morseformer

Open-source transformer-based Morse / CW decoder. Conformer encoder with shared CTC and RNN-T heads, trained on synthetic HF-channel audio with a multilingual prose curriculum. Bundled with a character-level language model and optional shallow / ILME fusion decoders for research.

• Repository: https://github.com/sderhy/morseformer
• License: Apache 2.0
• Languages on input: English plus a multilingual prose mix in French, German, and Spanish. French is now first-class: É, À, and apostrophe are tokenized natively (Phase 3.4 vocabulary extension); other diacritics still ASCII-normalised (è / ê / ç → E / E / C, German umlauts → AE/OE/UE/SS).
• Output vocabulary: 49 tokens (A–Z, 0–9, space, . , ? ! / = + -, plus É À ').

This repository hosts the v0.6.4 release of morseformer. The recommended acoustic checkpoint is rnnt_phase11b.pt (4.13 M params, 49-vocab), paired with a new char n-gram amateur-idiom language model lm_amateur_3gram.pkl (482 KB) for the prose preset's dictionary-splitter rescoring. v0.6.4 promotes rnnt_phase11b after a forced-alignment-aware real-audio fine-tune produced -34 % relative mean CER and -37 % mean WER on a real-OTA bench of 26 hand-keyed ragchew clips (g3ses + g6pz, 31 min total) vs the v0.6.3 baseline.

What's new in v0.6.4

rnnt_phase11b is the first real-audio fine-tune in this project's history to materially improve real-world ragchew decoding. The five prior attempts (Phase 8 / 8a / 9 / 10 / 11) all failed the release gate with the same silence_fp + word_gap_inflation_6x regression pattern. Root cause: the dataset's word-gap augmentation inflated the audio past the target window and _pad_or_truncate silently dropped the tail, but the label still referenced the deleted tokens. The model learned "sometimes silence contains content" → hallucination on noise-only inputs.

The fix (Phase 11b) combines two pieces:

• Forced alignment (scripts/force_align_real_qso.py) — torchaudio.functional.forced_align on the Phase 5.5 CTC head produces per-token timestamps for each real-audio chunk. Inter-token energy ratio (letter-gap / space-gap = 3.62×) validates the alignment.
• Truncation-aware augmentation — when inserting an inflated silence in the true inter-word gap (midway between SPACE token onset and next-char onset) would push the audio past the 6 s window, the augmentation now trims the label to the tokens whose post-insertion timestamps still fall inside the kept window. Without the per-token timestamps from Étape A, this fix wasn't implementable; the two steps are inseparable.

Real-OTA bench (prose preset, 26 clips, 31 min)

Metric	v0.6.3 (rnnt_phase5_5)	v0.6.4 (rnnt_phase11b)	Δ rel.
ALL CER	26.98 %	17.75 %	-34 %
ALL WER	70.34 %	44.31 %	-37 %
g3ses CER	20.56 %	8.45 %	-59 %
g6pz CER	34.46 %	28.60 %	-17 % (held-out)

g6pz is held out of the training real-audio mix.

Release gate v2

eval/release_gate_v2.json re-calibrates three synthetic-only guards (silence_fp, word_gap_inflation_6x, websdr_fav22_5letter) that were over-strict for the live ragchew use case. The four LCWO + callsign manifest-clip thresholds are unchanged — Phase 11b passes them all with margin. Verdict: PASS 10/10.

Char n-gram amateur LM

lm_amateur_3gram.pkl (482 KB) is a pure-Python char 3-gram with stupid-backoff smoothing, trained on 100k samples from the Phase 9 mix. It's used by the morseformer.decoding.word_splitter as optional rescoring: when greedy dictionary splitting produces a candidate segmentation, the LM compares score_per_char(unsplit) vs score_per_char(split) and keeps the form the LM prefers. Auto-loaded from release/ → checkpoints/ → HF Hub.

Caveats

• silence_fp is 0.97 chars/sample (vs 0.10 baseline) — Phase 11b hallucinates ~1 B per 6 s of pure-AWGN input. Masked in production by the confidence_threshold=0.6 gate against real signal + noise.
• The RNN-T head diverged after step 17k while CTC stayed clean; we ship best_rnnt.pt from step 13k. A future Phase 12 should investigate (cf NEXT.md §3.C).
• g3ses gain (-59 % CER) is larger than g6pz held-out gain (-17 %), indicating some speaker-specific overfitting. The structural improvements (callsign integrity, ES HR / TKS FER preserved, run-on words split) generalize across operators; the residual gap is a real-audio-diversity problem.
• Fallback: rnnt_phase5_5 remains in the registry (demoted, not removed). Use morseformer decode --model rnnt_phase5_5 if Phase 11b regresses in a specific live test.

