Matthieu Amiguet

My EuroPython'25 talk is online

Thu, 27 Nov 2025 10:28:26 +0100

The video of my EuroPython 2025 talk How to use Python on a RPi to develop a custom headless guitar FX box is online!

For context, code, slides and links, please refer to the companion post.

Note that I was also lucky enough to give a lightning talk titled From Switzerland to Europython Prague 2025 by e-bike… and back!. You can see it from minute 12:45 in the following video (it should start directly at the right place when you click the play button):

More details about this bike trip can be found here.

EuroPython 2025 in Prague was a great experience: high quality talks, outstanding organisation and, above all, an extremely friendly and welcoming atmosphere! Looking forward to the next edition!

Twint, une belle réussite de souveraineté numérique?

Wed, 17 Sep 2025 08:56:06 +0200

Il y a quelques jours, Le point J, émission d’actualité de la Radio Télévision Suisse, a invité Anouch Seydtaghia, en qualité de “journaliste spécialiste en technologie au journal Le Temps”, pour parler de Twint, l’application de paiement mobile qui cartonne actuellement en Suisse.

Par deux fois, M. Seydtaghia s’y enflamme pour décrire à quel point Twint représente une réussite pour la Suisse en termes de souveraineté numérique. Vous pouvez écouter l’entier de l’émission ici mais je vous ai préparé un petit extrait si vous ne voulez pas écouter le tout:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

La souveraineté numérique? Ça m’intéresse, ça! Je me précipite sur le site de Twint pour télécharger l’application sur mon fidèle laptop sous Arch Linux, et là, qu’est-ce que je vois?

“L’app TWINT est disponible pour iOS et Android”. Aha.

Est-ce que ça marche mieux sur mon smartphone sous LineageOS, une variante google-free d’Android? Scanner le QR de téléchargement me donne un lien vers Google play, que je n’ai pas (parce que, justement, je me soucie de souveraineté numérique - et de protection de la sphère privée). Encore raté, donc.

Je dois manquer quelque chose… iOS est un produit développé par Apple, entreprise éminemment américaine. Cela contraste bien sûr avec Android, produit par Google aux États-Unis. Les apps Twint ne sont installables que par les shops respectifs de ces deux OS, gérés - faut-il le préciser - depuis les États-Unis. S’il prenait l’envie aux États-Unis de bloquer l’utilisation de Twint, ils pourraient donc le faire du jour au lendemain sans le moindre problème.

Parlerait-t-on de souveraineté alimentaire pour du tofu élaboré en Suisse à base de soja 100% importé? Ou pour du gruyère AOP vendu exclusivement sur Temu?

Il en va de même pour une application: si les plateformes auxquelles elle est destinées sont en mains étrangères, on peut la rendre aussi incroyablement suisse que l’on veut, ça ne fera pas avancer la souveraineté d’un iota.

La solution? Elle est devant vos yeux en cet instant même. En développant un service web respectueux des standard ouverts, on rend le service accessible par quiconque disposant d’un accès internet, quelle que soit sa plateforme. Ou alors, dans le cas particulier de Twint, il y a aussi une autre alternative, presque parfaitement souveraine, mais il se peut qu’elle ne plaise pas à tout le monde: le paiement en espèce.

Some Reflections About Britten's "Harpejji Notation System" — part 1

Mon, 15 Sep 2025 14:40:00 +0200

In 2020, when I started to share my experience about harpejji technique, I realized that I needed a way to communicate precisely about fingerings. I used a tab notation that allowed me to express what I wanted. In order to encourage information sharing in the community, I developed an editor that was later included in Marcodi’s player resources on their website. In the meantime, this notation has become some kind of de facto standard for sharing harpejji fingerings.

But of course this kind of tab notation can only go so far, and I’ve been toying with the idea of annotating regular sheet music with fingerings ever since, without finding a satisfying way.

That’s why I was thrilled when Jenova shared on Discord a link to a post by Michael Britten that proposed a full, coherent fingering notation system for the harpejji.

We had some exchanges with Michael in the meantime, and I could have written him an email about what I think of his system. But I think think this discussion ultimately belongs to the community, and that’s why I answer publicly on this blog (I also think these discussions could still be of interest in a few years, which rules out messaging platforms and social media that are at best very fragile when it comes to going back in time…).

So this post presents my reflections about Michael’s proposition. If you haven’t read his post and/or watched his video, do it now because I’m afraid my post won’t make much sense otherwise. You probably also want to watch Jacob Collier’s video and might want to have a look at its original transcription by Suilven Chambers, without fingering notation.

General considerations

First, I’d like to thank Michael for the enormous amount of work he put in proposing a full and coherent system. The tremendous job of annotating the whole Jacob Collier’s transcription shows his level of dedication to finding a system that works.

I think Michael’s system fundamental ideas make perfect sense. In particular, I like the following insights:

We don’t need a “Y axis” in the notation as it is already present in the notes themselves.
Use relative notation for strings, which is both more readable and allows to take advantage of the isomorphic property of the harpejji fretboard.
Use “bookmarks” (what he calls “references slashes”) to simplify notation when there are a lot of “jumps”.

However, I also have a few concerns about the details of his proposition. Mainly:

I don’t like the notation for absolute string position.
I think the system as it is now graphically mixes notes and fingerings and spreads fingering information too much. I think it would be more readable to group fingering information in one place.
The glyphs chosen by Michael look great, but I’m not sure how to reproduce them. Maybe we should select things that everybody can use, whatever software/platform/font they’re using (or even when annotating by hand).
As impressive as the Collier’s transcription is, I guess most of the time the fingerings won’t be as detailed. For pedagogical purposes, one might want to notate every fingering note by note. But in day to day work, a musician is likely to only notate fingerings in a few tricky places. I think this should be taken into account.

In the rest of this post, I’ll develop these consideration and suggest adaptations to the system.

Two more remarks before diving in:

I do appreciate and respect Michael’s work a lot. I’m suggesting adaptations of his system because I think we can be more clever collectively than alone, but that’s by no way a criticism of the awesome work he’s done!
I don’t think that Michael, or me, or anyone really, can decide of a notation that will become standard. We can discuss of a good solution and suggest ideas to the community, but the standard notation will emerge from the community along the years. So keep this in mind when reading all this - these are only ideas, that want nothing more than living and evolving for making the harpejji community more alive!

With that in mind, let’s consider my concerns about the method one by one.

Absolute string positions

To be honest, this is the only point where I completely disagree with the proposed solution. What I don’t like:

First and mostly: OK to call a string C string because its lowest note is C, but I categorically refuse to call a string A string “because” its lowest note is B♭ (same for F and G, of course). If we are to name the strings after their lowest note, so let’s really use the lowest note, not some random approximation!
The proposed solution is specific to the G16, and there’s no way to adapt it to another model without looking at the G16 specs.
The proposed notation conflicts with standard chord notation. In Michael’s Mary had a little lamb, any musician who did not read his post would interpret the first “G” as a G major chord (and wonder what the “A1” is).

Actually I’m not sure at all what a good solution would be. Technically a fret number would do the trick, but I never think of my harpejji in terms of fret number, and I know I’m not alone.

What I’d try at this point would be

Use the full note name with real octave as string name. So the G16 strings would be called C2, D2, E2, F#2, G#2, B♭2, C3, …, F#4.
To avoid conflicts with chord notation, write the string name in lower case just above (or below) the fingering, in parenthesis.

So the beginning of “Mary..” would be

This notation doesn’t conflict (too much) with chord notation, it doesn’t lie about string lowest notes, and I can find the nearest string on my K24 or M5 without having to lookup range information about the G16 before interpreting it correctly.

But is it more readable than the original version? barely. Does it look nice? Not really. I think this would be a better solution than Michael’s original idea, but actually I hope someone in the community comes up with a better idea!

(And if you can never remember the octave number for a given note, don’t worry, you’re not alone. I can’t either, because it’s not very useful in my daily practise. But if the standard notation includes this information, I’ll learn it quickly - and so will you! And in case of real difficulties, it’s always possible to put stickers with string names on the tuner channel!)

