Creating Recordings

Before you can curate a dataset or train a model, you need recordings in the RIA Hub Library. This guide covers the two ways to produce SigMF recordings using the RIA Toolkit command-line interface:

• Synthetic generation: produce mathematically clean IQ signals with known labels and controlled SNR, using ria synth
• Live capture: record real RF signals from a connected SDR device, using ria capture

In both cases the output is a SigMF file pair (.sigmf-data + .sigmf-meta) that you push to a RIA Hub repository with Git LFS.

Note

If you want a fully in-browser, no-CLI workflow, the Dataset Manager → Generator produces a labelled synthetic dataset directly — no toolkit installation or recording step required. See Synthesizing a Dataset.

What you’ll need

• RIA Toolkit installed locally:

  pip install ria-toolkit-oss

• A RIA Hub repository with Git LFS initialised — see Working with Git LFS
• For live capture: one supported SDR device connected (PlutoSDR, HackRF One, BladeRF, USRP, RTL-SDR, or ThinkRF)

Part 1: Synthetic generation with ria synth

ria synth (also callable as ria generate) generates modulated IQ signals and writes them to SigMF files. The signal parameters are exact and reproducible, making synthetic recordings ideal for building a first labelled dataset.

Available signal types

Command	Signal
ria synth psk	Phase-shift keying (BPSK, QPSK, 8-PSK, …)
ria synth qam	Quadrature amplitude modulation (16-QAM, 64-QAM, …)
ria synth fsk	Frequency-shift keying
ria synth ook	On-off keying
ria synth oqpsk	Offset QPSK
ria synth gmsk	Gaussian minimum-shift keying
ria synth pam	Pulse amplitude modulation
ria synth chirp	Linear frequency-modulated chirp
ria synth noise	AWGN (noise-only baseline)
ria synth tone	Single-frequency tone

Common options

All ria synth commands accept these flags:

Flag	Short	Description
--sample-rate	-s	Sample rate in Hz (required)
--duration	-t	Duration in seconds
--num-samples	-n	Total number of samples (alternative to duration)
--center-frequency	-fc	Records the center frequency in the SigMF metadata
--add-noise	—	Adds AWGN to the signal
--noise-power	—	AWGN variance (default 0.1)
--output	-o	Output filename (required)
--metadata	-m	Write a KEY=VALUE pair into the SigMF global metadata; can be repeated
--overwrite	-w	Overwrite an existing output file

Modulation-specific commands also accept:

Command	Additional flags
psk, qam, pam	--order -M (modulation order), --symbol-rate -r, --filter, --filter-beta, --filter-span
fsk	--order -M, --symbol-rate -r, --deviation
ook	--symbol-rate -r
chirp	--bandwidth -b (required), --period -p (required), --type (up/down/up_down)

Example: modulation recognition dataset

The following commands produce three signal classes (QPSK, OOK, noise) that you can use as a starting point for a modulation recognition classifier. Run each block from your repository root.

QPSK (order 4, 500 ksymb/s, RRC pulse shaping, light AWGN):

for i in 1 2 3; do
  ria synth psk \
    --order 4 \
    --symbol-rate 500e3 \
    --sample-rate 1e6 \
    --duration 5 \
    --center-frequency 915e6 \
    --add-noise --noise-power 0.005 \
    --output "recordings/qpsk_${i}.sigmf" \
    --overwrite \
    -m label=QPSK \
    -m campaign=modrec-tutorial
done

OOK (200 ksymb/s):

for i in 1 2 3; do
  ria synth ook \
    --symbol-rate 200e3 \
    --sample-rate 1e6 \
    --duration 5 \
    --center-frequency 915e6 \
    --add-noise --noise-power 0.005 \
    --output "recordings/ook_${i}.sigmf" \
    --overwrite \
    -m label=OOK \
    -m campaign=modrec-tutorial
done

Noise (AWGN baseline class):

for i in 1 2 3; do
  ria synth noise \
    --sample-rate 1e6 \
    --duration 5 \
    --power 0.01 \
    --output "recordings/noise_${i}.sigmf" \
    --overwrite \
    -m label=noise \
    -m campaign=modrec-tutorial
done

You should now have 9 .sigmf file pairs (18 files) under recordings/.

The -m label=<value> flag writes label into the SigMF global block. This is the field the Curator reads to assign class labels when slicing — keep the value consistent across all recordings of the same class.

Tip

Add --iq-amp-imbalance and --iq-phase-imbalance to simulate receiver hardware imperfections, or --multipath-paths and --multipath-max-delay for fading channels. These make the dataset harder but produce more robust models.

Part 2: Live SDR capture with ria capture

ria capture streams IQ samples from a connected SDR device and writes a SigMF file. The toolkit handles device initialisation, sample streaming, and metadata creation automatically.

Discover connected devices

ria discover

Sample output:

Found 1 device(s):
  pluto  —  PlutoSDR (ip:192.168.2.1)

If no devices are found, check USB or network connectivity. Drivers for USRP (UHD), HackRF, BladeRF, RTL-SDR, and ThinkRF must be installed separately.

Capture command

ria capture \
  --device pluto \
  --frequency 915e6 \
  --sample-rate 1e6 \
  --duration 5 \
  --gain 50 \
  --output-dir recordings/ \
  -m label=QPSK \
  -m campaign=modrec-tutorial

Flag	Description
--device	Device type: pluto, hackrf, bladerf, usrp, rtlsdr, thinkrf
--ident -i	Device identifier — IP address for PlutoSDR, serial for HackRF/BladeRF. Omit if only one device is connected.
--frequency	Center frequency in Hz
--sample-rate -s	Sample rate in Hz
--duration -t	Capture duration in seconds
--num-samples -n	Capture length in samples (alternative to duration)
--gain -g	RX gain in dB
--bandwidth -b	RX bandwidth in Hz (if the device supports it)
--output -o	Output filename. If omitted, a timestamped filename is generated automatically.
--output-dir	Output directory (default: recordings/)
--metadata -m	Write a KEY=VALUE pair into the SigMF metadata; can be repeated

To capture from a PlutoSDR at a non-default IP address:

ria capture --device pluto --ident 192.168.2.2 --frequency 915e6 --sample-rate 1e6 --duration 5

For a multi-step campaign (changing transmitter state between captures), see Campaign Control.

Part 3: Pushing recordings to RIA Hub

1. Initialise Git LFS for SigMF files

If the repository is new, run this once:

git lfs track "*.sigmf-data"
git lfs track "*.sigmf-meta"
git add .gitattributes
git commit -m "track sigmf files with lfs"

2. Commit and push

git add recordings/
git commit -m "add modrec-tutorial recordings"
git push

Large uploads (> 100 MB total): use the dedicated upload endpoint to bypass the Cloudflare limit on the main domain:

git remote set-url origin https://upload.riahub.ai/<owner>/<repo>.git
git push
git remote set-url origin https://riahub.ai/<owner>/<repo>.git

Or configure lfs.url in .git/config permanently:

[lfs]
    url = https://upload.riahub.ai/<owner>/<repo>.git/info/lfs

3. Verify in the Library

Open the repository and click the Library tab. RIA Hub detects .sigmf-data + .sigmf-meta pairs on push, queues a background task to parse them, and registers them as recording assets within about a minute.

If a recording does not appear, check the Project metadata field — it must be a valid repository slug. Recordings without a parseable .sigmf-meta will not be indexed.

sidebar: order: 2

Next steps

With recordings in the Library you are ready to build a dataset:

• Curating a Dataset — slice, qualify, and label recordings into an HDF5 training set
• Reviewing and Labelling Recordings — inspect recordings visually and correct labels before curation