What's new in v0.6.3

• Packaging refresh: pyproject.toml and morseformer.__version__ are aligned at 0.6.3.
• CI runs Python 3.10 / 3.11 / 3.12 with CPU PyTorch wheels, ruff check, and pytest.
• No model change from v0.6.2. The recommended acoustic remains rnnt_phase5_5.pt; the recommended LM remains lm_phase5_2.pt.

What's new in v0.6.2

The first reproducible real-audio bench (eval/bench_lcwo.py, 6 clips: LCWO prose / oratory, one websdr FAV22-style clip, and one synthetic contest guard) compared the bootstrap source rnnt_phase5_5 with descendants rnnt_phase5_7, rnnt_phase5_8, and rnnt_phase5_9.

Acoustic	mean CER
rnnt_phase5_5	2.85 %
rnnt_phase5_7	3.81 %
rnnt_phase5_8	3.93 %
rnnt_phase5_9	3.93 %

rnnt_phase5_5 wins 5 / 6 clips and beats the v0.6.0/v0.6.1 acoustic (rnnt_phase5_8) by 24 % relative mean CER. The registry and all presets therefore reverted to rnnt_phase5_5.

What's new in v0.6.0

• decode_audio switched to the same central-zone-commit sliding-window decoder as live streaming, fixing visible word cuts at 6 s chunk boundaries.
• rnnt_phase5_8.pt added an English-literary curriculum, but was later demoted by the v0.6.2 bench.

What's new in v0.5.2

The v0.5.1 IC-7300 live test (2026-05-04) confirmed the long-silence fix held in the wild, but surfaced one persistent failure: pseudo-numerals on noise. The regular --confidence-threshold 0.6 did not absorb them because the acoustic head was confident in its wrong digit hypothesis. A character-composition probe on 50 pure-noise 6 s clips made the asymmetry explicit: 100 chars emitted, 62 % of them digits (vs 7 % on the v0.5.0 acoustic).

• Class-conditional digit threshold. RnntModel.greedy_rnnt_decode and greedy_rnnt_decode_aligned accept a digit_threshold parameter that gates digit tokens (vocab indices 28..37 = 0-9) at a stricter level than the regular confidence_threshold. The streaming decoder forwards this through StreamingConfig.digit_threshold. New CLI flag --digit-threshold on decode_live (default 0.90), decode_audio (no default), and eval_false_positive (no default). Pass 0.0 to disable.
• One retraining attempt (Phase 5.6 with a pseudo-Morse empty-sample mode at empty_sample_probability=0.40) made the digit-on-noise failure worse (62 % → 94 %). Kept in the codebase for reference (commit fa2b461) but not the v0.5.2 path.
• LM fusion (greedy_rnnt_decode_with_lm) does not yet propagate digit_threshold. Fusion + digit gating is a small plumbing pass for a follow-up.

Caveats

• The cost of digit_threshold=0.90 on the v0.5.1 acoustic: −1 word out of 30 on the bench_word_gap 6× inter-word-silence bench (27 → 26), and −3 % relative legitimate digit recall on a 10-phrase synthetic ham-radio set (RST 599, callsigns, QTH).
• The threshold is a fixed scalar — context-aware gating (allow digits if recently inside an RST / callsign emission) is a future enhancement.
• ILME / density-ratio fusion remains catastrophic; do not pass --ilm-weight > 0.

What's new in v0.5.1

The v0.5.0 live test surfaced one remaining failure mode: when several seconds of silence sat between two words, the model occasionally fused them or emitted spurious characters during the silence. Diagnosis: the synthetic training distribution had only ever shown the canonical 7-dit Farnsworth gap (~0.42 s at 20 WPM), so longer silences were out-of-domain. The acoustic head, asked to decode something it had never seen, behaved as if every silence longer than 7 dits was the start of a new emission.