One last thing about absolute string notation: as has already been noted on Discord, most of the time we don’t really care about absolute position. Given the isomorphic nature of the harpejji, starting on any instance of a given note should work as the rest of the system uses relative notation. So most of the examples below will omit this information.

However, I still think we should be able to include this information, because there are cases where it is relevant. Suppose you want to play Debussy’s Clair de Lune on the harpejji (good luck!). When reaching this part, you’ll probably have to select very carefully on what string you start the left hand arpeggios in order not to collide hands and/or not to run out of strings:

(Sheet music CC-BY-NC by Les Éditions Outremontaises)

Other possible uses of absolute positioning include repeats in classical music, where jumping back in the score might imply considerable hand shifting (and loss of context), and special effects like playing on very high frets of lower strings (I love how it sounds!).

Grouping fingering information

Michael’s string position modifiers are a great idea. They mostly match how I think about the instrument and allow to take advantage of the isomorphic property of the fretboard.

However, I’m not sure the placement is optimal. Let’s have a look at his simple “Mary had a little lamb”:

Let’s focus on the right hand. To know what to do, I must look at the note, then down to the string modifier, then up to the finger number. The information about the fingering is distributed over two relatively remote places.

I think it would be much easier to follow if the information was grouped in the same place. Something like

I find it easier to follow, with a few added benefits:

The fingerings don’t look like accents, and it is still possible to add accents to the music if desired.
The music itself and the fingerings are in separate places. That means the music is still readable by someone playing another instrument, with no fingering information in the way.

Using standard glyphs

Now I have another concern. Michael found very nice glyphs for his ideas, but as far as I know, they are not standard Unicode characters. I managed to reproduce something similar in my engraving software, but it’s nowhere as nice, and it was already difficult to do.

So I’m wondering if that wouldn’t be a good idea to use standard glyphs to make access to this system easier.

For notating “skip one string to the right (or left) and play on the following one with finger #3”, Michael uses something like •>3 and <•3, but with custom glyphs. To stay with Unicode characters, I considered

..3 and 3..
''3 and ,,3
²3 and 3²
… and some others.

But most of them are not very readable and/or need too much space.

Looking at existing Unicode characters, I found this solution which seems interesting:

1 string left or right: ․ (standard point or Unicode #2024)
2 strings left or right: : (standard colon)
3 strings left or right: ⁝ (Unicode #205D)
4 strings left or right: ⁞ (Unicode #205E)

Bigger leaps could be combinations:

5 = 4+1: ⁞․
6 = 4+2: ⁞⁚ (I’m using U+205A for the 2 here)

So we would get something like

I’m aware that this is slightly more abstract than the very intuitive version of Michael. But I think it is still very readable with some practise, doesn’t take much space and uses widely available Unicode characters. Note also that we will see another, potentially simpler, version of this example after having introduced bookmarks.

You might wonder how to type these characters on your computer. The specifics vary from platform to platform; depending on your preferences,you could simply copy-paste the above characters, search the internet to find out how to type Unicode characters on your machine, or (as I did) configure your keyboard to add key combinations that type them.

One thing that bothers me is that there’s an off-by-one incompatibility in the number of dots as compared with Michael’s version: > gets ․ and •> gets :. That means that the two systems wouldn’t cohabit very well so we as a community had better selecting a version quickly and sticking to it!

Potential conflicts

The notation proposed above removes some potential conflicts with existing musical notation, especially nuances and accents. Some confusion might remain in the specific case of figured bass but I think that using different places for fingerings and figuration removes any ambiguity in interpreting the notation:

Progressive enhancement

In pedagogical material or very complicated pieces (like the Collier’s transcription), it can be interesting to notate precisely every fingering detail over the whole piece. But in many practical cases, we are likely to only notate a fingering here and there, where we feel they’re not intuitive.

I think Michael’s proposition (with or without my suggestions) is perfectly usable in this case, except for one thing: there’s no way to notate “play this note on the same string as the last one”. If you notate every fingering, that’s not a problem: just don’t write any modifier. But if you want to notate only some fingerings, you need to be able to specify this. I’d suggest to prefix the finger number with an equal sign (=).

For instance, if I was annotating “Mary…” just for myself, and decided to begin with finger 3 and play the d and c on the same string, I’d probably only notate this, the other fingering being completely logical to me:

Slashes

If we try to group all fingering information in the same place, we still have to find something for the slashes of Michael proposition. He uses them in two ways: Slashed chords and reference slashes.

Slashed chords

I tried to use a simple slash (/) in the fingerings, but I don’t think it’s very legible. I’d suggest to rename slash chords into anchored chords (following Michael’s own terminology: “Slashed chords are anchored in space”) and draw a circle around the anchoring note. For instance, a F major cadence with the voicings described here would look like this:

Note that if your software makes it difficult to draw a circle around a fingering information, you could also use Unicode characters #2460 and following: ①, ②, …

Reference slashes

The other use of the slash is the great idea of reference slashes. If we try to get this back in the fingerings themselves, we cannot use exactly the same system, because the position of the slash in the staff matters.

My proposition would be to reserve some characters like *, +, @ and maybe others to serve as bookmarks, as follows: to bookmark a string position, write the symbol just after the fingering. To “use” this bookmark, write the same symbol just before the fingering. For instance, the same F major cadence as above, but arpeggiated could be notated like this:

For passages that have two voices in one hand, it’s possible to use several bookmarks. Let’s re-consider the example we used before when defining the position modifiers:

We could notate the exact same fingerings with one fixed bookmark * for the low C, and a “moving bookmark” + for the scale:

The 2* on the first note means “play with finger 2 and remember this position”. The subsequent * mean “play in same position, same finger”.

The .3+ on the second note means “move one string to the right, play with finger 3 and remember this position”. Then on the fourth note, +.4+ means “start from the remembered position, move one string to the right, play with finger 4 and remember this new position”.

This bookmark notation has an added benefit, especially when sight-reading: with Michael’s reference slashes, when you get to the slash, you have to look back to find where the last corresponding note was, and maybe remember with what finger you played it. With the bookmarks suggested above, you’ll already know when playing the “reference note” that you have to store this information in your mind, potentially removing the need to look back later.

Putting it all together

Well, I think that’s it. My new proposition might seem very different from Michael’s at first sight, but it’s actually using exactly the same concepts, with the following adaptations:

Make absolute string positions more logical;
Group fingering information in the same place;
Use widely available characters.

Let’s compare for example the two following measures from Collier’s transcription, first in Michael’s version:

Now with the modifications I’m suggesting:

I think the second version is more readable and a little bit less cluttered, but as I said before that’s not for me, nor for Michael, to decide, but ultimately for the community as a whole.

That could be it, but…

When writing this post, an idea got more and more insistent in my head: maybe we selected the wrong axis; maybe we should notate frets changes instead of string changes.

At first I thought that would be a rather equivalent way of doing things, but the more I think about it, the more reasons I see that it could work better.

This post is already quite long, so I won’t develop this idea here, but I intend to write a follow-up soon where I explore the pros and cons of using a fret-based notation system instead of a string-based one.

In the meantime, please react, comment, criticize - on Discord, on the Harpejji Hangout, on your own blog… Whatever, but we definitely need community interaction to collectively develop the best notation system. Since Michael opened the discussion (and thanks again for that!), now is the time, so express yourself!

Do you find this post useful? Would you like to see more content about the harpejji on this blog in the future?

Writing about harpejji technique is time-consuming. If you can afford it, please consider donating to support this work. Even a small donation helps, so don't hesitate!

You can support me through Les Chemins de Traverse, the cultural organization I co-created and I've been working for over the past 25+ years. Please go to their support page (in french) or use this direct Paypal link. Thank you!

