described as a set of connections, between modules, so I could see that angle working since the math is pulled out of focus, but the implicit connections part is still fuzzy to me. Is this more or less along the lines of what you have in mind?
@aeva it's also relational (in the einstein sense), which allows you (and the compiler) to radically shift perspective with few changes. you can invert a function just by reading it from the outside in. it's particularly well suited to graph processing. in soufflƩ, a datalog dialect, a toposort is 3 lines. a SCC analysis is less than 10 lines. a raytracer should be similarly dense. and you're thinking data oriented rather than functional - good for entity component systems. for audio:
-
@aeva it's also relational (in the einstein sense), which allows you (and the compiler) to radically shift perspective with few changes. you can invert a function just by reading it from the outside in. it's particularly well suited to graph processing. in soufflƩ, a datalog dialect, a toposort is 3 lines. a SCC analysis is less than 10 lines. a raytracer should be similarly dense. and you're thinking data oriented rather than functional - good for entity component systems. for audio:
@aeva say we generate a sample of a sine with:
samplerate(44100).
duration(0.5).
pitch(440).
sample(i, sin(t * TAU * p)) :- samplerate(hz), duration(s), range(i, 0, hz * s, 1), t = i / hz, pitch(p).sample() then becomes a wavetable.
challenge: also generate wavetables for pitch 220 and 880. how would you do it?
-
@aeva say we generate a sample of a sine with:
samplerate(44100).
duration(0.5).
pitch(440).
sample(i, sin(t * TAU * p)) :- samplerate(hz), duration(s), range(i, 0, hz * s, 1), t = i / hz, pitch(p).sample() then becomes a wavetable.
challenge: also generate wavetables for pitch 220 and 880. how would you do it?
@lritter am i doing it right
-
@aeva now do it for 120 pitches.
Gli ultimi otto messaggi ricevuti dalla Federazione
-
Romania continua la crisi economica. LāUnione valuta sanzioni per lo sforamento del deficit
@news
https://www.eunews.it/2025/11/07/romania-continua-la-crisi-economica-lunione-valuta-sanzioni-per-lo-sforamento-del-deficit/
Crescita ai minimi storici, deficit di bilancio oltre il 9 per cento e titoli di stato con interessi altissimi. La Commissione Europea presenterĆ a breve le note sulla situazione del deficit rumeno: Bucarest rischia sanzioni
-
Mi immagino i truffatori truffati che andranno piangendo dalla polizia
-
Dormi profondamente
e non ti preoccupare
per la mia insonnia ā¦ https://cctm.website/maram-al-masri-insonnia/Maram al Masri
#maramalmasri #insonnia #poesia #cctmwebsite #anoipiaceleggere #leggere
-
@elvecio sƬ ma alla fine sono funzionalitĆ che possono pure tornare utili, ma per il giro ristretto che bazzica il fediverso non so quanta differenza effettiva possano fare: alla fine sebbene il mezzo conti (e conta tantissimo, intendiamoci) ĆØ piĆ¹ un discorso di persone, come usano e quanto usano un sistema, poi le eventuali menate subentrano piĆ¹ o meno sempre in una forma o nell'altra
-
@Xabacadabra pensa che lĆ addirittura puoi scegliere di liberare a tutti la risposta ma allo stesso tempo puoi moderare i commenti, che saranno visibili solo dopo la tua moderazione, finite ma quasi infinite possibilitĆ , dopo sta alla persona usare in modo giusto le cose.
-
@Dio š poĀ®co D'io š
-
@Xabacadabra comunque considera che tutti gli aggiornamenti/aggiustamenti sviluppati su hubzilla, streams e altre sono arrivati dopo anni e anni di esperienza nella federazione/ fediverso, niente ĆØ venuto dal caso, c'ĆØ sempre un motivo.
-
@catavz sometimes, this is true, too
