Both faults reported on the air, both reproduced in synthetic signals first —
the existing tests all passed because each starts a fresh decoder and sends
textbook spacing, which is neither of the cases that hurt.
Speed change. A 32 wpm station followed by a 14 wpm reply decoded as
"TTT TTTT TTTTT TT…" for most of the over. Not the element classifier: the
CHARACTER-gap threshold is 2.2 dits, so at the stale fast estimate it stood at
81 ms while the newcomer's element gaps were 86 ms. Every single element was
flushed as its own character, and a lone element with no contrast reads as a
dah. A silence long enough to end an over now drops the estimate back to the
seed, which is exactly the state a freshly started decoder is in — and that case
was always fine. The threshold scales with the current estimate (12 dits, floor
600 ms) because a fixed one cannot serve both 10 and 40 wpm.
Callsign split. The word boundary sits at 4.6 dits, the geometric mean of a
3-dit letter gap and a 7-dit word gap. It assumes textbook spacing; a fist
leaving 5 dits between letters had every letter turned into a word, so a
callsign arrived as "O Y 1 C T". The boundary now also follows the letter gaps
this operator actually sends, whichever is larger. A wide sender's words may run
together — a far smaller price than a callsign in pieces.
The first RX-audio CW decoder decoded poorly on real signals (vs SDC) and was
removed in cafade0. This rebuilds the DSP core from scratch in
internal/cwdecode, keeping v1's good ideas (pitch lock, Flex cw_pitch
targeting, tiered acquisition) and fixing its proven failures:
- Two-way DEBOUNCE (pending-commit ~0.3 dit): v1 rejected short marks but not
short spaces, so a one-hop fade inside a dah shattered it into dits — the
single worst real-signal failure.
- Per-character BATCH dit/dah classification at flush time, using the batch's
own bimodal boundary — the first letter of an over decodes at any speed; the
timing clusters (muDit/muDah) are updated with the same attribution, killing
the classify-then-learn feedback spiral ("all dits at 60 WPM").
- dB-domain envelope with separate floor/peak trackers, hysteresis slicer,
span CAP (30 dB — else quiet backgrounds put the off-threshold in the
analysis-window skirts and letters merge) and window de-bias on durations.
- Double squelch: span >= 6 dB AND peak >= bank-median + 9.5 dB (span alone
cannot reject pure noise).
- Hamming-windowed Goertzel, 16 ms window / 5 ms hop (a 20 ms window left
40 WPM inter-element gaps with no envelope dip).
- Hardened acquisition (overlapping windows made "3 stable hops" meaningless)
plus SUPERVISED RE-LOCK: a clearly stronger tone on another pitch sustained
~1 s takes over — heals noise locks and follows a QSY.
- Gap-fed dit-length tracking (inter-element gaps are timing evidence too).
Proven by a synthetic-signal test suite (all passing): clean 12-40 WPM, three
pitches, added noise, QSB fading 0.35-1.0, 10 ms dropouts inside dahs, QRM in
auto and targeted modes, noise-only squelch, mid-over speed change 25->15 WPM,
and jittered hand keying (+/-20-25%, 3 seeds).
Wiring and UI restored from git (app_cw.go, Ear header button, decoded-text
strip with WPM/pitch/level + pitch lock + click-a-word-to-fill-callsign, Tools
menu entry) — strip strings translated (cwd.* i18n keys, EN/FR) this time.
- Removes the RX-audio CW decoder: deletes app_cw.go + internal/cwdecode, the
cwDecoder/cwPitchHz/cwMu/cwStop App fields, and all frontend state/effects, the
header Ear button, the Tools menu item and the decoded-text strip. The CW
KEYER (WinKeyer/Icom/Flex) is untouched.
- Fix: typing a callsign while a macro is keying now aborts the CURRENT engine
(Flex CWX / Icom / WinKeyer) on the first character, not just WinKeyer during
an auto-call loop — via a shared stopKeyerTx() helper.
rouggy