Experimental SSTV Studio

Time-compressed slow scan television: encode an image to SSTV audio, squeeze the audio to a fraction of its length before it hits the air, and let the receiving station stretch it back and decode as usual. Half (or a quarter) of the airtime, paid for with a few seconds of number crunching on each end.

The idea comes from an old trick with audio editors: speed a recording up 3–4×, save the much shorter file, and later time-stretch it back — with a decent algorithm the result was surprisingly close to the original. This project asks the obvious ham question: what happens if the "recording" is a Martin M1 transmission? Shorter bursts mean less exposure to QSB and QRM, easier fits into short openings, and less time holding a frequency. The catch is that this is not a standard mode: both ends need this tool (or an equivalent external time-stretcher). Treat it as an experiment between consenting stations, and identify your transmissions as your local regulations require.

It's on GitHub — va3jfl/SSTV-Studio, license and all. The project has a home now, and that deserves a moment. Now, the README trick: plain Markdown image syntax can't scale, but GitHub allows HTML tags with width attributes inside READMEs, and an HTML

gives you the grid. One wrinkle your filenames add: the spaces need to be %20-encoded in the paths. Here's a ready-to-paste block — first row shows two screenshots side by side at half width each, second row centers the third, every image clickable to open full size: html## Screenshots

FM turbo 8× — clean loopback	SSB 2.7 kHz + noise — 7×
Wideband — 8×

Quick start

pip install numpy scipy Pillow sounddevice
python3 sstv_studio.py

sounddevice is optional — without it everything works except Play and live listening. On Linux you may also need the Tk package: sudo apt install python3-tk. Python 3.9+.

Headless check that your install works end to end:

python3 sstv_studio.py --selftest

Using the studio

The window is split into a TX side and an RX side over a shared "on-air scope" (a spectrogram of whatever was last transmitted or received) and a log.

Transmit: load an image (or click Test pattern), pick a mode, a compression factor and a method, optionally type a callsign to burn into the image, then Encode + compress. The status line shows the trade: e.g. Martin M1 Baseband 115.7 s → on-air 57.9 s (2× stretch). Save the WAV and play it into your rig like any SSTV audio, or click Play to key via VOX.

Receive: either open a recorded WAV, or click Start listening — the RX side then waits with a VOX squelch (threshold slider in the toolbar, one second of pre-roll is kept so the start of a burst isn't clipped), records until the signal drops, restores the time base and decodes, painting the image line by line as it lands. With auto-detect on, the decoder finds the compression factor and method by trial-restoring the first seconds and looking for a valid VIS header, so the RX operator doesn't need to be told the settings in advance.

Loopback and channel simulator: Send → RX (loopback) pushes the freshly encoded burst straight into the receive chain, optionally through a simulated SSB filter (2.7 or 2.4 kHz) with additive noise at a chosen SNR. This is the quickest way to get a feel for what each factor does before going on air.

The two methods, and the physics

FM turbo (recommended). SSTV is frequency modulation, so the honest way to run it faster is to speed up the modulation itself: demodulate to an instantaneous-frequency track, resample that track — an exactly linear operation with no artifacts of its own — and resynthesize the carrier. The deviation range (1200–2300 Hz) is untouched, so pitch is inherently preserved and the signal fits a normal SSB voice channel; it is equivalent to transmitting the same mode with a faster scan clock. The scaler's inner demodulator is factor-aware: a k×-compressed signal has k× the modulation rate and therefore k×-wider sidebands (Carson), so its passband opens with the factor — capped at the measured spectral edge of the input, since demodulating past where the channel's filter cut off only admits noise, and with the low side floored at 900 Hz only when the spectrum shows actual noise below the carrier. This single change is what makes 5–8× practical: with the old fixed narrow band, 8× lost ~6 dB of sharpness that was sitting in the shaved sidebands all along. What remains is pure channel physics: compressing time multiplies the modulation rate, so by Carson's rule the sidebands widen and a 2.7 kHz filter shaves some of that energy — fine horizontal detail softens gradually as the factor rises. In loopback this method loses only a few dB to a direct transmission even at 4×, and it is also the best receiver for the vocoder method below.

Resample. The tape trick: play the audio N× faster. Round-trips essentially losslessly — but every frequency is multiplied, so a 2× Martin transmission occupies roughly 2400–4600 Hz and a 4× one reaches 9.2 kHz. An SSB filter will destroy it; use it only over channels that pass the widened spectrum (wide FM audio, direct cable, digital links), or as a lab reference.

The decoder is shared by all paths: it demodulates instantaneous frequency with a precision-safe conjugate-product method, parses the VIS header, then tracks the sync-pulse train line by line with a small PLL-style follower rather than assuming a rigid line grid — absorbing residual time-base wander and ordinary soundcard clock drift alike. The receiver measures channel noise on the VIS leader tone and picks its demodulator bandwidth accordingly — wide for clean signals (maximum edge sharpness), narrow for dirty ones — and the sync tracker refines every pulse to sub-sample accuracy before smoothing the line grid against the knowledge that the true time base is nearly linear. An optional cleanup pass adds content-based de-jitter (each channel straightened against its own rows, with a do-no-harm check), and a detail-preserving denoiser — inspired by VE3NEA's SSTV Image Denoiser — cleans off-air receptions: it measures the actual noise level, removes impulsive line streaks, then averages each pixel only with patches that genuinely match, searching mostly vertically where SSTV noise is statistically independent. Luma is treated gently, chroma firmly, and clean images pass through untouched; when the noise is heavy — including the horizontally-streaked noise that a high-factor FM restore makes out of channel hiss (stretching the track stretches the noise with it) — the filter strength escalates automatically. A separate Sharpen toggle applies a halo-clamped unsharp on luma: it measurably crispens hard graphics at high factors and is a matter of taste on smooth photographs, so it is off by default. Three settings in the RX panel: Off / Normal / Strong. The channel simulator also offers an "FM 3 kHz (FRS)" model — pre-emphasis, a 300–3000 Hz shelf, and de-emphasised noise — for rehearsing handheld and repeater paths.

