Training Transcripts
Human-reviewed dispatch audio paired with corrected transcripts — the dataset that teaches ThinLine Radio to actually understand scanner traffic. Free to browse, play, and download.
What is this?
Speech-to-text models are trained on podcasts and audiobooks — not on clipped, noisy, jargon-heavy public safety radio.
Off-the-shelf transcription mangles street names, unit designators, and dispatch codes because it has never heard them. ThinLine Radio fixes that by training on the real thing: actual dispatch audio, with text a human already verified.
Every clip here started as a scanner call that an operator listened to and corrected by hand. Those verified audio-and-text pairs are exactly what a speech model needs in order to learn the vocabulary and cadence of radio traffic — and the corrections feed straight back into better automatic transcription for everyone on the platform.
How a clip gets here
Nothing is scraped or auto-collected. A person reviews and approves every single sample.
1. Call is recorded
A ThinLine Radio server ingests a dispatch call and produces an automatic first-pass transcript.
2. A human corrects it
The server operator listens on their own /transcripts page and fixes the text to match exactly what was said.
3. Approved & contributed
On approval the clip is converted to 16 kHz mono WAV and uploaded to the shared collector.
4. The model improves
Verified pairs fine-tune the speech model, lifting transcription accuracy across every server.
What's in a download
Take the whole dataset or just the newest few hours — it arrives in the layout Whisper and Hugging Face fine-tuning pipelines already expect.
| Item | Description |
|---|---|
manifest.jsonl | One JSON object per line: {"audio": "...", "text": "..."} |
audio/ | 16 kHz mono 16-bit PCM WAV clips |
transcripts/ | Matching lowercase transcript text files |
README.txt | Sample count, total audio duration, and format notes |
export-meta.json | Machine-readable summary of the export |
Source stays anonymous
Exports and the public browser deliberately omit which scanner, server, agency, or reviewer a clip came from. You get the audio and the words — nothing that identifies the feed behind them.
Already-public traffic
The dataset contains only radio traffic that was broadcast unencrypted over public safety airwaves — the same audio anyone with a scanner could hear live.
Contribute from your server
Run a ThinLine Radio server? Open /transcripts in your admin panel, connect it to the collector, and start approving corrections. Every clip helps the whole community.