EuroPython 2025 - Some context about my talk

Thu, 17 Jul 2025 08:00:00 +0200

This is a companion post to my EuroPython 2025 talk How to use Python on a RPi to develop a custom headless guitar FX box to provide some context and additional information. In no particular order:

UPDATE 27.11.2025 The video of the talk is available here!
The slides of the talk are here
The full code is available on bitbucket
pyo is a great module for realtime audio processing developed by Olivier Bélanger in Montréal.
pyo is great for low-latency audio processing, although getting reliable sub-10-ms roundtrip latency may require “some” tuning of the OS. My RMLL'17 and RMLL'12 talks give some details about the required settings on a generic Linux system and this more recent post gives details about how to do it on a rpi with Debian.
In the talk I’m mentioning the pisound and Patchbox OS from Blokas.
For more details on why we chose to work with python/pyo for realtime audio processing, see my Europython'19 talk and/or my Swiss Python Summit 2016 talk.
pyo heavily relies on callbacks, which can sometimes be inconvenient. For a way of replacing some callbacks with generators to get more readable code, have a look at my PyCon DE 2024 talk.
Most harpejji videos are from the Harpejji solo project.
This research was supported by Les Chemins de Traverse.
And if you’re interested, some details about the trip to Prague and back on e-bikes are available here!

I hope you enjoyed the talk! Don’t hesitate to contact me if you still have questions!

Harpejji M5 Review - Part 2: Sound

Mon, 09 Jun 2025 16:54:24 +0200

The first part of this review was focused on the outer aspects of the instrument. In this post I’d like to focus on the sound of the new harpejji M5, using the K24 as a comparison.

What we are comparing

As a reminder, we will be comparing a 2013 Bamboo K24 with a 2025 Maple M5. Both are stringed with the exact references suggested by Marcodi (which are Ernie ball Strings). The M5 has the original strings which were shipped with the instrument (so ~6 months old strings that travelled oversea) and the K24 has strings that are ~2 months older than that.

Both instruments are plugged directly into a Presonus AudioBox 1818VSL sound interface and recorded at 48kHz on a computer, then exported to high-quality mp3’s (mono 48kHz, variable bitrate 220-260kbps). No additional effect or amp simluation is added, so it might sound a little bit dry… It’s normal — nobody would ever record a song this way! The goal is not to provide clips that sound nice, but clips that allow an analytical comparison of the two instruments.

Of course, listening to these clips mainly makes sense when using headphones.

Overall volume

The first thing I noticed when comparing the two instruments side by side is that the K24 has a slightly louder output than the newer M5. Here are two recordings with exactly the same settings (volume and tone fully open, same recording gain), where I played the same thing on the K24 first, and then the M5, trying as hard as possible to play in the same way (although I find it difficult not to compensate for volume!).

First the K24

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

Then the M5

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

It’s a well-known fact in psycho-acoustics that the volume does influence the perception of other characteristics of the sound, so from now on I’ll adapt the gain to try and get the same output volume on the two instruments.

Hiss

In a video presentation of the M5, Landon Meeks from Marcodi says that the new pickup design needs “less volume compensation”, “inevitably” leading to “less hiss”.

After adapting the volumes as explained above, and using only my ears without any kind of measuring instrument, I can say that there is a tiny difference when the tone knobs are fully open: the K24 produces very slightly more noise in the upper part of the spectrum than the M5. But I usually play with the tone knob fully closed (see below for an explanation), and in this case I can’t hear a real difference between the two instruments.

I don’t include recordings for this section, because I’m not sure what part of the noise comes from my recording rig and what part from the instruments. Maybe someone more experienced with studio recording will provide a better comparison and/or measurements (I’m mainly a gig guy, and by no way a sound engineer!).

It might be that when playing through an effect that boosts the higher end of the spectrum a lot (like a strong distortion) and with the tone knob fully open the difference becomes significant. But for my use of the instrument, I don’t expect to feel any difference at all.

Note however that the new pickup design, even if it doesn’t really influence hiss, does influence other characteristics of the instrument, in particular what I call “mute pops”. More on that later.

Tone filter

In his video, Landon also mentions that they “revoiced” the tone filter. This is another quite obvious change between the K24 and the M5. The former has a tone filter that goes from “an awful lot of high frequencies” to “quite a lot of high frequencies”. I think that the M5 has a slightly wider tone setting, but overall it’s “rounder” - or less “brilliant” - (less high frequencies) than on the K24.

Here is a small musical phrase played…

on the K24 with the tone knob fully open:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

on the K24 with the tone knob fully closed:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

on the M5 with the tone knob fully open:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

on the M5 with the tone knob fully closed:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

But there’s more to it…

I think this more or less covers what Marcodi advertises concerning the sound of this new model. But there are two more aspects that I want to talk about; to my knowledge, there’s next to zero resources on these aspects on the web, so I don’t even think they have an official name yet. In my head, I call them bass gap and mute pops, so that what I’m going to call them in this post.

Bass gap

One thing I like less on my otherwise beloved K24 is what I call the bass gap: the eight bass strings and the 16 treble strings are equalized very differently internally, so there is a considerable change in sound quality when going from one side to the other. I’m not sure why they did it like that, but that’s definitely a fact about the instrument.

Here is my K24 played on the 4th fret from string #7 to string #10. With the tone knob fully open, it really sounds like two different instruments:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

Still on the K24, but with the tone knob fully closed, it’s already better, but there’s still a considerable gap:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

To mitigate this, I usually run my K24 bass side through an equalizer to try and get a more balanced sound over the whole instrument. That also explains why I’m almost always playing with the tone knob fully closed.

So when the M5 came out, I was wondering whether I could play the instrument directly through an amp, without no need for an equalization.

Let’s see: this is the M5 played exactly on the same place (fret #4, strings #7-#10 - of course, it will sound a minor third higher than on the K24). First with tone open:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

Then with tone closed:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

So as you can hear, if the M5 does reduce the bass gap, it does not suppress it totally. Depending on the amp, music and situation, the gap might go unnoticed on the M5, but in a studio setting or with a high quality full-range amp, it’s still definitely noticeable.

Mute pops

The other big untold secret of the harpejji sound is what I call mute pops.

The muting system of the harpejji is a blessing, and I truly love it. But muting/unmuting a channel is bound to make some noise, especially if you want an immediate effect. On the lower part of the instrument, that’s hardly noticeable. But to get a balanced sound across the instrument, it is necessary to boost the higher strings much more than the lower ones. So when playing softly in the top register, the noise of the muting system becomes very noticeable. Then add some reverb and it can get even worse!

There are workarounds - for instance, when letting the sound die at the end of a piece, I often use a volume pedal to fade out the sound before lifting my fingers to avoid these pops that would otherwise break the atmosphere. It works, but it’s still a workaround, isn’t it?

So does the M5 bring any improvement there? Let’s compare!

Here is a soft, slow melody first played on the three highest strings of the K24:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

And now the same melody on the three highest strings of the M5:

Oh oh... il semble que votre navigateur n'est pas capable de lire ce fichier audio :-(

As you can hear, the M5 is much better here! While the mute pops are still noticeable, they are much softer. I might still use the fade-out trick at the end of a piece, but I expect that it will be easier to get a nice, soft melody in the higher register on the M5 than on the K24.

(Temporary?) Conclusion

In this post, I reviewed what I consider the most important aspects of the sound of the new harpejji M5, per se and in comparison with the K24.

The M5 has a very round and balanced sound, with clear improvements over the K24, especially regarding mute pops and the bass gap. The instrument is a pleasure to play and is very rewarding.

In comparison, the K24 feels a little rougher, a little bit less balanced, and somehow less well-behaved. Still… I kind of like it too! This roughness contributes to the character of what I’m playing and sometimes this is exactly what I need. Also, I’d say that my K24 sounds a little bit more acoustic than my M5, which can be nice sometimes.

As an important reminder though: I’m not sure what part is due to the model (K24 vs M5), what part to the wood (Bamboo vs Maple), to the age of the instrument (12 years vs a few months), etc. - so take this comparison with a grain of salt.

When I bought the M5, I was wondering if it would replace my K24 completely. I know now that this will not be the case. My K24 and M5 are two great instruments and I’ll probably end up playing one or the other depending on the context and musical style.

That’s it! I think these two posts on the harpejji M5 cover most of what I feel important to say about the instrument for now. Maybe in the upcoming months, I’ll find other aspects that I want to share - in that case I’ll extend this multi-part review with a new post.

