Skip to main content

Posts

How Verified Real-World KiwiSDR Data Cut NeuroCW's Error Rate Below ft6 (Training Report)

Training-results report. This post covers training and benchmarking only. Testing, ONNX conversion and the application come next; no new NeuroCW release is available yet . Earlier parts: Part 1 · Part 2 Figure 1. Headline result. CER on real on-air W1AW bulletins with real band noise added (test, 384 clips). CER: ↓ lower is better. Error bars are 95% bootstrap CIs; ★ marks the best model. TL;DR 18.8% CER ↓ on real noisy CW (test, 384 clips). ft6: 29.1%, DeepCW: 28.8%. Lower is better. −4.07 pp ↓ from the verified real-data fine-tune alone (paired bootstrap, p < 0.001). Lower is better. 126× real time ↑ CPU speed of the 5.25M-parameter Lite model on 2 threads. Higher is better. 333 h → 4.1 h receiver audio cut down to verified ground truth (≈1.2% yield), over ≈356 GPU-hours of work Call for data. The fine-tune in this report used English CW only. I am collecting more real CW training data in English, Korean ...
Recent posts

NeuroDMR Monitor: DMR Monitoring with AI — Live Subtitles & Translation

NeuroDMR Monitor listening to BrandMeister TG 91 (World-wide): the decoded audio as a spectrum and waterfall, with live English subtitles translated into Japanese. TEST BUILD — A work-in-progress demo by Ian Lee / KD8CEC. Features and settings may change. Executable release planned for this week — The Windows build of NeuroDMR Monitor will be posted here at hamskey.com under the NeuroDMR label . No download link yet. NeuroDMR Monitor: DMR Monitoring with AI — Live Subtitles & Translation NeuroDMR Monitor is a desktop app I am building to read DMR traffic as well as hear it. It listens to BrandMeister talkgroups, turns each transmission into text with speech recognition, and translates it into the language of your choice — live, while the operators are still talking. In this recording I hop across talkgroups in Germany, the world-wide TG 91, Malaysia and France, and change the target language several times along the way. How it fits together The pipeline as configured in t...

Nemotron 30B on Radio: Why I Still Prefer Whisper

A real snapshot from the test: radio captions alongside the GPU process monitor. TEST BUILD — A preliminary field test by Ian Lee / KD8CEC, not a controlled CPU-versus-GPU benchmark. Nemotron 30B on Radio: Why I Still Prefer Whisper A larger model is not automatically the better fit for a live radio-captioning workflow. This recording explores a backend exposed to NeuroCaption as nemotron-30b : useful English captions, uneven language handling, and substantial GPU-accessible memory use. What was actually configured? The visible alias is nemotron-30b. Auto language, translation off, four-second clips and gate off are shown. The exact checkpoint is not shown. The backend prompt and complete runtime configuration are unverified. The client tab labeled Whisper does not establish which model is running behind the server alias. A substantial memory footprint in this setup The GB10 snapshot above lists VLLM::EngineCore at 48,341 MiB — approximately 47.2 GiB of GPU-accessible proces...

Preliminary Study: Speech-to-Text on Shortwave AM Time Stations (WWV/WWVH) with a Gold Answer Key

Preliminary study, shortwave broadcast AM only. 8 clean clips (536 words). Rankings are not definitive: the top five confidence intervals overlap. Gold-key collection continues until November 9, and this post will be updated. In my first preliminary test , the answer key came from two Whisper models, so it could be wrong in places. This time the key comes from broadcasts whose words are known in advance: the WWV and WWVH time stations. Their announcements follow a published format, and the space-weather bulletin text comes from NOAA. That gives a fair answer key without anyone having to guess. Method 42 WWV/WWVH AM recordings captured through public web SDRs. Expected-content key for each clip, built from the published broadcast format and NOAA bulletins. Only 8 clips have clear enough voice to pass the quality filter. They are the headline set. WER with 95% clip-bootstrap confidence intervals (10,000 resamples). Result 1: accuracy on clean clips (preliminary) Engine ...