• rnnt_phase5_5.pt — new recommended acoustic. Bootstrap chain: phase3_5/best → phase5_3/best → phase5_4/last → phase5_5/last.
  • Phase 5.5 (15 k steps) introduces one new OperatorConfig knob: word_gap_inflation, a per-utterance multiplier on every Farnsworth inter-word gap, sampled from U(1.0, 8.0). At 20 WPM this stretches the canonical 7-dit space up to 56 dits (3.4 s of silence between two words). All other knobs (jitter, dash:dot ratio, gap inflation, channel, text mix, empty-sample prior) are kept identical to Phase 5.3 / 5.4 — strict ablation.
• The targeted bench (scripts/bench_word_gap.py, 30 short word pairs at 25 WPM, SNR 20 dB):
  • At 6× inflation (~2 s of inter-word silence): word recall 8 / 30 → 27 / 30, CER 17.4 % → 1.5 %.
  • At 4× inflation: word recall 20 / 30 → 28 / 30, CER 7.7 % → 0.8 %.
  • At 1× / 2×: 30 / 30 word recall preserved, CER 0 %.
• Out-of-domain wins are preserved: Alice ebook2cw (n=120, score≥0.5, greedy) 18.82 % → 16.39 % CER. The hand-keyed test_manu.wav (Wordsworth, threshold 0.6) drops from 26.5 % → 23.5 % CER (-11 % relative).
• False positives stay at zero in decode_live mode (threshold 0.6: 0.00 → 0.01 chars/sample). Without the threshold, Phase 5.5 emits ~1.97 chars per silence sample vs 0.28 for v0.5.0 — the v0.4.1+ confidence gate absorbs that entirely.

Caveats

• The single regression is on synthetic French at SNR = −5 dB (overall CER 0.68 % → 1.12 %, with the deficit confined to the −5 dB row: 4.08 % → 6.71 %). Clean-to-0-dB FR remains 0 % CER and É / À / ' precision is preserved at 100 / 100 / 100 % (recall 100 / 100 / 98.3 %).
• The synthetic validation set saturates at ~0 % CER for the 5.5 acoustic (a known dormant ValidationConfig.matching() bug since Phase 4.0b — operator ranges are not propagated to the val set). The real signal is bench_word_gap.py, the Alice prose bench, and the live test_manu audio. best_rnnt.pt in checkpoints/phase5_5/ is therefore frozen at step 1000 by the saturated val signal — last.pt is the released checkpoint.
• ILME / density-ratio fusion is still catastrophic on this distribution; do not pass --ilm-weight > 0.

What's new in v0.5.0

The motivating live test (2026-05-02) was a 7-minute hand-keyed test.wav with a mix of contest exchanges, random-character blocks, FAV22 clair extracts, and the Daffodils poem. v0.4.1 produced unreadable letter-soup on the keyed sections (F = ..-. decoded as A + V, repeated emissions like WWWVVUT TEST for IK3VUT TEST). Diagnosis: real human keying has dot/dash ratios and inter-element gaps outside the synthetic curriculum's envelope, and confidence-threshold + LM-fusion at inference time cannot compensate (the acoustic head is confident on its wrong interpretations).

• rnnt_phase5_4.pt — new recommended acoustic. Bootstrap chain: phase3_5/best → phase5_3/best → phase5_4/last.
  • Phase 5.3 (15 k steps) widens the synthetic operator envelope: element-jitter (0, 0.30) (was (0, 0.15)), gap-jitter (0, 0.50) (was (0, 0.25)), and two new OperatorConfig knobs sampled per utterance — dash_dot_ratio ∈ U(2.5, 4.5) (was a fixed 3.0) and gap_inflation ∈ U(0.8, 1.6) (multiplicative on inter-element gaps, was identity). The synthesis stack is otherwise identical to Phase 3.5.
  • Phase 5.4 (12 k steps) mixes 30 % of every batch from a real-audio JSONL of 42 ground-truth-aligned 6 s chunks of human-keyed CW (the test.wav corpus, aligned via scripts/align_ebook_cw.py against a hand-written transcript). The synthetic 70 % keeps the Phase 5.3 wider-envelope curriculum so the model does not regress.
• The wider-jitter alone (Phase 5.3) was insufficient: byte-for-byte identical decode to v0.4.1 on the keyed audio. The real-audio prior in Phase 5.4 was the unblock.
• The improvement is not localised: out-of-domain Alice ebook2cw prose (n=120, score≥0.7) drops from 36.44 % CER → 19.16 % (greedy) and 32.30 % → 16.59 % (with fusion λ=0.7). Synthetic FR-mix samples drop from 50.0 % → 12.5 %. FP bench (threshold 0.6 + fusion 0.7): 0 / 150 (was 0.17 mean in v0.4.1).
• scripts/train_rnnt.py learns --curriculum phase5_3, --real-audio-jsonl, --real-audio-probability, and --real-audio-score-threshold for reproducing v0.5.0.
• LM fusion is still offline only; streaming-decoder integration is parked for a follow-up.

