Repository navigation
SDR for Linux 0.6.0
The equalizer release, on a new DSP base: the vendored WDSP moves from 1.29 to 2.11 (neural noise reduction, PureSignal 3.0), both audio chains get a parametric equalizer on a drawn curve, synchronous AM arrives, the TCI server gets a Unix socket, and the transmit side learned a few things from a stuck-TX hunt in October.
Equalizer
- Parametric RX and TX equalizer (header-bar icon or
e, issue #48): up to eight bands per side — peak, low shelf, high shelf with frequency, gain and Q — on a drawn response curve: drag a node for frequency and gain, wheel on it for Q, double-click the plane to add a band, right-click a node to remove it; a row per band for exact values, Add band and Reset. The RX profile acts ahead of the AGC and is bypassed in the digi modes so TCI decoders get raw audio; the TX profile is the pre-EQ ahead of the speech processor, for the live mic in voice modes only. The band model (RBJ cookbook biquads) is computed here and realized by WDSP's continuous-gain FIR equalizer. Both profiles persist.
DSP base: WDSP 2.11
- WDSP 1.29 → 2.11 (issue #22). The third noise-reduction slot is now NNR, WDSP's neural noise reduction with two selectable built-in models — Standard / Premium in Preferences → Audio → Receive, with an adjustable mask floor — each with a CW-trained variant that is switched in automatically in CW modes. rnnoise left the tree. NNR costs about 10 % (Standard) or 20 % (Premium) of one core per receiver.
- PureSignal 3.0. On the ANAN G2E it calibrates only with a one-line local patch to the vendored WDSP (
PS_NF_MAG_PREFILT_YMAX1.8 → 4.0): this radio's coupler/DAC ratio rises about 6 dB with level, and the stock pre-filter discards the low-level points before the fit, so every calibration was rejected — the same happens in piHPSDR 3.1 on this radio. Measured offline on recorded feedback and reported upstream as TAPR/OpenHPSDR-wdsp#10; details invendor/wdsp/VENDOR.md, the patch is pinned by a gate. With it, live on the G2E: two-tone 90 of 90 calibrations accepted, voice 52 of 64, the auto-attenuate following the level both ways. The two-tone pair is now 700 + 1937 Hz (a spacing that divides the sample rate confuses the fitter — WDSP guide, appendix D). - The FFTW wisdom cache is rebuilt once on the first start after this update (2.11 plans one more transform per size and names the cache
wdspWisdom01): about eight minutes on the development machine, behind the usual progress window. Every later start imports it instantly.
Modes and controls
- SAM — WDSP synchronous AM (issue #30): PLL-locked carrier (capture ±2 kHz about the dial), both sidebands or LSB/USB only (a phasing selector in the Filter dialog, not a filter — QRM on one side goes away without re-tuning), and carrier levelling against fading for AM and SAM (Preferences → Audio, on by default). Key
acycles AM → SAM. Transmitting in SAM is AM. - Drive and TUNE power are remembered per band (issue #47): the footer sliders show and edit the current band, a band change restores that band's two values. An existing install seeds every band from its current values.
- TX phase rotator (issue #23) — WDSP PHROT with its auto-optimizer and an asymmetry readout on the TX HUD; Preferences → Audio → Transmit, off by default. On the author's mic chain (Neumann TLM 103 + SPL Channel One) it brought nothing: the speech there is already symmetric, and the average power at the same envelope peak moved within 0.6 dB at any corner frequency, never upward (±0.2 dB on the clean chain, −0.6 to 0 dB with the speech processor on). It is there for chains where speech is asymmetric.
TCI
- Unix socket listener (issue #46): the server also listens on
$XDG_RUNTIME_DIR/sdr-for-linux/tci.sock(mode 0600) for clients on the same machine; TCP stays on by default and can be switched off or bound to 127.0.0.1 only (Preferences → TCI). About → Debug Info prints what is actually bound. Throughput over the socket equals TCP (192 kHz IQ, measured). log-for-linux connects over it from its next release (itsmainalready does); skimmer-for-linux follows (OK1BR/skimmer-for-linux#39). - TCI 2.1 stream headers (issue #43): word 8 of every Stream header is the per-stream packet counter the 2.1 specification defines — IQ counting from
iq_start, RX audio restarting with everytrx:notification, TX_CHRONO per over. The IQ centre stamps move to words 9–11 underiq_stamp:2;iq_stamp:1still serves the old layout (stamps at 8–10, no counter) for the released skimmer and goes away in the next release.samjoinedmodulations_list.
Transmit safety and diagnostics
Learned from two "the rig keeps transmitting" events on 2026-10-24, which turned out to be a 10GBASE-T Ethernet link dropping under RF on SSB peaks (a cable problem, not the radio or the app) — but the software had let the pedal re-key the radio the moment the link came back:
- Footswitch freshness guard (issue #16): the pedal counts only while the radio's HP-status packets are live (0.5 s); a pedal-keyed over drops within 0.5 s of a link loss and the first real status packet after the link returns decides — the host never re-keys by itself. Link-loss release: more than 1 s without status while keyed releases MOX/TUNE/two-tone and flushes the CW/RTTY queues; the reason shows up as a line in the readout block. Verified live with a simulated link loss (inbound packets from the radio dropped by nftables for 10 s while keyed).
- SWR floor (issue #17): the SWR indicator, the two-reading trip and the display act only at ≥ 5 % of the PA rating; below that the reading is held. The reverse-power floor at 1–2 W used to read as SWR 1.7 and flash the alarm at every RX→TX transition.
- A failed send no longer tears the radio link down: an
ENETUNREACHduring a NetworkManager re-activation used to end the session; the sender logs once and keeps going, the link recovers by itself. - Noise gate hold 0.8 s → 0.3 s (issue #18): the expander never closed between words at 0.8 s. Not yet judged by ear whether word endings survive — say so in an issue if they do not.
- Run log, opt-in (issue #20): Preferences → TCI → Diagnostics → Run log, or
SDRFL_LOG=1—~/.local/state/sdr-for-linux/sdr-for-linux.log, five runs kept, every line timestamped to line up withjournalctl. Off by default, nothing is written unless you switch it on. Its TX trace lines (PTT edges, keying requests, link state, per-over statistics) are what a stuck-TX report needs. - ANAN G2E gateware: 11.1.0 was measured (A/B flash) to put about 1.4 W of RF on the antenna during SSB pauses — for any non-zero DUC data, where 11.0.5 gives 0 W — which makes an auto-tuner re-tune between words. Stay on 11.0.5 until TAPR/OpenHPSDR-Firmware#4 is resolved.
Known and unverified
- The ANAN 10E and the Hermes Lite 2 have not been run since the WDSP 2.11 re-sync (last live run on 0.5.x). FT8 via TCI and switching the AF rate at run time were not re-tested under 2.11 either.
- The
--serverremote head speaks the piHPSDR 3.0 client/server protocol; a piHPSDR 3.1 server refuses it. - PureSignal on the G2E depends on the local WDSP patch above until upstream answers; the G2 (Saturn) and the HL2 were not re-calibrated under PureSignal 3.0.
- The phase rotator's auto-corner search was never exercised live (nothing to find on this chain).
Downloads: AppImage (glibc ≥ 2.43), .deb (Ubuntu 26.04+), .rpm (Fedora 44+). Arch users: sdr-for-linux on the AUR. The headless probes (sdrfl-*probe) are dev-tree binaries — build from source for those.