NeuroCW Japanese Test: Same Wabun Morse, Two Results - Auto vs Manual Language

TEST BUILD. This is an early, in-development build of NeuroCW, not a finished product. The decoded text below is shown exactly as it appeared on screen, including spacing quirks, odd spellings and decoding errors. Watch the 7:11 test. Run 1 (Auto) runs to 3:56; Run 2 (same file, language set to JA) starts at 3:58. 1. The test, in one line What happens when the same Japanese Morse recording is decoded twice — once with NeuroCW guessing the language ( Auto ), and once with the language set to Japanese by hand ( JA )? Japanese uses its own Morse table, Wabun code, in which many dot-dash patterns stand for kana instead of Latin letters. The same sounds can therefore mean two different things, and the decoder has to know which table to use. Figure 1. The input is a prepared test recording, ja_qso_1.wav , with two Japanese QSOs at once. Multi mode opens track A at 450 Hz and track B at 700 Hz. 2. Run 1: Auto Both tracks started on Auto . The Latin parts of each QSO (C...

NeuroCW Korean Test: Real-Time Hangul Composition from Morse, Two Languages at Once

TEST BUILD. This is an early, in-development build of NeuroCW, not a finished product. The decoded text below is shown exactly as it appeared on screen, including spacing quirks, odd spellings and decoding errors. Watch the 5:13 test. Korean decoding begins around 1:05; both panels are full from about 2:45. 1. The test, in one line Can NeuroCW read Korean Morse code and turn it, in real time, into proper Hangul text? Korean Morse sends individual letters (jamo), but written Korean stacks those letters into syllable blocks. This test checks that composition step, called moa-sseugi (모아쓰기), while a second, English-language signal is decoded at the same time. Figure 1. The input is a prepared test recording, long_qso_test1.wav , containing two CW conversations at once. 2. Two signals, split automatically In Multi mode NeuroCW found both signals by itself and opened one track for each: track A at 650 Hz and track B at 850 Hz. Both tracks were left on language mode Auto . ...

Preliminary Test: Small Local Speech-to-Text on Real HF Radio (Shortwave Broadcast and Ham SSB)

Speech-to-text models are usually measured on clean podcasts and audiobooks. HF radio is different: fading, QRM, narrow SSB passbands, net-style exchanges and low SNR. I wanted a plain answer to one question — how do small, GPU-friendly ASR models handle real KiwiSDR audio? Preliminary test, not a final benchmark. This is a first validation pass of speech-to-text for shortwave broadcast and amateur radio use. The reference transcripts were built by cross-checking each broadcast with two Whisper models, so they may not be fully accurate. For a more rigorous check, I will record shortwave broadcasts whose content (script or transcript) is published, and run a second comparison against that known text. Dataset Kind Clips Notes English shortwave broadcast (SWBC) 14 WRMI / WINB / WWCR English segments, KiwiSDR captures English ham SSB 11 VK40 net ~7065 kHz (Australian KiwiSDR), audible speech only Total scored 25 Pure-noise / hallucination-only clips dropped Audio was resamp...

NeuroCW On-Air Test: Reading Weak, Fading CW Off a KiwiSDR (Test Build)

TEST BUILD. This is an early, in-development model (test build), not a finished product. Everything below is a real experiment with honest mistakes shown as they happened. Results are a snapshot, not a benchmark. Watch the ~5:36 test. The interesting part starts around 4:18 on the right-hand track. 1. The test, in one line I pointed NeuroCW at a live web receiver and fed it weak, fading signals on 40m. The question was simple: when a signal nearly disappears into the noise, can the model still read the callsign a human ear struggles to hear? Figure 1. KiwiSDR in Chichester, UK, tuned to the 40m CW band. The audio was routed into NeuroCW through a PC loopback. 2. Setup Item Detail Receiver KiwiSDR, Chichester, UK (public web SDR) Band / mode 40m, CW mode selected Audio path Browser audio → PC loopback → NeuroCW Model Test build / early model (not a released version) Signal level App showed SNR fluctuating roughly from the mid-teens to high-20s dB, with d...