Thank you both for the perspective on the radio function. If the radio function is built on AI/ML, and the size of the dataset in large part define the quality, does that mean the individual platform's radio should theoretically be better than Roon's since they would have a lot more user? In this intended implementation on the Extreme, would that mean the radio function can access only one platform at any given time, so no cross pollination between Tidal and Qobuz for example?
Taiko Audio SGM Extreme : the Crème de la Crème
- Thread starter CKKeung
- Start date
Thank you both for the perspective on the radio function. If the radio function is built on AI/ML, and the size of the dataset in large part define the quality, does that mean the individual platform's radio should theoretically be better than Roon's since they would have a lot more user? In this intended implementation on the Extreme, would that mean the radio function can access only one platform at any given time, so no cross pollination between Tidal and Qobuz for example?
Any Radio-function will be built using AI/ML.
AI/ML is dependent on two things (simplified view):
1) Data to enable it/train it - here both quantity of data (roughly correlating with # Users on a platform e.g. Qobuz) but also the quality of the data plays a big factor - people using Radio only to generate background music vs People like myself using Radio to actively discover new music I really like. When you use Roon Radio you might have used the little feedback widgets, e.g. skipping a tune you did not like and answering the question from Roon Radio whether you did not like the music or whether you just did not want to play it now. Radio will also pay attention to whether you played the entire song or if you skipped it half-way. All these items are training data that will be collected locally on your Roon Core and fed back to the central Roon 'brain' and use to improve the Roon Radio AI/ML.
2) The AI/ML algorithm itself - there are tons of different ML algorithms, each such algorithm comes with a large set of parameters (hyper parameters) - a combination of algorithm and actual parameters is the 'secret sauce', important to understand is that this is a constant moving target as new data is continuously collected and used to improve the AI/ML.
I used to use Tidal (only until Quobuz became available in Sweden) - Tidal's Radio was not very good at helping me Discover new music.
Qobuz Radio is a little bit better.
But - Roon radio is very, very good. So - it is not about which platform has largest # Users but also the quality of the data and of course the quality of the AI/ML algorithm itself.
I know that mass market platforms like Spotify have invested massively in their Radio - to them it is very clear that their users do not want to chose music but be fed/served music by an algorithm.
I suspect that for audiophile offerings like Qobuz and Tidal there is more focus on the content and users selecting music themselves -> Qobuz and Tidal have not (yet) invested so massively in AI/ML.
By integrating with the underlying platform's native Radio XDMS will leverage the platform's Radio API to retreive a suitable next song based on what have been played before. The underlying platform will still be able to pick up some useful data via XDMS' invocations of the platform's API (e.g. listening to the whole song or skipping parts of it). Each song played via an XDMS Playlist will cause XDMS to interact with one platform API -> no cross pollination.
Since # Users running XDMS will be << compared with # Users on the platform I doubt that the platform will care - they make money on the subscription which the XDMS User still need to have.
Each platform will decide whether they want to expose an open API for the outside world (like XDMS) to use or keep a walled garden allowing only their own apps to access the platform - this is a business strategy decision.
Last edited:
I understand that for AI/ML to occur optimally a large population of users whose actions can drive learning is necessary. However I hope in XDMS there is room for my personal feedback for Radio-type feeds since at the end of the day I am not so much interested in what the majority of users out in the world like/dislike or find congruent or not for Radio content nearly as much as I am interested in my Radio function learning *my* likes/dislikes, musical taste and favorite music.
Steve Z
Happy new year everyone and hope all are well.
Had a blissful 2 hours to listen just now and felt compelled to post as I step back in appreciation of the TAS software.
While I am so excited for XDMS, of course, I have to admit I already have nostalgia for TAS. What was a beta software that was pulled off so relatively quickly (but no doubt w hundreds of sweat hours through the night) to demonstrate, as Mike L uses the term, 'bleeding edge' SQ improvements from in house software, continues to sound just jaw dropping. I have not used Roon either in 18 months (I forget if TAS came out before or after Covid ramped up).
Sure, I did love Roon's interface/search but I don't miss it as I gather new music recs from friends and still can access the latest Qobuz playlists for new releases etc. And, in my system, I have never experienced a glitch with TAS. Any issue I ever had was solved by switching back and forth between batch players. In my limited software knowledge, that is amazing, especially in a beta version.
For what TAS was meant to do, it sounds like it has uber over delivered from programming/testing genius, the feedback from this group and the SQ it served in the interim. Such an extraordinary accomplishment, that I think will always be remembered by those that experienced it and I imagine pointed to down the road as a successful approach to harness engineering talents while harmonizing with end users real time feedback. Also a very transparent effort w decently high stakes for all of the hifi community to see that can be equally complementary as critical.