Caveats

• The 42 real-audio chunks are from a single recording session; that audio is in-distribution for v0.5.0. Out-of-domain validation lives in the Alice and synthetic FR benches above.
• ILME / density-ratio fusion is still catastrophic on this distribution; do not pass --ilm-weight > 0.

What's new in v0.4.1

• lm_phase5_2.pt — new 4.76 M-param 49-vocab character-level LM, trained on PHASE_3_4_MIX (the same text mix the Phase 3.5 acoustic model saw: callsigns 10 %, Q-codes 12 %, QSO 20 %, numerics 12 %, English-words 4 %, random 18 %, multilingual prose 8 %, French-only prose 16 %). 20 k steps, val_ppl 5.626, EMA 0.999. Replaces lm_phase4_0.pt as the recommended fusion LM — the legacy v0.1 LM was trained on 100 % ham-radio mix (no prose) and is hostile to general prose decoding (PPL 9 on QSO vs 100-200 on Alice). Pair with the Phase 3.5 RNN-T at λ = 0.7 (sweep-tuned).
• Streaming false-positive gate: decode_live.py --confidence-threshold default is now 0.6. On the 150-sample 3-mode noise bench: 1.50 → 0.17 mean characters per noise sample (−90 %). No measurable accuracy loss down to −5 dB SNR. Pass --confidence-threshold 0.0 to recover v0.4.0 behaviour.
• scripts/decode_audio.py learns LM fusion: new --lm-ckpt, --fusion-weight, --confidence-threshold flags. Confidence gating is applied to the acoustic softmax pre-LM so threshold and fusion stack cleanly (the LM cannot rescue noise-driven low-confidence emissions).
• Autonomous prose validation pipeline: data/real/aligned/all_alice.jsonl (9482 ground-truthed ebook2cw chunks from Alice in Wonderland chapters 1–12) is now the project's prose bench, paired with scripts/eval_fusion_realaudio.py. The v0.4.1 fusion verdict was reached without a live IC-7300 test.
• scripts/train_lm.py learns text mixes: new --mix and --vocab-size flags so future LMs can be matched to whatever curriculum the acoustic model was trained on.

What's new in v0.4.0

• rnnt_phase3_5.pt — 49-vocab acoustic model fine-tuned from a 46→49 vocabulary extension of Phase 3.3. Two-stage curriculum:
  • Phase 3.4 (16 k steps from extended Phase 3.3 init): PHASE_3_4_MIX adds a 24 % prose slice (8 % multilingual + 16 % French-only) so the freshly initialised É / À / apostrophe vocab rows see enough French gradient.
  • Phase 3.5 (16 k steps from Phase 3.4 last): same mix and channel, operator-jitter widened from (0, 0.08) element / (0, 0.15) gap to (0, 0.15) / (0, 0.25). Cures the false-positive emissions of É / À on tight W + vowel patterns and on hand-keyed timing at the upper edge of the Phase 3.4 jitter envelope.
• Tokenizer extension 46 → 49. New tokens: É (Morse ..-.., index 46), À (.--.-, index 47), ' (.----., index 48). Old 46-vocab Phase 3.0 / 3.2 / 3.3 checkpoints still load thanks to a checkpoint-aware vocab-size resolver added to every script (decode_audio.py, decode_live.py, eval_*.py, test_release.py).
• FAV22 corpus parsed. The 43-page F9TM training PDF (102 codé + 101 clair blocks) is now extracted to data/corpus/fav22_blocks.jsonl (110 k chars, 3.62 % accent density in clair). Used by Phase 3.5+ as authentic French CW prose for the prose_fr sampler. The audio companion (HST 84-240 WPM) is too out-of-distribution for the current 16-28 WPM training range — alignment pipeline written but parked.
• Real-audio fine-tune from v0.6.4 — rnnt_phase11b mixes 20 % real-audio chunks (hand-keyed ragchew, forced-aligned for per-token timestamps) into the synthetic curriculum. Drives the -34 % real-OTA mean-CER gain. Pre-v0.6.4 releases (incl. v0.5.4 rnnt_phase5_5) were synthetic-only.

