I had the chance to listen to some good gear in Japan, and there was a combo that was interesting: Topping D900/A900 stack + HE1000se by HIFIMAN. I listened to "Come Away With Me" by Norah Johns and the beginning of the song (3 first seconds), seemed like it came from 5PM, 1 meter away, to 1PM. Pretty cool, I also tried that with STAX SR-X9000 (and its amp) and couldn't hear it in that way.
I have now the HE1000se at home and silly me, I couldn't reproduce that spatial sound (it is there, but not that crazy). So I though that it probably was the stack then. I wanted though, try to understand why I couldn't reproduce that experience with my gear (time domain and all that jazz).
I started asking Gemini (3 pro), ChatGPT (5.2), Claude (Sonnet 4.5) and getting in a rabbit hole. I ask it to be technical and drop the jargon. Then I started asking it to start showing me the math. And it did.
Long story short, my manual experimentation started with the following components and some reasonable number combinations:
* PCM / DSD local files
* RoonLabs
* HQPlayer
* NVIDIA RTX 5000 Ada (GPU offloading)
* Denafrips Ares II
* Singxer SA-1 Class A
* McIntosh HA200
* Dan Clark Stealth
* HIFIMan HE1000se
I also explored on paper Holo Audio Cyan 2 DAC.
As I asked for numbers, it walked me through R-2R ladder temporal resolution, resistor tolerance, bit depth, etc.
I ended up with 3 setting for HQPlayer and my stack. The following comes from the "AI":
-------------------
Jazz/Acoustic
Nx Filter: poly-sinc-gauss-xla
Dither: LNS15
Bits: 24
Volume: -1.5 dB
Output: Adaptive (705.6/768 kHz)
Rock/Metal
Nx Filter: poly-sinc-ext2
Dither: TPDF
Bits: 24
Volume: -3.0 dB
Output: Adaptive (705.6/768 kHz)
Electronic
Nx Filter: poly-sinc-short-mp
Dither: RPDF
Bits: 24
Volume: -3.0 dB
Output: Adaptive (705.6/768 kHz)
DAC
Ares II: NOS mode
Connection: XLR (Ares II → SA-1)
SA-1 Gain: LOW
--------------------
Some formulas used by AI along the way:
--------------------
τ (time constant) = R_total × C_parasitic
Settling time (0.01% accuracy) = 9.2 × τ
For 16-bit ladder:
Switch impedance: 16 stages × 15Ω + 10Ω (PCB vias) = 250Ω
Parasitic capacitance: ~75pF
τ = 250Ω × 75pF = 18.75 ns
Settling time = 9.2 × 18.75ns = 176 ns
For 24-bit ladder (my current hardware):
Switch impedance: 24 stages × 15Ω + 15Ω (PCB vias) = 375Ω
Parasitic capacitance: ~75pF
τ = 375Ω × 75pF = 28.1 ns
Settling time = 9.2 × 28.1ns = 258 ns
Resistor Tolerance vs Bit Depth:
Formula: Tolerance ≈ 1 / (2^N × 2)
20-bit: 1 / (1,048,576 × 2) = 0.48 ppm
24-bit: 1 / (16,777,216 × 2) = 0.03 ppm
--------------------
Well, the idea was not to dump the entire chats, but sharing how LLMs can guide you though the process, and you can ask it to show the math, that forces the AI not to use jargons, but trying hard to talk to an engineer.
This may be off if you are an electrical engineer (I'm a software one), but the interesting part was to start comparing results between LLMs and start getting to a consensus base the actual capabilities of the hardware.
How does it sound? Oh yeah great, but I started with the spatial (jazz) filter, and ended up creating 2 more as other genres / material didn't sound that impressive.
I guess one day we will have an adaptive DSP based on LLMs.
I have now the HE1000se at home and silly me, I couldn't reproduce that spatial sound (it is there, but not that crazy). So I though that it probably was the stack then. I wanted though, try to understand why I couldn't reproduce that experience with my gear (time domain and all that jazz).
I started asking Gemini (3 pro), ChatGPT (5.2), Claude (Sonnet 4.5) and getting in a rabbit hole. I ask it to be technical and drop the jargon. Then I started asking it to start showing me the math. And it did.
Long story short, my manual experimentation started with the following components and some reasonable number combinations:
* PCM / DSD local files
* RoonLabs
* HQPlayer
* NVIDIA RTX 5000 Ada (GPU offloading)
* Denafrips Ares II
* Singxer SA-1 Class A
* McIntosh HA200
* Dan Clark Stealth
* HIFIMan HE1000se
I also explored on paper Holo Audio Cyan 2 DAC.
As I asked for numbers, it walked me through R-2R ladder temporal resolution, resistor tolerance, bit depth, etc.
I ended up with 3 setting for HQPlayer and my stack. The following comes from the "AI":
-------------------
Jazz/Acoustic
Nx Filter: poly-sinc-gauss-xla
Dither: LNS15
Bits: 24
Volume: -1.5 dB
Output: Adaptive (705.6/768 kHz)
Rock/Metal
Nx Filter: poly-sinc-ext2
Dither: TPDF
Bits: 24
Volume: -3.0 dB
Output: Adaptive (705.6/768 kHz)
Electronic
Nx Filter: poly-sinc-short-mp
Dither: RPDF
Bits: 24
Volume: -3.0 dB
Output: Adaptive (705.6/768 kHz)
DAC
Ares II: NOS mode
Connection: XLR (Ares II → SA-1)
SA-1 Gain: LOW
--------------------
Some formulas used by AI along the way:
--------------------
τ (time constant) = R_total × C_parasitic
Settling time (0.01% accuracy) = 9.2 × τ
For 16-bit ladder:
Switch impedance: 16 stages × 15Ω + 10Ω (PCB vias) = 250Ω
Parasitic capacitance: ~75pF
τ = 250Ω × 75pF = 18.75 ns
Settling time = 9.2 × 18.75ns = 176 ns
For 24-bit ladder (my current hardware):
Switch impedance: 24 stages × 15Ω + 15Ω (PCB vias) = 375Ω
Parasitic capacitance: ~75pF
τ = 375Ω × 75pF = 28.1 ns
Settling time = 9.2 × 28.1ns = 258 ns
Resistor Tolerance vs Bit Depth:
Formula: Tolerance ≈ 1 / (2^N × 2)
20-bit: 1 / (1,048,576 × 2) = 0.48 ppm
24-bit: 1 / (16,777,216 × 2) = 0.03 ppm
--------------------
Well, the idea was not to dump the entire chats, but sharing how LLMs can guide you though the process, and you can ask it to show the math, that forces the AI not to use jargons, but trying hard to talk to an engineer.
This may be off if you are an electrical engineer (I'm a software one), but the interesting part was to start comparing results between LLMs and start getting to a consensus base the actual capabilities of the hardware.
How does it sound? Oh yeah great, but I started with the spatial (jazz) filter, and ended up creating 2 more as other genres / material didn't sound that impressive.
I guess one day we will have an adaptive DSP based on LLMs.
Reference (all genres)
Option A (recommended)
Option C
DSD Handling
