Transcription and CW decoding
The player can turn recordings into text on your own computer, offline:
| Recording mode | Engine | Download |
|---|---|---|
| Voice: LSB, USB, AM, FM, WFM, DV | OpenAI's Whisper speech model | Once, the first time it's needed |
| CW, CW-R | DeepCW neural Morse decoder | None: ships with the app |
| RTTY and data modes | Skipped |
Transcription runs in a separate background process, so playback stays smooth while it works.
Transcribing#
- TXT on the play screen (next to ☆): while it's lit, each recording is transcribed the first time it plays. The text appears as a two-line caption under the frequency and follows the audio. Click the caption to see the whole transcript; click a line there to jump to it.
- QUICK › Transcribe This File: just this one.
- QUICK › Transcribe All in This List: every recording in the folder, search results or bookmarks, in the background. Progress shows as ✎ n/total in the status bar. Stop Transcribing cancels it.
- ✎ TRANSCRIBE ALL on the selected row of the date-folder list: queues every recording in that folder that has no transcript yet, without opening the folder.
If the app is closed while a Transcribe All is running, it carries on at the next start with the recordings still to do ("Continuing the transcription — n left"). Stop Transcribing cancels it for good.
Transcribed recordings show a cyan T in lists. Bookmark exports include the transcript.
Other Quick menu items: Show Transcript…, Re-transcribe This File, Delete Transcript.
Voice: Whisper#
QUICK › Speech Model chooses the model:
| Model | Size | Use it for |
|---|---|---|
| SMALL | ~250 MB | Best accuracy (default) |
| BASE | ~80 MB | Faster, on slower computers |
The model is downloaded once from Hugging Face the first time it's needed, then loaded from disk.
What to expect on HF SSB:
- Phonetics ("November", "Bravo") and common phrases come through well.
- Call signs are often only partly right. The call-sign finder works hard to repair them.
- Audio is band-passed and normalised first, and the sound-effect captions and repeated lines Whisper invents from noise are removed.
Call-Sign Boost (on by default; QUICK or CALL SIGNS › LISTS) nudges Whisper towards the phonetic alphabet and spoken digits, which it otherwise often hears as ordinary words on noisy SSB. The nudge is mild on purpose: stronger settings made Whisper invent phonetics. It applies to new transcripts.
CW: DeepCW#
CW recordings are decoded by DeepCW, a neural Morse decoder by e04 (MIT licence). Everything above works the same for CW: TXT, the caption, Transcribe This File, ✎ TRANSCRIBE ALL, the database, and the CALL SIGNS screen.
- The model (15 MB) is part of the app. Nothing is downloaded, and Whisper isn't needed.
- The player finds the CW tone (the radio's CW pitch) by itself. DeepCW listens between 400 and 1200 Hz, so a pitch outside 450–1150 Hz is shifted to 700 Hz first. The tone is shown with the transcript, e.g. DeepCW · 600 Hz tone.
- Characters the decoder is unsure of (noise, speech) are dropped. A recording with no CW shows No Morse decoded.
- Only a call sent as one word counts as a call sign:
DE K6ABC Kgives K6ABC, never K6ABCK. - Speed: about a second per minute of CW on a typical PC.
Transcript database#
Every transcript is also kept in a SQLite database, transcripts.sqlite, in the app's data folder. QUICK › Transcript Database… shows what's in it and the most-heard call signs, and opens the folder.
It's a standard SQLite 3 file, so DB Browser for SQLite or the sqlite3 command line can open it.
- The JSON files in
transcripts\are the master copy. At start-up the database picks up new or changed transcripts and drops deleted ones, so deletingtranscripts.sqlitesimply rebuilds it. - Tables:
recordings(one row per transcribed file: radio, file, path, date and time, frequency, mode, RX/TX, S-meter, length, engine, full text),segments(each timed line),segments_fts(full-text index of the lines;docid=segments.id) andcallsigns.
Example queries
-- lines that mention "antenna", with the time in the recording
SELECT r.file_name, r.recorded_at, r.freq, s.start_s, s.text
FROM segments_fts f
JOIN segments s ON s.id = f.docid
JOIN recordings r ON r.id = s.recording_id
WHERE segments_fts MATCH 'antenna'
ORDER BY r.recorded_at;
-- every recording where a call sign starting K6 was heard
SELECT c.callsign, r.file_name, r.recorded_at, r.freq, r.s_level
FROM callsigns c
JOIN recordings r ON r.id = c.recording_id
WHERE c.callsign LIKE 'K6%';
Models, transcripts and the database live on your computer. Nothing is written to the SD card, and no audio leaves your computer.