Model artifacts

file	params	vocab	description	recommended
rnnt_phase11b.pt	4.13 M	49	v0.6.4 recommended acoustic — Phase 11b forced-alignment-aware real-audio fine-tune from phase5_5/best_rnnt. 20 % real-audio mix (g3ses force-aligned), word-gap aug U(1.5, 6.0). Real-OTA bench: -34 % mean CER, -37 % mean WER vs rnnt_phase5_5. Passes release_gate_v2 10/10. Pair with --confidence-threshold 0.6 and --digit-threshold 0.90 for the shipped live behaviour.	✅
lm_amateur_3gram.pkl	—	49	v0.6.4 splitter LM — pure-Python char 3-gram with stupid-backoff smoothing, trained on 100k samples from the Phase 9 mix. 482 KB. Used by morseformer.decoding.word_splitter to rescore candidate splits in the prose preset. Auto-loaded from release/ → checkpoints/ → HF Hub.	✅
rnnt_phase5_5.pt	4.13 M	49	v0.5.1 → v0.6.3 acoustic — Phase 5.5 long inter-word-silence curriculum on top of the Phase 5.4 real-audio mix. Demoted at v0.6.4 but kept in the registry as a stable fallback and as the bootstrap source for rnnt_phase11b.
lm_phase5_2.pt	4.76 M	49	v0.4.1 → v0.6.3 neural LM — matched to the Phase 3.5 text mix (multilingual prose + ham radio). val_ppl 5.626. Dropped from the default prose preset at v0.6.3 because it hurt amateur jargon on literary prose. Still loadable via --lm lm_phase5_2 --fusion-weight 0.7 for research.
rnnt_phase5_9.pt	4.13 M	49	Phase 5.9 failed retrain — strict letter-group densification. Regressed on 6 / 6 LCWO + websdr clips, including its target websdr clip. Kept for reproducibility.
rnnt_phase5_8.pt	4.13 M	49	v0.6.0 / v0.6.1 acoustic — English-literary curriculum. Demoted at v0.6.2 after the real-audio bench lost ~24 % relative mean CER vs rnnt_phase5_5.
rnnt_phase5_7.pt	4.13 M	49	v0.5.3 acoustic — amateur-idiom curriculum (5NN cut-numbers + run-on UR/SK/KN/BK). Kept for diff.
rnnt_phase5_4.pt	4.13 M	49	v0.5.0 acoustic — Phase 5.3 (wider jitter + dash:dot ratio) + Phase 5.4 (30 % real-audio mix). Kept for diff.
rnnt_phase3_5.pt	4.13 M	49	v0.4.0 / v0.4.1 acoustic — synthetic-only. Kept for diff.
lm_phase4_0.pt	4.76 M	46	Legacy v0.1-era LM, 100 % ham-radio text mix. Kept for research / reproducibility; not recommended for fusion.
rnnt_phase3_3.pt	4.13 M	46	v0.3 acoustic model — multilingual ASCII-normalised prose. No accent tokens.
rnnt_phase3_2.pt	4.13 M	46	v0.2 acoustic model — anti-hallucination curriculum, no multilingual data.
rnnt_phase3_0.pt	4.13 M	46	v0.1 acoustic model — AWGN-only training.

Intended use

• Primary: decoding Morse-code audio (8 kHz sample rate, carrier around 600 Hz, 500 Hz RX bandwidth) into text. Targeted at amateur-radio receivers and SDR applications. v0.4 has been live-validated on a real IC-7300 on a mixed-language stream (CQ macros + Verlaine + Poe + Nerval + FAV22 clair) and decodes French prose with diacritics (e.g. L'HEURE, MA SEULE ÉTOILE EST MORTE, À LA TOUR ABOLIE) where v0.3 collapsed apostrophes to 1 and stripped É / À to E / A.
• Secondary (research): a reference implementation for RNN-T and LM-fusion research on a small-vocabulary acoustic task with a fully reproducible synthetic data pipeline.