Anyway, while I could live with TAS for good, I greatly look forward to settling into XDMS and peeling back the next layer.
Take care everyone. Congrats again to team Taiko for moving so many balls forward at once and the skills and foresight to create something like TAS. Toasting 2022 w hopeful optimism on all fronts.
Had a blissful 2 hours to listen just now and felt compelled to post as I step back in appreciation of the TAS software.
While I am so excited for XDMS, of course, I have to admit I already have nostalgia for TAS. What was a beta software that was pulled off so relatively quickly (but no doubt w hundreds of sweat hours through the night) to demonstrate, as Mike L uses the term, 'bleeding edge' SQ improvements from in house software, continues to sound just jaw dropping. I have not used Roon either in 18 months (I forget if TAS came out before or after Covid ramped up).
Sure, I did love Roon's interface/search but I don't miss it as I gather new music recs from friends and still can access the latest Qobuz playlists for new releases etc. And, in my system, I have never experienced a glitch with TAS. Any issue I ever had was solved by switching back and forth between batch players. In my limited software knowledge, that is amazing, especially in a beta version.
For what TAS was meant to do, it sounds like it has uber over delivered from programming/testing genius, the feedback from this group and the SQ it served in the interim. Such an extraordinary accomplishment, that I think will always be remembered by those that experienced it and I imagine pointed to down the road as a successful approach to harness engineering talents while harmonizing with end users real time feedback. Also a very transparent effort w decently high stakes for all of the hifi community to see that can be equally complementary as critical.
Anyway, while I could live with TAS for good, I greatly look forward to settling into XDMS and peeling back the next layer.
Take care everyone. Congrats again to team Taiko for moving so many balls forward at once and the skills and foresight to create something like TAS. Toasting 2022 w hopeful optimism on all fronts.
Last edited:
Steve, your personal, manual song selections, played via platform's API invoked by XDMS (or by the platform's native player) will be added to the platform's data trove and used to improve the platform's AI/ML over time. This will impact future Radio suggestions made to yourself but also to other users that the platform's AI/ML believes have similar taste to yourself.
Your personal choices for a suggested song from platform's Radio: skipping it | playing a full-song | playing a partial partial song will be fed back to the platform's Radio via the platform's Radio API invoked by XDMS.
The power of the AI/ML lies in the rather uncanny ability to leverage not only your own, personal data but finding other users that are similar to yourself and also leverage that data - i.e. leverage the wisdom of the crowd but intelligently selecting who is part of 'your crowd'.
To me at least - Roon Radio does this very, very well.
Final point - AI/ML is a completely different and separate domain compared with 'traditional' software engineering - ways of working and skills are VERY different from those of a software developer. It is the most in-demand skill set and commands ridiculous salaries, Spotify's top AI/ML resources are on CEO-level salaries.
Last edited:
Thank you for this overview. The Taiko software Team know when we are out of our depth, AI / ML is certainly a field we leave to others with the support of deep pockets
That being said, we are looking at ways which the recommendations of other subscription streaming services can seed fetching of tracks on our favored / supported streaming services
I have been working in this field of expertise for a few years (running data science related processes on the large scale) and this somehow is a personal field of interest, I agree with your comment 100%.
Implement such AI driven process would require a significant investment, finding skilled engineer/scientist in the field it's something every data companies struggle everyday (including Spotify).
My take on this is that beside the volume of data that is required in order to build a good recommender system, the most dominant factor would be the "quality" of these training datasets, and their dimensionality, at least which important feature of data exercise a significant weight in the recommendation outcome and how that is finely tun-able. That IMHO what makes the difference from Roon over more simplistic form of recommendation that other streaming services offer, perhaps there they don't even run ML process at all.
Taking things further, Spotify has now the best in class state of the art approach, but that involved the acquisition of a company (EchoNest) that provided music intelligence/audio analysis (A bit unfortunate I must say, as this data now becomes proprietary):
One of the reasons I have stopped listening from Spotify in the past (way before my interest on HIFI and EchoNest acquisition) was their recommendation system was overly repetitive, same famous songs over and over.
As consequence of the data that was acquired (audio features and analysis of the audio spectrum), the recommendations are no longer limited by the classic approach (users that have listened this song have listened to that song), but potentially recommendations can be generated over similarities of given audio features ( Energy, Tempo, Valence, etc, I believe these are also available/taken into account to Roon as part of their underlying API) and potentially features extracted from the audio analysis spectrum.
So result I'd expect there is some sort of hybrid approach which is finely tun-able, and it is somehow reflected in their API documentation.
Spotify for Developers
developer.spotify.com
Ideally this API can be used to mix-match songs recommendation for other platforms but I would be surprised their T&Cs would allow that, particularly if Spotify it's not the main/sole mean of audios streaming.
