The acoustic shape
A trained operator recognises the whole character, then groups, then words. Below 10 wpm, according to Koch, the shape falls apart.
Koch, 1936
No chart: sound only. Two characters at the target speed, one more at 90%. Order reconstructed by the LICW: L F C K R D P X Y Q B G A N Z U V W J H S I E O M T.
Farnsworth
Characters at c, silences stretched for the effective speed s: tf = (300c − 186s) / (95cs) s. At 20 / 10: 654 ms between letters.
The filter
Ideal gain over 2.5 kHz, white noise: 7 dB at 500 Hz, 10 at 250, 14 at 100, 17 at 50. But a narrow filter smears the dots.
QRM, QSB
QRM: other stations. QSB: fading, about twenty dB on the real extract. Against both: filter, context, and repetitions.
The SNR
A signal-to-noise ratio only means something with its bandwidth. Here: in 500 Hz. Published readability thresholds depend on the protocol and the bandwidth: there is no single figure.
Montnémery
Five papers (1991–1995). According to the 1995 abstract, humans do better than the decoders tested at low SNR and below 100 characters per minute. Decoders of the 1990s.
Going further
Farnsworth, derived
On PARIS, 31 character units at tp = 1.2 / c s and 19 space units at tf must fit in 60 / s s: 31 × 1.2 / c + 19 tf = 60 / s, hence tf = (60 / s − 37.2 / c) / 19 = (300c − 186s) / (95cs). Check: s = c gives tf = tp. Examples: 20 / 10 → 218 ms per unit, 654 ms between letters, 1,525 ms between words; 25 / 6 → 1,344 ms between letters; 18 / 12 → 463 ms.
The simulated filter
Four identical band-pass stages (Web Audio's BiquadFilter, formulas from the RBJ « cookbook »), their Q chosen so that the −3 dB width is exactly the one displayed. Its equivalent noise bandwidth is 1.13 times the −3 dB width: the real gain is about 0.5 dB below the ideal gain (6.5 dB at 500 Hz, 9.5 at 250, 13.5 at 100, 16.5 at 50). Rise time from 10 to 90% of a dot at 20 wpm: about 3 ms at 500 Hz (the 5 ms keying edge accounts for most of it), about 4 ms at 250, 7 to 8 ms at 100, about 14 ms at 50 (calculated for this episode; the exact value depends on how the envelope is measured). Attenuation at +200 Hz from 650 Hz: 1.6 dB at 500 Hz, 5.6 at 250, 21 at 100.
Ground-truth signals
Made by cw-synthese.js: a 650 Hz tone with 5 ms cosine edges, Gaussian white noise set for the desired SNR in 500 Hz (σ² = (A²/2) · (fs / 1,000) / 10SNR/10), a fixed seed: each signal can be regenerated sample for sample.
Montnémery et al., in detail
1991: 10 men with normal hearing, trained, from 13 to 103 characters per minute; recognition falls as speed rises and SNR falls. 1995: five decoders, an 800 Hz tone in 185 Hz noise then wideband, SNR lowered in 1 dB steps, from 20 to 200 characters per minute; at 80 characters per minute, they needed +8 to +12 dB (narrow-band noise) and −4 to +3 dB (wideband) for 100% recognition. These figures come from the abstracts: the papers themselves were not read.