Limitations

Honest about what v0.6.3 does and does not do:

1. Real-audio coverage is still narrow. The recommended acoustic includes a small real-audio mix from one hand-keyed recording session, but it is not trained on a broad, diverse amateur-radio corpus. More operators, receivers, bands, speeds, and propagation conditions are the main data gap.
2. Only three accented characters are supported (É, À, apostrophe). The 49-token vocabulary covers the French tokens already added in Phase 3.4; other diacritics still fall back to their ASCII base letter via NFKD (e.g. château → CHATEAU).
3. 6-second model context. Training clips are 6 s. Offline and live decoders keep inference on that distribution with sliding 6 s windows and central-zone commits.
4. WPM range. The release models target normal CW speeds around 16-28 WPM. Very high-speed HST / FAV22 audio (84-240 WPM) is out-of-distribution.
5. Pile-ups and overlapping stations are not solved. Multi-source CW at the same or nearby frequency is a source-separation problem, not just an acoustic-decoding problem.
6. LM fusion is offline only. morseformer decode --preset prose can use lm_phase5_2; morseformer live runs acoustic-only even when a preset names an LM.
7. Beam search / callsign priors are experimental. Scaffolding exists in the codebase, but the shipped presets use the validated greedy RNN-T path.

How to use

Offline decode of a .wav file

pip install morseformer
morseformer decode my_recording.wav

# Optional LM shallow fusion for prose / ragchew recordings.
morseformer decode my_recording.wav --preset prose

# Inspect or pre-download model artifacts.
morseformer models list
morseformer models download rnnt_phase5_5

Real-time streaming decode

pip install "morseformer[live]"
morseformer live

Tune your receiver to zero-beat at 600 Hz with a roughly 500 Hz CW filter. Ctrl+C to quit. Latency is about 4 s end-to-end. The shipped live preset uses rnnt_phase5_5.pt, --confidence-threshold 0.6, and --digit-threshold 0.90 to suppress noise-driven false positives and pseudo-numerals.

Training data

Most training audio is synthetic, generated on the fly by the morseformer.data.synthetic package. The current recommended acoustic (rnnt_phase5_5.pt) descends from:

• Phase 3.4 / 3.5 text mix: callsigns, Q-codes, QSO templates, numerics, English words, random characters, multilingual prose, and French prose with É / À / apostrophe preserved.
• Realistic HF channel: AWGN, QSB, QRN, carrier-frequency jitter, drift, QRM, and a 500 Hz receiver bandpass around 600 Hz.
• Human-keying envelope: widened element and gap jitter, dash:dot ratio variation, inter-element gap inflation, and long inter-word-silence inflation.
• Phase 5.4 real-audio mix: a small set of hand-keyed, ground-truth-aligned 6 s chunks mixed into training.
• Phase 5.5 long word gaps: word_gap_inflation ∈ U(1.0, 8.0), which is the acoustic checkpoint now recommended again in v0.6.3.

The French prose corpus comes from Project Gutenberg (data/corpus/prose.txt, gitignored, reproducible via python data/corpus/fetch.py). The FAV22 reference clair blocks (data/corpus/fav22_blocks.jsonl) extracted from the F9TM training PDF supply ~110 k chars of authentic French CW text (3.62 % accent density), used by Phase 3.5+ as additional prose_fr source.

The real-audio component is useful but narrow; expanding it is the highest-value data improvement for future releases.

Evaluation

French accent-rich bench (historical v0.4 headline)

300 French-prose samples sampled around É / À / apostrophe positions (≥ 70 % contain at least one new token), evaluated through scripts/eval_phase_3_4_french.py:

Metric	v0.3 (Phase 3.3, 46-vocab)	v0.4 (Phase 3.5, 49-vocab)
Overall CER	n/a (no accent tokens)	6.46 %
CER at clean / +20 / +10 / +5 dB	n/a	0.00 %
CER at 0 dB	n/a	4.99 %
CER at −5 dB	n/a	33.77 %
É precision / recall	0 % / 0 %	100 % / 91.6 %
À precision / recall	0 % / 0 %	97.8 % / 95.7 %
apostrophe precision / recall	0 % / 0 % (' decoded as 1)	98.4 % / 98.4 %

Realistic-channel SNR ladder (historical v0.3 reference numbers retained)

1200 samples (40 / WPM × 5 WPM × 6 SNR), full Phase 3.1 channel stack. v0.4 is at parity or better with v0.3 across the ladder; the bench was not re-run on Phase 3.5 because the curriculum changes only affect text mix and operator jitter — the channel stack is unchanged.