One thing is still missing, though: I’d like to record some real music (not only 10-seconds clips) on the K24 and M5, to demonstrate how the instruments behave in real situations. I’ll try and post something in the upcoming months, but don’t expect it to soon: other projects (and the way towards them) are likley to keep me busy for the next few weeks!

Do you find this post useful? Would you like to see more content about the harpejji on this blog in the future?

Writing about harpejji technique is time-consuming. If you can afford it, please consider donating to support this work. Even a small donation helps, so don't hesitate!

Harpejji M5 Review - Part 1: Outer Aspects

Tue, 27 May 2025 11:32:51 +0200

Context

I was lucky enough to discover the harpejji relatively early after it was commercialized - my beloved K24 has serial number #00063. I’ve been playing it for more than 10 years, documenting my progress and sharing tips through various posts on this site. Along the way, I also proposed a tab notation to share fingerings in the harpejji community and contributed to develop the tab editor that Marcodi later included on their website.

The “full-range” aspect of the instrument has always appealed to me, so when the g16 and the u12 came out, I wasn’t really interested, although they did include some technical innovations that I was curious about.

But of course when Marcodi announced the new M5 series that was completely different! A new full-range instrument (5 octaves) with improvements brought by almost 15 years of experience building harpejjis? I must admit I was extremely curious.

After some hesitations, I decided to take the jump and I ordered a M5. It seems that I’m relatively early again as I received an instrument with serial number #00018!

This post is the first in a series that should constitute a full in-depth review of the new instrument, with detailed comparison to the earlier K24.

Full disclosure Marcodi was kind enough to offer me a discount on the instrument so that I could write this review. I did accept the discount, but apart from that I’m not affiliated in any way to Marcodi. The things I write in this series of posts are purely my opinions, expressed completely freely!

Unpacking the M5

As far as I can remember, unpacking a harpejji in 2013 was just that: unpacking a harpejji (with cable and power supply). In 2025, I also got a “quick start guide”, a set of tools for tuning and tweaking the harpejji, a couple of string cleaning wipes and a nice, handwritten note from Joy Meeks.

The “quick start guide” includes basic information about harpejji parts and accessories, setup, tuning and string gauges. It also contains a QR code linking to a youtube playlist of harpejji How To videos. This is a good idea, although I feel the playlist could be updated and refreshed a little bit

I think the addition of these items really enhances the unpacking experience. Someone getting her first harpejji will probably feel a little bit less lost than 10+ years ago.

I got my instrument with a Gator gig bag. Although it feels a little bit less sturdy and definitely less comfortable than the Mono case I bought for my K24, I think it’s definitely up to the task - and much cheaper!

The harpejji M5, physical aspects

Dimensions

The M5 is slightly shorter and, oddly enough given the 2 strings less, slightly wider than the K24. But of course the string spacing is exactly the same and I don’t think the physical dimensions would make any big difference in any real situation (except that the M5 does of course not fit into the custom K24 flight case, which is discontinued anyway).

Wood

I selected the 13-ply maple model for my M5. When I ordered it, it was the middle model between the cheaper birch and the “premium” bamboo models. When writing this post, I see that the bamboo model is not available and the maple one is the “premium” model. I don’t know if this is just now or if they gave up producing the bamboo M5’s.

My K24 is out of bamboo. Funnily enough, when I ordered it, bamboo was the budget option, and maple was considered as an upgrade. Not sure what made Marcodi invert their value scale.

Anyway, that means that when comparing the two models, I will be comparing woods as well electronics and frets and whatnot. I’ll have to keep this in mind!

Note Markers

Maybe it’s out of habit, but I prefer the historic “rhombus” note markers rather than the newer “circle” ones (and I truly dislike “ABC” markers, so that was never an option). Hence rhombus is what I chose for my M5. Note that the markers on the new instrument are a few millimeters apart from the frets, unlike the older K24 where the markers were right behind the frets. I think the new design might lure beginners into hitting the strings a little bit too far from the frets, so I’m not sure this was a good design decision. But at least for an experimented player, the difference is very small and quickly gets forgotten.

Tuning Wrench

The K24 has a slot for keeping a small tuning wrench inserted into the instrument. I do like this setup, as the wrench is always there when you need it. For some reason Marcodi did not include this feature in the new M5, which come with a bigger, more comfortable but separate wrench. That’s a point where I prefer the older model, but of course that’s a tiny detail.

Knobs placement and shape

Now that’s a point where the new M5 clearly is an improvement. On the K24, the tone and volume knobs are on the player side of the instrument, and it happened more than once that I would bump the knobs - especially the volume one - and modify the settings while playing. Also, the knobs where round with no marker, so reproducing a given setup was rather difficult.

On the M5, the knobs are on the left side of the instrument, so accidental bumping by the player is impossible. Also, the new drop shape of the buttons make it much easier to quickly get to a given setting without guesswork.

(Note that my K24 has a custom mono switch, so the top image is not a typical K24 representation)

Frets

The M5 benefits from the new “pepperoni” frets Marcodi designed specifically for their instruments. I’m not sure what part of it is to be attributed to the new frets, but the instrument does feel smoother to play than the K24. The feeling is as if the string tension was slightly lower and/or the string wider: the strings do get a little bit less “into” the fingers when playing.

Of course to be sure I’d have to try two instruments with exactly the same specs except for the frets (and that might very well be that two such instruments simply do not exists), but in the end, who cares? The M5 provides a very nice feeling under the fingers, and that’s what’s important!

Saddles & pickups

The K24 is using standard guitar saddles that have been filed to accommodate the non-standard gauges. It works, but it means that the lower strings are closer to the pickups than the higher strings, because the saddle has been filed deeper. As the lower strings also naturally have a higher output volume, it necessitates a lot of volume compensation in the electronics to get a balanced volume throughout the instrument.

The M5 has the new, custom saddles that don’t require filing. Not only do they look nicer, but they need less volume compensation. According to Marcodi, that should mean less hiss, which I don’t find very perceivable, but I hear a huge difference regarding “mute pops”, which is great news. More on that later.

It might also be that the new saddles could simplify experimenting with different string gauges. I never felt the need, but I guess that someone wanting to put strings with other gauges on a K24 could have troubles with the specially filed saddles. If I understand well, all saddles on the M5 are identical, so experimenting with alternate gauges and/or tuning should be much easier. This is just a guess, though, that would have to be validated by Marcodi.

Cosmetics

The M5 also features a new Marcodi logo, and the so-called accent plates, which were plain white on my K24, are now “Glossy hand painted tortoise with deep embedded logos” (to quote Marcodi’s website).

To be completely honest, the way an instrument sounds is for me way more important than how it looks. I guess the new cosmetics give the instrument a more professional finish, and I’m sure this is important to many harpejji owners, but frankly I don’t really care.

Harpejji M5: Range and note placement

I left aside until now one of the most visible difference between the K24 and the M5 because, given its importance, I wanted to dedicate a full section to it.

Marcodi has decided to change the string number, scale length and overall range of the instrument. The K24 has 24 strings and 15 frets, with exactly 5 octaves ranging from A0 to A5. The M5 has fewer strings (22), more frets (19) and also covers 5 octaves, but a minor third higher (C1 to C6).

To me, the range shift is a blessing. I’m rarely using the very lowest notes of my K24, but I often find that a highest note at A5 is a little bit limiting, especially when trying to sight-read piano music. C6 is still not very high (remember I was trained as a classical flautist 😉), but it’s already much better.

The shorter scale length yields a narrower fret spacing. In the lower part of the strings, I feel the difference is barely perceivable. If anything, it makes things a tiny bit easier. I was afraid however that the very small spacing in the higher part of the strings, due both to the shorter scale length and the higher number of frets, would be difficult to master. But although it does take some time to get accustomed to it, the space between frets is still much larger than my fingers and there’s no real problem playing up there.