What to expect (measured)

Numbers from --selftest, which round-trips a deliberately nasty synthetic test pattern (color bars, gradient, fine checkerboard, pixel text) — real photographs fare visibly better. PSNR against the transmitted image, "lock" is the fraction of scan lines whose sync pulse was confidently found:

Path	On-air time	PSNR	Sync lock
Martin M1 direct (no compression)	115.7 s	36.7 dB	100 %
Martin M1, resample 2×	57.9 s	36.7 dB	100 %
Martin M1, FM turbo 2×	57.9 s	31.1 dB	100 %
Martin M1, FM turbo 3×	38.6 s	27.7 dB	100 %
Martin M1, FM turbo 4×	28.9 s	27.6 dB	100 %
Martin M1, FM turbo 5×	23.1 s	25.9 dB	100 %
Martin M1, FM turbo 6×	19.3 s	25.2 dB	100 %
Martin M1, FM turbo 7×	16.5 s	24.6 dB	100 %
Martin M1, FM turbo 8×	14.5 s	23.7 dB	100 %
Martin M1, FM 3× through SSB 2.7 kHz at 10 dB SNR, auto-detected	38.6 s	21.4 dB	100 %
Martin M1, FM 5× through SSB at 10 dB SNR, auto-detected	23.1 s	19.0 dB	100 %
Martin M1, FM 8× through SSB at 12 dB SNR, auto-detected	14.5 s	17.0 dB	96 %
Robot 36, FM turbo 2×	18.7 s	22.3 dB	100 %
Scottie S2, FM turbo 2×	36.3 s	24.8 dB	100 %
Photo, FM 3× through SSB at 10 dB SNR — denoised	38.6 s	29.3 dB	100 %
Photo, FM 3× through SSB at 6 dB SNR — raw / denoised	38.6 s	22.6 / 27.2 dB	100 %

Command line

python3 sstv_studio.py --encode photo.jpg --mode "Martin M1" \
    --factor 3 --method fm --callsign "VE3XYZ" --out onair.wav

python3 sstv_studio.py --decode onair.wav --factor auto --out rx.png

--factor auto on decode probes factors and methods until a valid VIS appears. --selftest (optionally with --demo DIR to dump PNGs, or --quick) runs the full QA matrix and exits nonzero if the core codec misbehaves.

The FRS "kid lab" setup

FRS handhelds (licence-exempt walkie-talkies) are narrowband FM with a hands-free jack — which happens to be exactly the channel this program likes. A workable across-town picture link:

Radio speaker / hands-free output → small audio isolation transformer (600:600 or similar; isolation matters more than exact matching — it kills ground-loop hum) → PC line-in. PC headphone output → roughly 100:1 resistive divider (e.g. 22 kΩ series into 220 Ω across the mic pins) → radio mic input, with the radio's VOX turned on. Then tick VOX header in the TX panel (or pass --vox on the CLI): it prepends a 0.7 s 1900 Hz tone so the radio's VOX attack and the far end's squelch open on a sacrificial tone instead of eating the VIS header — in simulation, 2×–8× all decode through an FM voice channel with 250 ms of VOX clipping, and without the header the same transmission fails outright. Rehearse first with the channel simulator set to "FM 3 kHz (FRS)".

Keep the radios themselves stock: FRS certification requires the integral antenna, so the fun experimental antenna (a hand-wound UHF loop!) belongs on a receive-only device instead — a cheap RTL-SDR dongle plus a kid-built loop makes a superb dedicated receiver, and receiving is unrestricted.

Rules honesty: FRS is a licence-exempt voice service in both Canada (ISED) and the US (FCC), and the provisions for non-voice content are narrow and change over time — check the current wording before making it a habit, keep transmissions short (another argument for the high factors), identify, pick a quiet channel, listen first, and always yield to voice traffic. Supervised kids sending 15-second pictures across town, courteously, is very much in the spirit of why these services exist — and it is a gateway drug to the ham licence.

Practical notes

Feed and take audio at sane levels — the decoder is FM-based and tolerates a lot, but clipping in a soundcard or an over-driven mic input hurts the stretch mode disproportionately. On SSB use FM turbo. 2× and 3× are essentially free; 4× is comfortable on any decent path; 5× through 8× lock reliably in the simulator down to ~10–12 dB SNR — a two-minute Martin M1 frame in 14.5 seconds at 8× — but the SSB filter is eating real sideband energy at those rates, so expect gradually softer horizontal detail and step up only as conditions allow. Keep the receiver's filter as wide as it goes. If VOX triggers on band noise, raise the threshold slider until the level meter's marker sits above the idle level. The vocoder briefly allocates a few hundred megabytes while restoring the long modes on the RX side; on a Raspberry-Pi-class machine prefer Robot 36/72 or Martin M2. And since the whole point is a nonstandard waveform on the air: announce what you're doing, keep your callsign in the image and in your ID practice, and check your local rules about experimental transmissions.

Files

• sstv_studio.py — everything: GUI, encoder, decoder, vocoder, channel simulator, self-tests.
• requirements.txt — Python dependencies.