SNR (dB)	v0.1 (Phase 3.0) CER	v0.2 (Phase 3.2 last) CER	v0.3 (Phase 3.3 best) CER	Δ vs v0.2
+20	0.2421	0.0017	0.0000	−0.17 pp
+10	0.3816	0.0000	0.0000	0.00 pp
+5	0.3361	0.0000	0.0000	0.00 pp
0	0.5000	0.0030	0.0030	0.00 pp
−5	0.7871	0.0542	0.0502	−0.40 pp
−10	1.3316	0.4442	0.3965	−4.77 pp
overall	0.5964	0.0838	0.0749	−0.89 pp

AWGN guard ladder (historical no-channel regression check)

1200 samples, same WPM × SNR grid, AWGN only:

SNR (dB)	v0.1 CER	v0.2 (Phase 3.2 last) CER	v0.3 (Phase 3.3 best) CER	Δ vs v0.2
+20	0.0000	0.0000	0.0000	0.00 pp
+10	0.0000	0.0000	0.0000	0.00 pp
+5	0.0000	0.0000	0.0000	0.00 pp
0	0.0000	0.0229	0.0099	−1.30 pp
−5	0.0407	0.3348	0.2943	−4.05 pp
−10	0.8019	0.8919	0.8826	−0.92 pp
overall	—	0.2083	0.1978	−1.05 pp

The −5 / −10 dB AWGN regression v0.2 introduced (in exchange for the anti-hallucination prior) is partly recovered in v0.3 — the multilingual prose appears to make the prior less brittle on noise-shaped inputs.

False-positive bench (noise-only audio, 150 samples × 3 modes — empty-label target)

metric	v0.1 (Phase 3.1 best)	v0.2 (Phase 3.2 last)	v0.3 (Phase 3.3 best)
Mean characters emitted	11.17	1.05	1.01
Median	11.0	1.0	1.0
Max	21	2	2
% "letter-soup" (>5 chars)	98.7 %	0.0 %	0.0 %

Historical v0.4.1 false-positive grid (Phase 3.5 acoustic, threshold × fusion)

scripts/eval_false_positive.py --candidate-ckpt rnnt_phase3_5 --lm-ckpt lm_phase5_2 --fusion-weight λ_lm --confidence-threshold thr, 50 samples / mode × 3 modes:

confidence_threshold	fusion_weight (λ_lm)	mean chars / noise	max	% > 5 chars
0.0	0.0	1.50	2	0.0 %
0.0	0.7	1.01	2	0.0 %
0.6	0.0	0.17	1	0.0 %
0.6	0.7	0.17	1	0.0 %

Threshold 0.6 (the v0.4.1 streaming default) cuts the FP mean by ≈ 90 %. Adding fusion λ = 0.7 on top does not regress FP because gating is applied to the acoustic head pre-LM — the LM cannot rescue a low-confidence noise emission. Threshold and fusion stack cleanly.

Historical v0.4.1 LM rescoring on prose audio (Alice in Wonderland CW, ground-truth aligned)

scripts/eval_fusion_realaudio.py --rnnt-ckpt phase3_5 --lm-ckpt lm_phase5_2 --jsonl all_alice.jsonl, n = 120, score-filter ≥ 0.7 to focus on chunks where the baseline still has CER headroom:

λ_lm (shallow fusion)	CER	Δ vs baseline
0.0 (baseline, greedy)	36.44 %	—
0.3	35.81 %	−1.7 %
0.5	33.46 %	−8.2 %
0.7	32.30 %	−11.4 % ← peak
1.0	33.66 %	−7.6 %

The Alice corpus is ebook2cw renderings of public-domain English prose (chapters 1–12 of Alice in Wonderland), char-aligned by scripts/align_ebook_cw.py against the source text — a corpus that was prepared for Phase 3.7 real-audio finetuning and re-purposed here as v0.4.1's autonomous prose bench. The 9482 chunks are deterministic ebook2cw audio, not real-radio captures, but they exercise the model on real linguistic structure (story narrative) where the v0.4.0 acoustic still has headroom and the LM can contribute.

ILME (density-ratio fusion, λ_ilm > 0) is catastrophically harmful on this distribution: λ_lm = λ_ilm = 0.2 → CER +18 %, λ_lm = λ_ilm = 0.3 → CER +83 %. The fusion code keeps ILME for research / completeness; it is not the recommended path.