When improvising in the middle range of the instrument, I don’t really notice any difference in note placement between the K24 and the M5. When playing music I wrote for the K24 however, I sometimes run out of strings on the right side of the instrument. That means I’ve got either to modify my fingerings to play more “vertically” on the fretboard, or I’ve got to play on higher frets (hence more to the left to get the same notes). The second solution is sometimes slightly less ergonomic, because playing on higher frets means stretching the arms further away from the body and/or leaning forward a little bit more. It might be that the real solution there, if I want to keep the same ergonomics, is to transpose the music 1 or 2 tones lower (although it obviously has an impact on the music too!).

But all in all, I feel the changes in notes geometry completely make sense and make the instrument more suited to many musical situations.

(Temporary) Conclusion

In this post, I reviewed what I feel are the most important outer aspects of the new M5. I also compared them with the K24. But of course this is only the tip of the iceberg: the really important part is how it sounds, and how it helps you to make music. That will be the focus of a next post on this blog, so… stay tuned!

Do you find this post useful? Would you like to see more content about the harpejji on this blog in the future?

Writing about harpejji technique is time-consuming. If you can afford it, please consider donating to support this work. Even a small donation helps, so don't hesitate!

Ne dites plus IA, dites SALAMI!

Sat, 12 Apr 2025 17:14:09 +0200

L’expression “intelligence artificielle” introduit un biais cognitif dû à l’utilisation de termes mal définis mais fortement connotés. Stefano Quintarelli, l’un des spécialistes de la tech les plus en vue en Italie, propose dans un billet datant de 2019(!) de remplacer les termes “Artificial Intelligence” par “Systematic Approaches to Learning Algorithms and Machine Inferences” (qui pourrait se traduire en français courant par quelque chose comme “Approches systématiques des algorithmes d’apprentissage et et de raisonnement automatique”).

Essayez! Dorénavant, dites “J’écris tous mes mails avec un SALAMI”, “J’ai demandé à un SALAMI de me résumer la séance d’hier soir” ou encore “Je ne pourrais plus coder sans l’aide d’un SALAMI”. Affirmez qu’en vertu d’une nouvelle loi de Moore, le SALAMI devient deux fois meilleur tous les 7 mois. Clamez haut et fort que Le SALAMI s’impose comme un acteur central dans l’évolution de la médecine, offrant des perspectives prometteuses pour améliorer la qualité des soins et transformer la prestation des services de santé. Essayez enfin, lors de votre prochaine soirée mondaine, d’affirmer d’un air confiant: “Oui bien sûr, le SALAMI utilise une grande quantité d’énergie, mais je ne doute pas qu’il nous permette de découvrir des moyens d’en économiser encore plus”.

Sérieusement, essayez! Vous verrez que la connotation des mots qu’on utilise contribue considérablement à la crédibilité que l’on accorde - ou non - aux énoncés qui les contiennent.

… et tout en discutant, n’oubliez pas de servir pour l’apéro quelques tranches d’Ingrédients Alléchants, ça aide à faire passer les discussions animées! “Une p’tite tranche d’IA pour accompagner le prosecco?”

Technodépendance

Mon, 20 Jan 2025 09:17:13 +0100

On lit un peu de tout et pas mal de n’importe quoi au sujet de l’IA ces derniers temps. J’ai donc décidé d’apporter ma modeste pierre à l’édifice et de publier quelques textes à ce sujet. Docteur en IA, ayant fait de la recherche dans ce domaine de 1998 à 2003, et ayant enseigné cette matière en école d’ingénieurs de 2004 à 2010, je suis outillé pour comprendre de quoi on parle. Actuellement, je suis toujours avec intérêt ce qui se passe dans le domaine, mais de plus loin. Cela me donne un recul que n’ont pas certains spécialistes actuels de la branche, alors qu’eux sont imbattables sur les détails des développements récents. J’espère donc que mon éclairage atypique pourra aider quelques personnes à y voir plus clair dans ce sujet actuellement très chaud.

Après des textes sur la définition de l’IA, sur sa consommation énergétique, et sur les risques (ou non) que l’humain soit rattrapé par la machine, voici quelques considérations sur la dépendance à la technique.

Je me souviens de ma grand-mère qui préparait ses paiements. Après avoir fait le total à la machine à calculer, pour savoir combien d’argent prendre avec elle à la poste, elle recalculait la somme de tête, pour être sûre…

Trente ans plus tard, je vais rendre visite à mon neveu en études à Berlin. Il y habite depuis trois mois, mais quand je lui demande dans quel quartier il loge, il ne peut pas me répondre. Plutôt au sud? au nord? à l’est? à l’ouest? Aucune idée. Se reposant systématiquement sur le GPS de son smartphone pour se déplacer, il n’a aucune représentation géographique de la ville.

Les outils que nous propose la technologie sont extraordinairement pratiques. Mais en quelques décennies nous leur avons accordé une confiance énorme, au point de ne même plus garder, dans un coin de notre esprit, l’idée qu’ils pourraient nous induire en erreur.

Calculette? plus besoin de savoir compter. GPS? à la poubelle le sens de l’orientation. Véhicule autonome? plus besoin de savoir conduire - ni peut-être même marcher.

La grande promesse de l’IA est d’avoir des machines qui pensent pour nous. Avant de s’y lancer corps et âme, cela ne mérite-t-il pas une petite réflexion?

J’exagère? J’aimerais que ce soit vrai. Aux États-Unis, un avocat a présenté au tribunal un mémoire défendant son client en s’appuyant sur plusieurs cas de jurisprudence, avec citation d’extraits de jugement et force détails. Le seul problème: le mémoire avait été rédigé par ChatGPT et aucun des précédents cités n’existait en dehors de ses neurones artificiels. Il semble néanmoins que le brillant avocat ait soumis ce mémoire en toute bonne foi, convaincu de son bien-fondé.

L’affaire n’est pas isolée. En fait, depuis la sortie en grande pompe de ChatGPT, ce genre d’histoire se multiplie, dans tous les domaines.

OpenAI, clame en substance: “Nous vous vendons une machine qui pensera pour vous”. Conquis, les clients mettent leur cerveau sur off. Je n’ose pas imaginer l’odeur des appartements quand OpenWC vous vendra l’app qui prétend pisser pour vous…

New Harpejji Videos

Wed, 18 Dec 2024 16:52:38 +0100

After the gig I played with my harpejji in L’atelier des Carrels end November, Les Chemins de Traverse published a series of 9 videos with short excerpts of what I played there.

In this gig, I experimented with as diverse styles and atmosphere as I could, so the excerpts are quite different one from another (at least that’s how they sound to me), ranging from classical to jazz, to ethno and contemporary music…

Most harpejji videos on the net are more or less in a pop/rock style. I hope these videos contribute to show that the instrument is able to do many other things too!

Turing pris à revers

Tue, 26 Nov 2024 16:50:14 +0100

Après un premier texte sur la définition de l’IA et un second sur sa consommation énergétique, voici une petite réflexion sur les risques (ou non) que l’humain soit rattrapé par la machine.

Vous est-il jamais arrivé, lors d’un contact avec le service client d’une grosse entreprise ou d’une administration, de vous demander si la personne que vous avez au bout du chat, voire même au téléphone, est réellement un humain?

Les procédures qualité et la volonté que tout employé soit un simple rouage remplaçable dans la mécanique de l’entreprise transforment parfois les employés en machines.

Alan Turing, l’un des pères fondateurs de l’informatique, imagine en 1950 déjà un test pour déterminer si une machine est intelligente ou non.

En gros, l’idée est qu’un examinateur mène en parallèle deux conversations - par écrit - avec un humain d’une part et un ordinateur d’autre part. S’il est incapable de déterminer lequel de ses interlocuteurs est une machine, cette dernière passe le test et est donc déclarée intelligente.

Nous sommes probablement proche du moment où des ordinateurs seront capables de passer le test de Turing. Mais peut-être pas exactement pour les raisons imaginées par Turing lui-même.

Les machines ont fait des progrès époustouflants ces dernières années, c’est un fait. Mais les méthodes de management moderne, les économies dans le secteur de la formation et les logiciels d’aide à la rédaction de textes lisibles (risibles?) ont également pour effet de baisser le niveau de langage - et parfois de pensée - d’un certain nombre d’humains.

