The chain
Filter, envelope, threshold, durations, two clusters (dots, dashes), unit hence speed (1200 ÷ unit), silences cut into letters and words, table.
The threshold
Fixed, it fails in QSB. Adaptive, it tracks noise and peak: here, halfway over 1.5 s, never below 2.5 times the noise floor.
Three decoders
TellusMe, the site author's software: fast (threshold), fine (Bayesian), neural (CTC). They agree on the regular CQ. One extract per condition: we observe, we don't rank.
Their limits
Neural: closest to the reference at 44 wpm and on the weak signal (reference partly drawn from its reading), worst on the moderate CQ, ≈ 2 s latency, errors on hand keying. Fast and fine chop up hand keying; the fine one invents text in the noise.
Confidence
Fine and neural are saturated and poorly calibrated: 6% of their characters at 0.9 or more are wrong. But they do sort: only about 40% right below 0.9.
State of the art
Guenther 1973, Bell 1977 (Kalman, trellis), Mills 1977 (Viterbi), CW Skimmer 2008, AG1LE's CTC 2019 (synthetic). We cite only what the abstracts say.
RBN
Repeats: 2, 3 or 4 depending on the callsign; 256-character buffer; 10-minute ignore window on the same frequency. CC Cluster rejects single spots.
Going further
What is measured here
For each extract, each decoder's output is aligned with the reference reading (minimum edit distance); a character is « in agreement » if it lands on the same character. A text's agreement is 1 − distance ÷ length of the reference, leaving out its illegible parts. The reference is not guaranteed: it is a consensus of the three decoders and of listening, retained when the structure confirms it (repeated callsign, CQ, coherent exchange).
The figures for the three decoders, over the eight extracts
Share of characters displayed with a confidence of at least 0.9: fast 2%, fine 76%, neural 85%. Among the characters that disagree with the reference, 4%, 24% and 37% keep 0.9 or more. Read the other way round: a fine or neural character at 0.9 or more is right 94% of the time; below 0.9, about 40%. Area under the ROC curve: 0.87 (fast), 0.84 (fine), 0.79 (neural). Confidences are saturated and poorly calibrated, yet they sort clearly. Mean latency between the estimated end of a character and its output: 0.3 s, 1.1 s, 1.9 s.
The limits of these figures
Small sample: one extract per condition, 27 to 37 wrong characters per decoder. Partly circular reference: it is a consensus of the same decoders and of listening; on the weak signal and the very fast CQ, it retains readings from the neural decoder alone. On the moderate CQ, the neural decoder agrees at 65%, against 84% and 77% for the fast and fine.
The early work
MAUDE (1959), Guenther (AFIT, 1973, on a PDP-12), Bell (NPS, 1975 then 1977), Hickey (NPS, 1976), Bell (NPS, 1977: memory-conditioned Markov process, optimal Bayesian state estimation, Kalman filters on an evolving trellis, performance simulated on random letters), Mills (Maryland, 1977: Viterbi decoding, matched filtering, Markov model). Only their abstracts were read: no performance figure is therefore quoted.
CTC and networks
AG1LE (2019): CNN-LSTM-CTC, 4 s extracts turned into 128 × 32 images, 5,000 synthetic samples (a word of five characters at most, 20 to 30 wpm, 600 Hz tone); error rate of 0.1% on validation, but inconsistent logs and one take per condition. An April 2020 follow-up reports 25,000 extracts and 1.5% errors. Other projects (DeepFist, MorseNet) remain to be evaluated on real traffic.