30 years ago, the state of the art for assessing voice quality for wireless communications was to make audio recordings under various conditions (noise background, voice encoder algorithm, etc) and have a panel of "expert listeners" give a numerical rank on a scale where 5 was the noise-free, full bit rate source material. the resulting average scores were call MOS (mean opinion scores) and vocoder algorithms (the voice analogs of mp3) were chosen for use in cellular based on their MOS score.
about 25 years ago, we developed and patented a neural network (using matlab) that trained on the audio recordings and the MOS scores given by the 'expert listeners.' it was able to accurately predict the MOS of later and more advanced voice codecs used in 2G and 3G mobile communications, without the needed for time consuming (and expensive!) 'expert listeners.' of course, even better methods have been developed over these past decades.
so let's dream (hallucinate?):
1. we already have the best panel of 'expert listeners' you could hope for right here
2. imagine if we had the ability to make recordings (or listening sessions or measurements) of some sort that people could 'score', on one (totality) or more (soundstage, speed, timble, etc) scales of interest
3. train the neural network based on these scores and the 'recordings'/'listenings'/'measurements'
3. then, you could vary the parameters (switch, router, equipment footers) and use the NN to assess the impact
of course, we already have this as Emile et al. serve as our highly calibrated 'listerners' who are already implementing this, and whose judgement of the MOS scores have been verified by the extremely obsessive members of this forum.
just some random thoughts to keep discussion moving while we sit here quietly listening to music, in a state of anticipation trying to imagine how the sound could be better with XMDS.
stay positive
test negative
This is a really cool idea
One application could be for a mastering engineer test out his Masters
We have a lot of new technology, updates and upgrades releasing this and next year, on the hardware side it is mainly holding for sufficient parts availability, on the software side it's mainly squashing the last few resilient bugs. As we've been copied a bit to often now we've chosen to not disclose what we are releasing till we are there as some of this is currently non existing technology. I can disclose we are very excited over what we are sitting on here, most of these are not minor improvements/optimizations in the ~10% sound quality delta range but, al always IMHO, rather substantial advances, perhaps even revolutionary.
Thank you, all very compelling points indeed. During my audition, I had not the time to pinpoint the single most relevant improvement on my system, probably because of a combination of a too brief exposure and the extent of holistic step-up compared to my current server being too big.
I experienced an overall release of tension in the music, a more visceral depiction of the midrange, and - more apparently - an increase of nuances / information in the bass region.
As for the *fast* thing, I read a very interesting interview from Rob Watts (the Chord DAVE mastermind), where he insists that the accuracy of timing is crucial not only for the obvious effects on rhytmic drive, propulsive energy, attack / decay realism, imaging, etc., but he mentions psychoacoustic experiments where the pitch, hence the timbre, of instruments was perceived very differently when the timing accuracy of the reproduced transients was altered, even very slightly.
I seem to recognize a similar design intent on the Taiko Extreme, in the obsessive research of latency reduction (?).
As I listen to 90% acoustic music (mostly classical) and often attend to live concerts, my quest is for realism, as I found that sound presentations that are steered towards the fun-side of music listening, and those veering towards an embellishment of the recording may be rewarding on the short term, but become annoying (for me) in the longer run.
Hope the Extreme will bring me one step closer to my quest
With this setup how are you adding new music to the internal drives of the Extreme? Are you using a laptop which you can temporarily connect to the Amplifi wifi network so you can see the Extreme?
I think taiko has the superior server product for many reasons
but two make it above all for me
1- a closed system that is there own and keeping users out of it with limited adjustments is key.
2- to develop all new methods of better sound and not just using off the shelf ideals Roon being one of them
i use Roon it’s a great GUI but it’s sound. Alone is subpar compared to many other players. why they don’t go after great sound perplexes me. but it seems there objective is looks and features alone
but two make it above all for me
1- a closed system that is there own and keeping users out of it with limited adjustments is key.
2- to develop all new methods of better sound and not just using off the shelf ideals Roon being one of them
i use Roon it’s a great GUI but it’s sound. Alone is subpar compared to many other players. why they don’t go after great sound perplexes me. but it seems there objective is looks and features alone
It makes no difference if there are only 2 these main reasons. Many companies did it for a long time as Aurender, Lumin, Dcs...
Yes, this is hardly ideal, it only serves as an example on how to get more consistency. We are working on a better solution for this right now.
steve williams
Site Founder, Site Owner, Administrator
you can bet money on it. I have yet to hear a disgruntled Extreme user....plus the best is yet to comeHope the Extreme will bring me one step closer to my quest
A bit cumbersome but doable. I guess one would need to also make a change so someone could remote into the machine, if needed.