Ainsi lorsque l’on clamera, dans une année ou dans cinq ans, que pour la première fois une machine a passé le test de Turing, il sera bon de se demander: est-ce parce que nous avons réussi à hisser la machine au niveau de l’intelligence humaine, ou à broyer des humains pour en faire des machines?

Une question d'énergie...

Wed, 06 Nov 2024 12:51:58 +0100

Après un premier texte sur la définition de l’IA, voici une réflexion sur sa consommation énergétique.

En mai 2017, le logiciel AlphaGo bat le champion du monde du jeu de Go. Ce jeu a longtemps été réputé très difficile, voire impossible à automatiser du fait du gigantesque nombre de combinaisons possibles ainsi que de la subtilité de ses règles et des stratégies en découlant.

Alors, l’humain est dépassé, ou s’il ne l’est pas encore, c’est pour demain?

Pas si sûr!

D’abord parce que si l’on sait faire jouer une machine au go, ce n’est pas demain ni même après demain qu’on verra un robot gagner un match de badminton ou une équipe de machine gagner un match de foot contre des humains.

Mais il y a autre chose: les 1'200 processeurs et 175 coprocesseurs graphiques d’AlphaGo consommaient environ 1'000kW pour fonctionner. Le cerveau humain, lui, tourne à moins de 20W. Pour gagner, AlphaGo a donc grillé 50'000 fois plus d’énergie que le cerveau de son adversaire. Même en comptant l’énergie du corps entier - ce qui n’est pas une mauvaise idée, le cerveau tout seul ne pouvant pas faire grand chose - on reste à un facteur 10'000.

10'000 fois plus d’énergie pour une même tâche… peut-être qu’un jour les IA (quoi que cela signifie) prendront le pouvoir sur les humains et les asserviront. Il faudra à ce moment-là serrer les dents, baisser la tête et attendre. En quelques mois, les machines à court d’énergie s’éteindront d’elle-mêmes et l’humain, aussi naturellement bête qu’avant, pourra librement recommencer à développer de nouvelles technologies le mettant en danger.

Ce que l'intelligence n'est plus

Wed, 23 Oct 2024 15:21:48 +0200

Ce premier texte de la série amène quelques réflexions sur la définition de l’IA.

Intelligence artificielle…

Deux mots qui font couler beaucoup d’encre
qui créent des espoirs fous et des peurs insensées
qui promettent des lendemains qui chantent
ou un avenir maudit.

Plus un produit ne peut se vendre
sans qu’il ait été conçu par l’IA
ou qu’il fonctionne grâce à elle.

Les chaussettes, les voitures, le service après vente,
jusqu’aux sex toys, rien n’y échappe,
tout est intelligençartificiellisé.

Pourtant une chose me frappe:
Si tout le monde en parle
peu de monde semble se poser la question
de ce qu’est vraiment l’IA.

Marvin Minsky, l’un des pionniers de l’intelligence artificielle, en a proposé en 1956 la définition suivante:

La construction de programmes informatiques capables d’accomplir des tâches qui sont, pour l’instant, accomplies de façon plus satisfaisantes par des êtres humains.

Cette définition a l’avantage d’être très générale et correspond assez bien à l’intuition que l’on peut avoir de ce concept. Mais c’est pourtant une définition paradoxale ou, pourrait-on dire, une définition hara-kiri. En effet, chaque succès de l’IA sort automatiquement de cette définition: si une machine parvient à faire aussi bien qu’un humain pour une tâche donnée, cette tâche n’est plus “pour l’instant accomplie de façon plus satisfaisante par des êtres humains”, et donc ce n’est plus de l’IA.

De fait, cela va même plus loin: chaque fois qu’une machine arrive à remplir une tâche que l’on considérait comme nécessitant de l’intelligence, on contemple le résultat et on se dit: “ça ne peut quand même pas être ça, l’intelligence…”

En contemplant l’histoire de cette discipline durant les dernières décennies, on pourrait donc tenter la définition suivante: l’IA, c’est ce que l’intelligence n’est plus.

My PyConDE 24 talk is online!

Mon, 14 Oct 2024 11:26:51 +0200

It took some time, but the video of my PyConDE 24 talk is now available on youtube!

For some reason, embedding is disabled for this video, so please click on the image below to see the video on youtube:

You might also be interested in the companion post I’m mentioning in the talk (with slides and references).

I’m not very fond of the AI-generated description text they’re using on youtube, so let me quote below the description I wrote for the conference program:

Callbacks have become an ubiquitous programming technique that we use every day without even thinking about it. They are definitely handy in many situations, but sometimes they feel more like a burden than a help. In developing an interactive realtime audio processing system for use on stage in live music, we encountered such a situation. This talk will present how a few dozen lines adding a thin abstraction layer allowed us to replace a complex callback mess with tremendously more readable generators (yes, you know, those functions which yield results instead of returning them…).

At Les Chemins de Traverse we explore ways of “augmenting” acoustical musical instruments with new sonic possibilities offered by computers (think “augmented reality” for live music). For doing so, we are using Olivier Bélanger’s great pyo module for realtime audio processing. To make the system interactive, this module allows to register callbacks on some events. While this works great in many situation, it can get very cumbersome when we design a stateful system, where the same event must trigger different callbacks depending on the system’s inner state.

This talk will present how we developed a thin abstraction layer that allows us to replace many callback functions together with many registering/unregistering of these functions by a nice, streamlined generator definition that’s incomparably more readable than the many-callbacks version. This allows us to keep our mind focused on what’s important, namely supporting the music we want to play, instead of tedious boilerplate code.

While our use case is admittedly very specific, we believe that the ideas we present could be adapted in many other situations where callbacks are used for technical reasons, but lead to bulky and contrived code.

Voice Harpejji Improvisation

Fri, 13 Sep 2024 09:10:12 +0200

Dimanche dernier, Les Chemins de Traverse organisaient un plateau de lecture.

Ahmad Motalaei y est venu avec un poème de Rûmî qu’il s’est proposé de chanter plutôt que de simplement le lire à haute voix. Je lui ai proposé un accompagnement à l’harpejji et nous sommes partis dans une impro sans aucune préparation.

Les Chemins de Traverse viennent de publier la vidéo de ce moment.

L’inspiration de la musique persane m’a poussé à chercher de nombreuses sonorités différentes, en plus du tapping “normal”: des effets de sourdine, gratter les cordes, du slap, divers effets percussifs, …

J’ai réussi à obtenir quelques sonorités qui me plaisent bien et je me réjouis de creuser pour développer ce type de jeu “non standard” (si tant est qu’il y ait un jeu standard sur l’harpejji!)

J'aurais voulu vous lire...

Mon, 09 Sep 2024 13:34:11 +0200

Hier Les Chemins de Traverse organisaient un plateau de lecture intitulé Art de la science, science de l’art. J’y ai lu un texte que j’avais écrit pour l’occasion et que je reproduis ci-dessous.

J’aurais voulu vous lire aujourd’hui un extrait d’un ouvrage majeur du début du XX^ème siècle.

J’entends d’ici vos esprits fureter: Proust? Appollinaire? Joyce peut-être? Ou, pour rester dans la thématique du jour, Einstein, Freud? Marie Curie?

Ne cherchez plus, je sais que vous ne l’avez pas lu. Comment puis-je en être aussi sûr? S’agirait-il d’un livre disparu? ou jamais publié? Pas du tout, il est paru chez Cambridge University Press en 1910. S’agirait-il d’un essai écrit dans une langue exotique et jamais traduit? À moins que l’anglais n’entre dans cette catégorie, ce n’est pas vraiment ça non plus. Ou d’un ouvrage tellement dangereux qu’il ne doit pas être mis entre toutes les mains, comme le second volume de la poétique d’Aristote dans Le nom de la Rose? Pas du tout, bien que la référence à Aristote soit assez pertinente en l’occurrence.

Non, vous ne l’avez pas lu parce que ce livre est illisible.

