measure a frequency

Phase drift

Slowing things down

DSP pipeline

phase coherent STFT

Continuous self calibration with NTP/chrony

Soundcards clocks are commonly off by several ppm (12ppm 1sec/day)

• chrony calibrate the system's clock
• absolute time is hard to get, but relative clock's rate is easy
• we use the system's clock to calibrate the soundcard's clock

Soundcard's clock tracking after 1 hour

2€ usb soundcard on a Raspberry pi zero v1.1 (armv6)

Pulse train's energy isn't at their rate

Beautiful Hilbert transform algorithm

(I learned it existed by reading faust's lib)

Hilbert transform

void hilbert(std::size_t size, float* input, float* output_r, float* output_i) {
    assert(size % 4 == 0);
    for (std::size_t i = 0; i < size; i += 4) {
        // Modulation by exp(-j*pi/2*n) : sequence 1, -j, -1, j
        output_r[i + 0] =  input[i + 0];   // *1
        output_i[i + 0] =  0.0f;

        output_r[i + 1] =  0.0f;           // *(-j)
        output_i[i + 1] = -input[i + 1];

        output_r[i + 2] = -input[i + 2];   // *(-1)
        output_i[i + 2] =  0.0f;

        output_r[i + 3] =  0.0f;           // *j
        output_i[i + 3] =  input[i + 3];

        // lowpass at à sr/4
        float* c[] = {output_r + i, output_i + i};
        lowpass.process(4, c);

        // demodulation by exp(j*pi/2*n) : sequence 1, j, -1, -j

        // i+0 : (*1) (nothing to do)

        // i+1 : (*j) -> (-b + j a)
        float tmp_r1 = output_r[i + 1];
        float tmp_i1 = output_i[i + 1];
        output_r[i + 1] = -tmp_i1;
        output_i[i + 1] =  tmp_r1;

        // i+2 : (*-1) -> (-a - j b)
        output_r[i + 2] = -output_r[i + 2];
        output_i[i + 2] = -output_i[i + 2];

        // i+3 : (*-j) -> (b - j a)
        float tmp_r3 = output_r[i + 3];
        float tmp_i3 = output_i[i + 3];
        output_r[i + 3] =  tmp_i3;
        output_i[i + 3] = -tmp_r3;
    }
}

coherent STFT for quartz watch

Linear regression on phase

STFT mechanical watch

RT sequence

Async sequence

there's no client's request implemeted yet, its periodic request in the main loop.

Results

• GPS's pps (very accurate 1Hz) measured with sub ppm accuracy with 2 hours measurement
• radio clock DCF77's carrier (77.5KHz) measured with sub 0.1ppm accuracy with 5 minuts measurement (with a simple wire hooked up to the soundcard)
• good at picking watch's quartz (32.768kHz) sound with even with standard microphone

Limitations / further work

• Assumes steady soundcard clock during measurement
• Assumes steady signal frequency
• Use more ressource than needed as it is. (64hz is way too much)
• No coherent integration for signal detection
• use only one reference clock and one soundcard
• pure real time clock would be better for very low frequency
• Implement actual client/server architecture (soon)

LLM's help

Mostly Qwen3.6-35B-A3B with Hermes agent

• writing test boilerplate
• simple refactor
• understanding a small snipet of code
• "add a CLI option for the parameter XXX"
• "use vcpkg for the dependancy XXX instead of cmake's fetch"
• write a whole module if its well specified (first interface, then tests, then implementation)

clanker's sabotage attempt

• "fix" bugs by relaxing tests constraint (eg : would change a tolerence from 0.0001hz to 0.25hz "because the bins are 1hz, so 0.0001hz is too strict")
• "this test doesn't pass but its a known issue" (it wasn't)
• mistake nominal samplerate with real samplerate (add an explicit cast)
• throws mutexes everywhere as soon as the word "thread" is in the context (the rt thread dont want mutex, the non-rt thread dont need any).
• add non-rt safe operation in the rt thread
• Ill designed benchmark that didn't reflected actual behaviour of two implementations
• break up the git repository and lose data
• poor code quality in general

LLM conclusion

I lost a lot of time trusting it too much, but things should be different with a frontier LLM

Thank you very much!

I am open for questions and for work!