sdr_notes_old.md

(very) old SDR comments

Steps

1. Get some FM radio data with RDS available:

• find with gqrx
• save with rtl_sdr -f 92.5M -s 2.4M -n 24000000 filename - used 92.5MHz (Radio Wanda Złote Przeboje), 1.8MHz, 54000000 (30s duration)
  • note: samples are saved as pairs (I/Q) of 8-bit integers (can infer from filesize - it's 2x num_samples)

2. FM demodulation with rtl_fm:

• command that works (without wideband-FM mode it won't work! pure static noise): rtl_fm -M wbfm -s 1800000 -f 92500000 | play -t raw -r 32k -es -b 16 -c 1 -V1 -
• just fine too: rtl_fm -M fm -s 170k -r 32k -f 92500000 | play -t raw -r 32k -es -b 16 -c 1 -V1 -

3. RDS with Redsea:

• command that works (171k is necessary!): rtl_fm -M fm -s 171k -f 92.5M | redsea

Useful commands

• play just left channel: play file.wav remix 1 0 (0 1 for right channel)

Notes

• complex signal (e.g. I/Q representation) sampled at Fs can represent frequency range [-Fs/2, Fs/2] - Nyquist holds, because each complex sample contains 2 real samples (~doubled sampling rate)

Questions

• rtl_fm:
  • why -s 170k is louder (and cleaner?) than -s 1800k?
    • possible answer: RTL SDR is set to sample around 1e6/s and downsample(!) to -s - does it handle such large values?
    • this is effectively bandwidth we take to demodulation too? (170k is reasonable, GQRX uses 160k by default)
  • does it tune to other (close) frequency to avoid DC spike interference?

Log

09.01.22

Figuring out how rtl_fm works

• settings:
  • demod.{rate_in, rate_out} = 171k, demod.rate_out2 = output.rate = 32k
  • demod.downsample = (1_000_000 / 170_000) + 1 (6 here) - in controller_thread_fn
  • demod.output_scale = 1 (for mode: FM) - in controller_thread_fn
  • controller_thread_fn - periodically changes dongle's freq and rate - WHY?
    • dongle_state.rate - demod.downsample * demod.rate_in - approx. 1_000_000...
    • dongle_state.freq - frequency + dongle_state.rate / 4 (controller_state.edge is 0 without -E edge)
• full_demod - the meat:
  • low_pass:
    • interesting: it groups RTL I/Q samples into packs of demod.downsample length and adds them - 6x shorter output
    • interesting because it's not convolution (not really FIR box filter)
      • update: it is convolution followed by downsampling (pretty ingenious)
    • input and result in demod_state.lowpassed
  • fm_demod:
    • moves (I/Q) sample by sample, computes phase difference between current and prev
      • atan2 + multiplication by complex conjugate
    • phase diff in [-pi, pi] -> divide by pi, multiply by 2^14 (WHY 14 and not 15?), cast to int16
  • low_pass_real (because rate_out2):
    • like low_pass, groups into packs of approx. demod.rate_out / demod.rate_out2 samples (5 here) and AVERAGES them
      • still not full convolution (followed by decimation as above)
    • casts to int16

13.01.22

Successful demodulation of Mambo no. 5

• implemented 2 algos - rtl_fm (works) and Lab6 (doesn't work)
• turns out SNR of ~20dB is required for decent quality (antenna higher, closer to window), previous sample was about ~10dB and barely intelligible
• but rtl_fm -M fm -s 170k -r 32k -f 92500000 | play -t raw -r 32k -es -b 16 -c 1 -V1 - does it with antenna in default location too (so I GUESS around 10 dB as well) - but how?
  • settings of rtl_sdr - e.g. tuning to different frequency and then shifting?
  • de-emphasis filter?
  • these rectangular simple decimator filters are actually better than scipy.signal.decimate???
  • rtl_fm uses 170 kHz instead of my 240 kHz for filter bandwidth (note: with 1.8MHz I have used 300 kHz and it yielded even worse results - maybe that's the thing??)

18.01.22

SSB modulation - figuring stuff out

• pieced together why I(t) and Hilbert transform of Q(t) give us both USB and LSB - ingenoius
• great mistery - why (and how, how well etc.) rtl_fm doesn't use Hilbert transform (nor filter, nothing) - just adds/subtracts I and Q?