Figure 1. Everything NeuroCW learned came from the bottom panel: a simulated channel. Now it needs real signals.
1. Why real data
In Part 1, NeuroCW scored 5.07% CER on synthetic tests, but its accuracy on real on-air signals hasn't been measured. Just two real recordings were enough to expose a blind spot:
Figure 2. On a synthetic heavy-weight keying test, one targeted fine-tune cut the error from 13.4% to 0.1%. Verified real recordings would turn findings like this into a benchmark.
2. What we need
Figure 3. Checklist for one submission.
3. How to contribute
Figure 4. Four steps. Several recordings per email is fine, and share links work for large files.
Transcript rules: write exactly what was sent, including the sender's mistakes. Put prosigns in angle brackets (<AR> <SK>). For Korean or Japanese, write Hangul or kana. If you are unsure of a part, trim it; never guess.
Figure 5. Paste this template into your email. The values are illustrative.
4. FAQ
- Sample rate? Send your recorder's native file, with no re-encoding. It should be fine.
- Other callsigns in the audio? They're fine since they were sent on air. You can also mask them (XXXX).
- Credit? Optional. Give your callsign or stay anonymous. Contributors are credited by callsign unless they ask to stay anonymous.
- Will the data be public? Yes. The recordings and transcripts will be released as an open real-air CW benchmark dataset, and every measurement result on it will be published.
By submitting, you confirm you made the recording and allow its use for training and evaluating NeuroCW, and its public release as part of the open dataset under an open license (details announced with the release). Please send amateur-band signals only.
73, Ian Lee / KD8CEC
Comments
Post a Comment