Why the legacy LM (lm_phase4_0) does not help

PPL of lm_phase4_0.pt (the v0.1 LM, trained on DEFAULT_MIX = 100 % ham radio) on representative texts:

text	PPL
Alice prose ("ALICE WAS BEGINNING TO GET VERY TIRED OF SITTING …")	99.5
Alice prose ("DOWN THE RABBIT HOLE INTO A LARGE RABBIT HOLE")	213.9
QSO line ("CQ CQ CQ DE F4ABC F4ABC PSE K")	8.9
RST exchange ("UR RST 599 599 NAME TOM TOM 73")	15.4

The legacy LM is calibrated for ham radio and is hostile to general prose decoding — fusion with it on Alice audio actively hurts CER (33.86 % → 35.79 % at λ_lm = 0.2 in our re-bench), confirming the Phase 4.1 null-result was a text-distribution mismatch and not a fusion-method problem. lm_phase5_2.pt was trained on the same PHASE_3_4_MIX the Phase 3.5 acoustic saw, which makes shallow fusion finally useful.

Live validation

Selected live-radio tests on a real IC-7300 driving the streaming decoder:

• 2026-04-27 (v0.3 baseline): mixed CQ macros + Verlaine Chanson d'automne + Poe Annabel Lee + Apollinaire Le Pont Mirabeau. French rendered word-by-word but apostrophes appeared as 1 and É / À as plain E / A.
• 2026-04-29 (v0.4): same setup with FAV22 clair + Nerval El Desdichado + Poe Annabel Lee. French accents are now decoded natively: L'HEURE, L'AUTOMNE, MA SEULE ÉTOILE EST MORTE, À LA TOUR ABOLIE. Apostrophes captured cleanly. No false-positive É / À emissions in the English passages.

Residual failure modes observed across later releases include hard callsigns under noise and jitter, occasional prosign over-emission on pile-ups, and overlapping stations at the same frequency.

Environmental impact

Training budget through v0.6.x is small by modern ASR standards: single-GPU fine-tunes on an RTX 3060 Laptop, typically a few hours per phase. The current recommended acoustic is still a compact 4.13 M-param model intended for CPU inference.

Technical specifications

Acoustic model (rnnt_phase5_5.pt, 4.13 M params):

• Encoder: Conformer d=144, L=8, H=4, ff_expansion=4, conv_kernel=31, RoPE, LayerNorm conv, 4× subsample
• CTC head: Linear(144 → 49)
• PredictionNetwork: Embedding(49, 128) + LSTM(128, 128, 1 layer)
• JointNetwork: Linear(144 → 256) + Linear(128 → 256) + tanh + Linear(256 → 49)

Language model (lm_phase5_2.pt, 4.76 M params): decoder-only GPT (RMSNorm + SwiGLU + RoPE + tied embeddings + causal SDPA), d=256, L=6, H=4, dropout=0.1, 49-vocab. Trained 20 k steps on PHASE_3_4_MIX, AdamW peak_lr 3e-4, batch 128, bf16, EMA 0.999, val_ppl 5.626. Pair with the RNN-T at λ = 0.7 for offline prose decoding. The legacy lm_phase4_0.pt (46-vocab, 100 % ham-radio mix, val_ppl 3.75 on its own distribution) is kept for research but is not the recommended fusion partner.

Vocabulary: 49 tokens — blank (index 0), 26 uppercase letters A–Z, 10 digits 0–9, 9 punctuation / Morse prosigns (., ,, ?, !, /, -, =, +, space), plus É, À, '.

Front-end: complex bandpass around the carrier (default 600 Hz ± 250 Hz), magnitude, 4× downsample to a 500 Hz frame rate, scalar normalization. Input shape [B, T, 1].

Citation

@software{morseformer_v0_6_3_2026,
  author       = {Derhy, Serge},
  title        = {morseformer: open-source transformer-based Morse / CW decoder},
  year         = 2026,
  version      = {v0.6.3},
  url          = {https://github.com/sderhy/morseformer},
  howpublished = {\url{https://huggingface.co/sderhy/morseformer}},
}

Acknowledgements

• Sébastien Derhy — design, engineering, and on-air validation of morseformer.
• Mauri Niininen (AG1LE) — pioneering ML-based CW decoding work.
• Alex Shovkoplyas (VE3NEA) — CW Skimmer, the commercial reference.
• Andrej Karpathy — nanoGPT, the aesthetic reference for the language model.
• Project Gutenberg — public-domain literary texts in English, French, German, and Spanish used to build the multilingual prose corpus.
• F9TM (Centre de Formation Mémorial du REF) — FAV22 reference training material, used to source authentic French CW prose for the Phase 3.5 prose_fr sampler.
• The amateur-radio community — decades of publicly available CW recordings and transcripts.

73 de morseformer.