J’aurais voulu vous lire aujourd’hui un extrait des Principia Mathematica de Whitehead et Russel, mais je me suis rendu compte que je n’ai pas la moindre idée de comment prononcer son contenu. Les cinq années de ma vie que j’ai passé à faire des maths à plein temps n’y changent pas grand chose.

Prenons un extrait au hasard du côté du début de cet ouvrage monumental de près de 2'000 pages.

S’il est facile de prononcer les numéros de proposition et le mot abrégé “Dem”, vous conviendrez avec moi que toute tentative de faire une lecture à haute voix de cet extrait ressemblerait au mieux à de la poésie sonore abstraite, mais plus probablement à un informe galimatias.

Si vous avez l’impression de voir une suite de symbole sans signification c’est que vous avez bien compris. En effet en un certain sens c’est exactement le but de l’exercice.

Mais pourquoi, me direz-vous, passer dix ans de sa vie à aligner sur des centaines de pages des symboles sans signification? Pour bien comprendre l’intention de Whitehead et Russel, il nous faut remonter un peu le temps.

On attribue habituellement à Euclide l’invention de la méthode axiomatique, probablement autour de 300 avant JC. On pose des définitions et des axiomes, puis toute la théorie se déploie depuis ces quelques bases sans ajouter de nouvelle notion. Mais les mathématiques restent avant tout une question de bon sens. “Une ligne est une longueur sans largeur”, nous dit par exemple Euclide, définissant ainsi une des base de sa théorie à l’aide de notions intuitives elles-mêmes non définies au sein de la théorie.

Au XII^ème siècle, Le mathématicien indien Bhaskara propose la preuve suivante du théorème de Pythagore: un dessin, accompagné du simple mot “Voyez”. Là aussi, les mathématiques sont plus une prolongation du bon sens qu’une méthode purement formelle.

Mais au fil des siècles les mathématiciens se mettent à manipuler des nombres imaginaires, à inventer des géométries courbes, à se promener dans des espaces à n dimensions (n pouvant même être infini). En 1832, un certain Évariste Galois se fait tuer en duel, victime selon ses mots d’une infâme coquette. Le soir d’avant, se sachant pour ainsi dire condamné, il avait couché fiévreusement sur le papier les idées qui l’habitaient, posant ainsi les bases de la théorie des groupes et ouvrant le chemin à l’algèbre moderne.

Bref, à l’aube du XX^ème siècle, les mathématiques ne sont plus une question d’intuition et de bon sens. Les notions manipulées deviennent si abstraites que l’intuition, parfois, devient un obstacle et le bon sens un frein. Et quand la théorie de la relativité et la physique quantique viennent ajouter leur grain de sel (ou peut-être plutôt déverser des sacs entiers de sel), le moins qu’on puisse dire est que cela n’arrange rien.

Dans ce contexte, des mathématiciens comme Peano, Frege ou justement Whitehead et Russel se lancent dans une aventure inouïe: vider les mathématiques de toute intuition et mettre en place une mécanique purement formelle permettant, par une simple manipulation de symboles selon quelques règles de départ, de dériver un à un tous les résultats mathématiques imaginables.

Commence alors cette odyssée insensée, au sens propre, d’aligner des symboles vides de sens les uns après les autres de manière purement mécanique, sur des centaines et des centaines de pages.

Cette entreprise titanesque a été à la fois un ratage total et un succès incroyable. Ratage, parce qu’en 1931, Kurt Gödel mettra en évidence une limite fondamentale de cette approche. Tout système formel de ce type est forcément incomplet, en ce qu’il contient des propositions dont on ne peut prouver ni la véracité ni la fausseté à l’intérieur du système. L’axiomatisation du monde, rêvée par Whitehead et Russel, est donc intrinsèquement vouée à l’échec.

Et pourtant, le travail de fourmi de ces deux fous furieux a planté une graine qui aller germer quelques décennies plus tard, et, bon grain ou ivraie, se répandre d’une manière incroyable à travers le monde entier. Par leur travaux, Whitehead et Russel ont rendu imaginable l’idée qu’une manipulation purement formelle de symboles pouvait ramener au niveau de l’automatisme une partie du raisonnement humain. Les contraintes techniques de l’époque ont fait qu’ils ont écrit les Principia mathematica à la main, mais je suis persuadé que s’ils avaient disposé d’une “machine à manipuler des symboles”, ils l’auraient utilisée.

Et je suis également persuadé que si, trois décennies plus tard, des von Neumann ou des Turing ont pu poser les bases d’une machine à raisonner sur la base d’opération logiques élémentaires, c’est aussi grâce aux portes ouvertes par les Principia. Cette machine, vous l’utilisez tous les jours, vous l’avez même dans votre poche - et je ne parle bien évidemment pas de votre mouchoir.

J’aurais voulu vous lire aujourd’hui un extrait des Principia Mathematica de Whitehead et Russel, et je me suis rendu compte que je n’ai pas la moindre idée de comment prononcer son contenu. Je ne résiste cependant pas, en guise de conclusion, à vous montrer la célèbre proposition 54.43 - ce n’est pas le 49.3, mais vous verrez, c’est encore mieux! - la proposition 54.43, disais-je, qui apparaît à la page 379 du premier volume de cet ouvrage.

Ce qui a rendu cette proposition célèbre, ce n’est pas tellement sa formulation des plus poétiques, ni même la limpide démonstration qui la suit. Non, ce qui la distingue des centaines d’autres suites de symboles qui l’entourent, c’est la petite remarque ajoutée juste après: “De cette proposition découlera, quand nous aurons défini l’addition arithmétique, que 1+1=2”.

Source des illustrations: University of Michigan Library pour les Principia et S’il vous plaît, dessine-moi une preuve pour Pythagore par Bhaskara.

Harpejji Baroque Improvisation

Mon, 05 Aug 2024 11:12:28 +0200

I’ve been practicing classical improvisation on the harpejji for some months now, and while experimenting I came up with a prelude in C major that I thought was quite nice.

So I practiced it a little bit before recording it. Here is what it sounds like:

I’d like to dedicate this piece to the late Gaël Liardon, who was the first to make me realize that this kind of improvisation was even possible.

The 'Harpejji Solo' project

Fri, 19 Jul 2024 17:05:52 +0200

For various reasons this site has been mostly dormant recently, but that doesn’t mean that I’ve been doing nothing!

Among others, after having played mostly accompaniment on the harpejji for years, I decided to explore more what can be done with the instrument in a solo setting. (This exploration has been partly supported by a grant offered by the City of Neuchâtel, many thanks for that!)

I did not document my research on this website, but I did write about it on the website of Les Chemins de Traverse.

Here are two videos that were published there:

first a rendition of Olhos de Gato, a nice standard written by Carla Bley.

Your browser does not support the video tag.

then a quick improvisation in classical style, following what’s called the Rule of the octave.

Your browser does not support the video tag.

If you want more context or discover other resources related to harpejji solo playing, like suggestions for sight-reading or reviews on books about improvisation, head to the Harpejji solo page.

matthieuamiguet.ch v3.0!

Thu, 27 Jun 2024 17:04:52 +0200

J’ai créé ce site en 2009 à l’aide de MODx, puis je l’ai migré à un CMS fait sur mesure en Django en 2011.

13 ans plus tard, il faut bien admettre que cette deuxième version montrait des signes de fatigue, aussi bien au niveau du design que des technologies (d’autant plus que je n’avais pas vraiment mis le code à jour depuis… bien trop longtemps!).

Intrigué par le travail de Nicolas Friedli avec Hugo et toujours amoureux, sur le principe du moins, des sites statiques j’ai décidé de tenter le coup et de reconstruire le site avec Hugo, selon une approche assez minimaliste.

Le résultat est sous vos yeux, j’espère que vous l’appréciez… et n’hésitez pas à me signaler si vous rencontrez des problèmes!

