Offline 2m-band monitor. Captures the entire amateur 2m band (144–148 MHz) from an SDRPlay RSP1, detects and demodulates live NBFM transmissions across all channels in parallel, transcribes them locally, archives everything searchable, identifies speakers, and detects and summarizes nets — with no cloud dependencies.
See plan.md for the roadmap and design rationale, and CLAUDE.md for project rules.
RSP1 ──IQ (~6 MSPS)──► Capture Daemon (C#)
│ polyphase filterbank channelizer (~300 ch)
│ per-channel adaptive squelch + NBFM demod
│ segmentation markers (squelch crash, courtesy
│ tone, RF edges, discriminator DC-offset jumps)
│
├──► Opus (OGG) clips on disk
└──► metadata/events ──► Web Host (ASP.NET Core)
│ EF Core → SQLite (WAL)
│ ** single writer **
│ job queue API
│ Vue 3 dashboard
▼
Processing Workers (C#) ◄── claim jobs via host API
│ transcription — Whisper.net (whisper.cpp)
│ speaker embed — ECAPA-TDNN via ONNX Runtime
│ net analysis — deterministic heuristics
│ net summaries — LLamaSharp (llama.cpp, 8B Q4)
└──► results posted back to host API
| Component | Minimum | Recommended | Notes |
|---|---|---|---|
| SDR | SDRPlay RSP1 | RSP1 + 2m bandpass filter | RSP1 has minimal preselection; FM broadcast/pagers can overload it. Run ~6 MSPS (2m band is only 4 MHz wide). |
| CPU | 6 cores with AVX2 | 8+ cores (Zen 3+/12th-gen Intel+) | LLM inference is memory-bandwidth bound — dual-channel RAM matters more than core count; >8 threads gives diminishing returns. CPUs without AVX (e.g. QEMU default vCPUs — use -cpu host!) work via bundled no-AVX runtimes but inference is several times slower. |
| RAM | 16 GB (tight) | 32 GB | See per-model budget below. |
| Disk | 50 GB | 100 GB+ SSD | Audio recorded directly as 16 kHz Opus-in-OGG, 32 kbps VBR voice mode (~14 MB per hour of recorded audio). Optional raw WAV/IQ capture behind a debug flag for DSP development. |
| USB | USB 2.0 port | — | 6 MSPS fits USB 2.0. |
| OS | Linux, .NET 8+ | — | Requires SDRPlay API service (proprietary, free download). |
| Workload | Model | RAM | CPU cost | Latency expectation |
|---|---|---|---|---|
| Channelizer + demod | — (custom PFB) | < 1 GB | 1–2 cores, continuous | Real-time |
| Transcription | Whisper small.en (q5) | ~1 GB | Bursty | Faster than real-time on 8 cores; 1.12× real-time on 3 threads of a 4-core box (measured) |
| Transcription (higher accuracy) | Whisper medium.en (q5) | ~2.5 GB | Bursty | ~Real-time; backlog drains overnight |
| Speaker embeddings | ECAPA-TDNN (ONNX) | ~100 MB | Negligible | Per-clip, milliseconds |
| Net summaries | Qwen 2.5 1.5B Q4_K_M (default) | ~1 GB | Batch, after net closes | Under a minute per summary on four cores |
| Net summaries (better prose) | Qwen 2.5 / Llama 3.x 7–8B Q4_K_M | ~5 GB model, 6–8 GB process | Batch | 5–15 tok/s → 2–4 min per net summary; wants 8 cores |
Everything above runs concurrently on one 8-core / 32 GB machine. Four cores is enough for the
defaults, but leaves no headroom: give inference cores − 1 threads (Settings → CPU threads) so the
capture daemon, which consumes a 6.4 MSPS stream in real time, is never descheduled. Summary quality
holds up at 1.5B because the facts come from the deterministic layer — the model only phrases them.
All inference is local; nothing requires network access after model download.
| Layer | Tech |
|---|---|
| DSP / capture | C# (.NET 8), sdrplay_api P/Invoke, System.Runtime.Intrinsics |
| Audio encoding | Opus in OGG via Concentus (managed Opus, native 16 kHz speech mode) |
| Transcription | Whisper.net (whisper.cpp bindings) |
| Speaker ID | ONNX Runtime + ECAPA-TDNN embeddings, clustered per session, labeled via extracted callsigns |
| Summaries | LLamaSharp (llama.cpp bindings) |
| Storage | SQLite (WAL mode) via EF Core; FTS5 transcript index (see CLAUDE.md for the pattern) |
| API / UI | ASP.NET Core (controllers, api/v0/) + Vue 3 + Vuetify |
US amateur transmissions carry no expectation of privacy (Part 97; encryption prohibited), so recording, transcribing, and publishing this traffic is permitted.