v7: standalone Sing Detect app — full Tauri project tree

Promotes the v6 detection slice into a complete, buildable Tauri 2 app and
renames Stream Studio → Sing Detect (detection-only; video-concat split to its
own project). Proper src/ + src-tauri/ layout, Python sidecar under sidecar/,
app icons, package/Cargo manifests, and dev/build/pack scripts.

Pipeline unchanged in spirit: Rust drives ffmpeg to a 16 kHz mono WAV, the
Python sidecar (inaSpeechSegmenter) emits line-JSON segments, and Rust writes
EDL / markers CSV / CSV / JSON exports. Long ops run on spawn_blocking; cancel
kills the whole py -3.10 process tree.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-06-26 23:12:25 +08:00
parent 765816e701
commit 4887981a2d
51 changed files with 8250 additions and 52 deletions

106
README.md
View File

@@ -1,41 +1,85 @@
# sing_detect
# Sing Detect
Singing/music-segment detection for long live-stream recordings: find the
segments where someone is singing and export markers (CSV / EDL / JSON) for fast
editing in DaVinci Resolve.
A fast, native-feeling desktop app that **finds singing segments in a stream
recording** and exports editor-friendly markers (DaVinci Resolve EDL / markers
CSV / CSV / JSON). Built with Tauri 2 (Rust backend + TypeScript UI) and a small
Python ML sidecar.
**This branch is the latest version (v6).** A plain `git pull` gives you only the
current code — no old versions clutter the working tree. The earlier iterations
live in the git history and are tagged, so you can inspect or check out any of
them:
This is the detection-only spin-off of Stream Studio (the video-concat feature
lives in its own project now).
```bash
git checkout v3 # or v1 … v6
git checkout - # back to latest
## How it works
```
recording ──▶ [Rust ▸ ffmpeg ▸ 16 kHz mono WAV]
[Python sidecar ▸ inaSpeechSegmenter] ──json──▶ segments
[Rust ▸ EDL / markers / CSV / JSON exports]
```
## Evolution
| Layer | Tech | Why |
|-------|------|-----|
| UI | **WebView2** via **Tauri 2** | GPU-composited, tiny app, not Electron-sized. |
| Core | **Rust** | Audio extraction, drives the sidecar, writes exports. No console windows. |
| Audio | **ffmpeg / ffprobe** | 16 kHz mono WAV extraction. |
| ML | **Python 3.10 sidecar** | `inaSpeechSegmenter` (TensorFlow) classifies singing as "music"; TensorFlow needs Python 3.10. |
The history is ordered by **technical & functional evolution** (not file
timestamp). Each tag is a complete, standalone version at the repository root.
## Prerequisites
| Tag | Engine | Interface | Leap |
|-----|--------|-----------|------|
| `v1` | HuggingFace `matthijs/svd` (torch.jit) | tkinter file picker | First prototype — single file, proves the idea |
| `v2` | `transformers` pipeline + pyannote | CustomTkinter GUI | Overlapping-chunk inference, CSV timecode export |
| `v3` | inaSpeechSegmenter | CLI (argparse) | Robust pipeline: gap-merge, min-duration, EDL+CSV |
| `v4` | inaSpeechSegmenter | CLI batch | Process whole folders, `--recursive`, `--auto-cut` |
| `v5` | inaSpeechSegmenter | CustomTkinter multi-tab app | Modular `core/` engine + UI tab, JSON + ffmpeg-script export |
| `v6` | inaSpeechSegmenter (Python sidecar) | Tauri (Rust + TS) | Cross-platform rewrite; Python reduced to a thin ML sidecar emitting line-JSON |
- **Node** (18+) and **Rust** (`rustup`, MSVC toolchain) — to build.
- **ffmpeg + ffprobe** on PATH (or beside the exe) — bundled by `pack.bat`.
- **Python 3.10** on the machine that runs detection. The app's built-in
**"Test detection engine"** button installs the ML packages + downloads the
model for you on first run (or run `sidecar\setup_sidecar.bat` manually).
Engine evolution: `svd model``transformers pipeline`
**`inaSpeechSegmenter`** (settled from v3 on, which classifies singing voice as
"music").
## Run / build
## Notes
- **v4** carries a copy of `detect_singing.py` because `batch_detect.py` imports it at runtime.
- **v5** keeps the original package layout (`core/`, `ui/`, `utils/`) so the relative imports resolve; it was extracted from the larger `stream_tools` app.
- **v6** is the detection slice only (`sidecar_detect.py`, `detect.rs`, `detect.ts`), extracted from the `stream-studio` Tauri app — `detect.ts` still references the app's `../dom`, `../ui`, `../api` modules.
```bat
dev.bat REM hot-reload dev (UI + Rust)
build.bat REM release MSI installer → src-tauri\target\release\bundle\msi\
pack.bat REM STANDALONE portable folder → dist-portable\SingDetect\
```
Excluded: `song_detection_cn.py` (an ACRCloud-style fingerprint *song-identification*
client) — it names a track rather than detecting singing segments.
`pack.bat` produces a copy-paste folder with `Sing Detect.exe`, ffmpeg, and the
Python `sidecar/` bundled. Copy it to any Windows PC; the only thing the target
needs is Python 3.10 (the in-app engine test handles the rest).
Detection auto-discovers `py -3.10`; override the interpreter with the
`STREAM_STUDIO_PYTHON` env var, or the script path with `STREAM_STUDIO_SIDECAR`.
## Project layout
```
sing_detect_v2/
├── index.html
├── src/ # Frontend (TypeScript, no framework)
│ ├── main.ts # shell: single view, theme, drag-drop
│ ├── api.ts # typed bridge to Rust commands + events
│ ├── dom.ts / ui.ts # tiny DOM + toast/progress/song helpers
│ ├── styles.css # Fluent-ish design system (light/dark)
│ └── views/detect.ts # the detection view + engine setup card
├── src-tauri/
│ ├── src/
│ │ ├── lib.rs # commands, state, app wiring
│ │ ├── tools.rs # ffmpeg/ffprobe discovery + helpers
│ │ ├── detect.rs # drives the Python sidecar; env check/setup
│ │ └── export.rs # EDL / markers / CSV / JSON
│ ├── tauri.conf.json # window + bundles sidecar/ as a resource
│ └── capabilities/ # dialog + opener permissions
└── sidecar/
├── detect.py # the ONLY Python — ML step only
├── requirements.txt
└── setup_sidecar.bat
```
## Detection parameters
| Parameter | Default | When to change |
|-----------|---------|----------------|
| Merge gap | 10s | Raise if the singer pauses >10s mid-song |
| Min duration | 20s | Lower for short covers/snippets |
| Padding | 2s | Raise if songs get clipped at the edges |
| FPS | 29.97 | Match your video (25 PAL, 23.976 film) — affects timecode export only |
Singing is classified as **"music"** by inaSpeechSegmenter; talking is "speech".