PyConDE 2024 - Some context about my talk

Wed, 24 Apr 2024 12:53:58 +0000

This is a companion post to my PyConDE 2024 talk Replacing Callbacks with Generators: A Case Study in Computer-Assisted Live Music to provide some context and additional information. In no particular order:

The slides used in the talk are here.
The full code for the Scenario class is here, as part of the much broader Charming Snake project. Examples of use can be seen in the code for Passamezzo and Helvetic Trip.
pyo is a great module for realtime audio processing developed by Olivier Bélanger in Montréal.
For more details on why we chose to work with python/pyo for realtime audio processing, see my Europython'19 talk and/or my Swiss Python Summit 2016 talk.
pyo is great for low-latency audio processing, although getting reliable sub-10-ms roundtrip latency may require “some” tuning of the OS. My RMLL'17 and RMLL'12 talks give some details about the required settings on a Linux system.
The auto-wah effect in pyo quickly demonstrated in the talk is described, along with other effects, in my blog post Create your own real-time guitar effects with python. This post can also be considered as a very quick pyo 101 for this kind of use (the scope of pyo being much broader than that!).
The solo bass flute video is from the project Old Wine in New Bottles
The audio excerpt with augmented contrabass and bass flutes is from the Dragonfly album.
The video with the dancer and harpejji is from the Digital Analogies project.
For a comparison of several languages for realtime audio processing, including ChucK and python/pyo see my blog post on the subject.
As far as I know, using generators to simplify callback-based code is not very well documented on the web. One notable exception are a couple of talks by the excellent David Beazley like this one.

I hope you enjoyed the talk! Don’t hesitate to contact me if you still have questions!

Using a raspberry pi as a DIY fx pedal

Wed, 23 Aug 2023 14:31:28 +0000

Not quite ten years ago (!), I wrote about creating your own real-time guitar effects with python. Since then, I’ve been using pyo more and more in my gigs for low latency audio processing. But until now, I’ve always been using it on a laptop, with a bulky sound interface, foot controllers, a reverb rack, etc.

All those things make sense for a project like Dragonfly where the quantity of stuff to take with me for a gig is huge anyway, but I’ve been dreaming more and more to have a smaller, much portable version, e.g. for occasions where I only take my harpejji with me and would like to travel by train.

So I looked at this raspberry pi that was sleeping on a shelf and thought that maybe…

Many hours (days!) later, I can confirm that it is absolutely possible to turn a rpi into a low latency audio processing device. It’s even relatively easy… if you know how to do it!

This post will describe succinctly the main steps necessary for getting there. It’s more a reminder for the later myself than a tutorial… but if you’re familiar with the Linux audio world, you should be able to follow these steps yourself.

Warning Although this post solves all computer-related problems, I’m still having problems with electrical noises in the output. See the update at the end of this article for more details before you spend time trying to replicate this process!

Material

I’ve been using

A rpi 3b in a case
A ESI U24 XL USB sound interface
A Korg nanoKontrol2 for the UI

Kernel and distrib

I’ve tried, measured, tweaked, measured again… but I couldn’t get under 10ms roundtrip latency with a mainline kernel. I won’t play on a device that has a higher latency than this threshold, so I had to switch to an rt-kernel.

After trying different things, I found that the easiest way to do that was switching from Raspberry Pi Os to Debian, because the latter provides an rt-kernel in its repos.

Basic install

(https://raspberrytips.com/install-debian-on-raspberry-pi/ saved me some time there)

Download and flash the image
Boot the rpi and - please! - begin with setting a root password (passwd)
Edit /etc/network/interfaces.d/wlan0 for getting online
update your system (apt update and apt upgrade)
create a normal user (adduser <name>)
apt install sudo
add your user to groups audio and sudo

Post-install tweaks

install a realtime kernel (apt install linux-image-rt-arm64) and reboot (test with uname -a)
install python3, jackd2, linux-cpupower, alsa-utils, python3-pyo
setup passwordless sudo (sudo visudo and add the line user_name ALL=(ALL) NOPASSWD:ALL)

Some soundcards have volumes turned to 0 on boot. Launch alsamixer, set the volumes to your liking, quit and run sudo alsactl store. We will handle restore later.

You should now be able to set your cpu governor to performance (sudo cpupower frequency-set -g performance) and launch a low-latency jackd without xruns (jackd -S -P70 -dalsa -dhw:U24XL -r48000 -p64 -n3 -D, which results in a very honorable 4ms ‘jack’ latency).

If you encounter xruns, rtcqs should help you nailing down what’s going wrong.

If you feel so inclined, you can now measure the roundtrip latency with jack_iodelay. In my case I got ~6.7ms, which is very comfortably under my 10ms requirement!

The main script

Write the pyo script you need for your effects (take inspiration there if required).

I’m using pyo-nk2 to interact with the nanoKONTROL2. In particular, I added

quitter = pyo.TrigFunc(nk2.Press('stop'), quit)

to my script, with quit() being the function that terminates the server and stops the script.

I also like to have

def cycle_blink():
    nk2.led_blink('cycle')

alive = pyo.Pattern(cycle_blink).play()

Now I can verify that my script is running by checking the “cycle” led on my nanoKONTROL2 :-)

Launching the script at boot time

After some research I decided to create a custom systemd unit.

I want to be able to connect to the rpi and see what’s happening in case of a problem; that’s why I decided to launch jackd and my script in a byobu session with tmux backend (don’t forget to apt install byobu!). Using the technique described in this answer that yields:

#!/usr/bin/bash

sudo alsactl restore
sudo cpupower frequency-set -g performance

# start new detached byobu session
byobu new-session -d -s harpejji
byobu split-window -h

byobu select-pane -t 0
byobu send-keys "JACK_NO_AUDIO_RESERVATION=1 jackd -S -P70 -dalsa -dhw:U24XL -r48000 -p64 -n3 -D" ENTER

sleep 1

byobu select-pane -t 1
byobu send-keys "python3 /home/amiguet/harpejji/harpejji.py; sudo poweroff" Enter

I must admit this is not very elegant. I’m still thinking about ways to get the same result in a nicer way.

Anyway, we can now create a systemd unit:

# copy to /etc/systemd/system/harpejji.service
[Unit]
Description=Harpejji

[Service]
# see https://bbs.archlinux.org/viewtopic.php?pid=1306587#p1306587
LimitRTPRIO=infinity
LimitMEMLOCK=infinity

# Type is forking because of byobu
Type=forking
User=amiguet
Group=audio
ExecStart=/home/amiguet/harpejji/harpejji

[Install]
WantedBy=multi-user.target

# networking startup takes too long, let's start this one before
WantedBy=networking.service

Enable the service (sudo systemctl enable harpejji.service) and reboot. All going well, the “cycle” led should start to blink after a while. You can then use your fx box. When you’re finished, just press “stop” on the nanoKONTROL2 and the rpi will shut down.

Limitations

When ssh’ing into my rpi, there seem to be some interference between network and audio that causes noise (not xruns!) in the output of my sound interface. I will probably completely disable networking when I’m sure everything works well (that will also allow for a faster boot), but it’s handy to keep it as long as I’m tweaking things. Anyway, the noise disappears as soon as I disconnect from the machine.

Conclusion

This is a puzzle with many small pieces and getting them all together was not an easy task, but there’s nothing really difficult in there. And transforming a $40 single board computer into a low latency audio processing unit is quite amazing!

Your browser does not support the video tag.

Update 28.08.2023

I thought I was there, but I’ve been having electrical noise problems that I could solve yet. After some analysis, and with the kind help of the excellent Dario Ciani, I could identify two different sources of noise:

A loud low hum, very typical of a ground loop… except that that is impossible as the pi is not grounded. As it happens, it’s not a ground loop, but a hum created by a floating ground. The hum disappears if I plug a monitor through hdmi into the pi, as the pi is then grounded through this connection. It also disappears if I route the signal through a mixer instead of plugging my sound interface directly into the amp.
A high hiss, provoked by the electrical noise of the Switched mode power supply of the pi.

If I replace my U24XL with my Presonus audio box 1818VSL, which is powered from the wall socket, it solves both problems. But it defeats the initial project of having something very portable.

(And of course I also tested the sound interface from a laptop to be sure it is capable of delivering a good sound!)

I’ll have to think about it and find how I can solve these problems and have my raspberry pi deliver a usable sound… I will describe the solutionin a new post on this blog soon (hopefully!).