Nightingale karaoke scoring: what the first launch downloads, and what it refuses to do
Machine learning powered Karaoke app (with scores!)
At a glance
- What is it?
- Nightingale takes a folder, Plex, Jellyfin or Navidrome library, separates the vocals, aligns the lyrics and scores your singing. The interesting part is not the singing, it is the toolchain the app pulls onto your machine before it can start, and the four experimental backends it hides behind one fallback.
- Who is it for?
- Nightingale earns a trial if your music already lives in Plex, Jellyfin, Navidrome or a folder and you want star ratings without a subscription karaoke service. Check three things before you commit: how much model weight the first launch pulls and which folders it lands in, whether your hardware has a CUDA, Apple Silicon or MPS path or will run everything through WhisperX on CPU, and whether you can live with a Linux build you update by hand.
- Can I use it commercially?
- Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 20, 2026, and from our analysis. They are not legal advice.
Star history
- rzru/nightingale1,559 ★ · +104
| Repository | 2026-09-10 | 2026-11-20 |
|---|---|---|
| rzru/nightingale | 1.5k | 1.6k |
104 Stars in the last 71 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
First launch downloads ffmpeg, uv, Python and PyTorch
The headline promise is a single binary with no manual installation of Python, ffmpeg or ML models, because everything is downloaded and bootstrapped automatically on first launch. Setup asks you for a data folder, then installs the Python environment and the ML models without further input. The self-contained bullet is more specific: ffmpeg, uv, Python, PyTorch and the ML packages are all fetched during setup, and the video backgrounds are pre-downloaded so the first session has something to show. What sits on disk after that run is an executable, a Python environment, model weights for separation and recognition, and cached stock video, which is a different shape from the single binary the project advertises. Setup also fixes the folder layout for good: the main data folder is chosen once, and cache, models, videos and vendor tools can each be redirected later from Settings.
A TypeScript label over a four member Rust workspace with an all deny lint table
The repository is catalogued as TypeScript, but the build is a Cargo workspace whose members are app-core, client/src-tauri, client/src-server and xtask, pinned by a rust-toolchain.toml and formatted by a rustfmt.toml, with a deny.toml sitting alongside for dependency and advisory policy. The workspace lint table denies unsafe_code, future_incompatible, rust_2018_idioms, nonstandard_style, unreachable_pub, unused_lifetimes and unused_qualifications, then every clippy group from all through correctness and perf. Three entries matter more here than the rest for an app whose actual work happens in a Python sidecar: unwrap_used, todo and unimplemented are all deny, joined by print_stdout, print_stderr and dbg_macro. Exactly one line in that table is not a deny. multiple_crate_versions is allow, so duplicated crate versions across the four members are tolerated on purpose.
Four experimental surfaces, all falling back to WhisperX
Recognition is pluggable in two layers, and both defaults are narrow. Whisper is the default ASR engine for broad language coverage; Parakeet v3 is offered for about 25 European languages and marked experimental, running through NeMo on CUDA and ONNX Runtime everywhere else. Forced alignment defaults to WhisperX's aligner, with an experimental GPU path built on torchaudio `forced_align` that only applies on CUDA and Apple Silicon, plus an experimental Qwen aligner (Qwen3-ForcedAligner-0.6B) that timestamps 11 languages including CJK in one pass on CUDA, MPS or CPU. Both experimental alignment backends fall back to WhisperX automatically. A fourth experimental surface sits outside the model stack: UltraStar Deluxe songs take their pitch and lyric data straight from the files. You drop a song folder holding a `.txt` or `.usdx` file plus sibling audio, vocals, instrumental and video, and no analyzer pass runs at all.
Analysis is skipped or replaced in three different ways
A song's backing track does not come out of one fixed pipeline. Stem separation runs the UVR Karaoke model by default, or Demucs, with an adjustable guide vocal volume, and the karaoke model is the one that keeps backing vocals inside the instrumental track for a more natural sound. But if a song carries timed LRC lyrics, that timing is used as it stands and stem separation can be skipped so you sing over the original mix. Plain untimed lyrics go through alignment instead. UltraStar Deluxe folders bypass the analyzer entirely. The lyrics themselves arrive from WhisperX transcription with word level alignment, or are fetched from LRCLIB when a match exists, and a `.lrc` or `.elrc` file sitting beside the source audio is loaded directly from the song's Actions button. For Japanese, Chinese, Cantonese and Korean tracks, alignment runs per character and a romanized reading sits above each token: Hepburn, pinyin, Jyutping and Revised Romanization respectively.
One shared cache folder, and Rescan library adopts the other machine's work
Storage is deliberately splittable, and the interesting case is several machines pointing at one shared cache folder. Analyses finished on one machine are found by Rescan library on the others and adopted without re-running, which only works if a cache entry is portable. That puts a hard constraint on everything the analysis side writes: separated stems, alignments, the key and tempo variants chosen after analysis, and the cached playback variants kept around for quick retries. Key and tempo can be adjusted after a song is analyzed, so a retry in a different key reuses a stored variant rather than the analyzer output. The three server sources feed the same store. Jellyfin songs cache locally on first play so karaoke behaves the same as a folder library, Navidrome login details are kept encrypted on disk, and Plex playlists import read only, which covers playlists coming from Plex, Jellyfin, Navidrome or a folder library's `.m3u`, `.m3u8` or `.pls` files.
Monitor gain reaches 200 percent, which is why the beep test exists
Pitch scoring runs off real-time microphone input, produces star ratings and keeps a per-song scoreboard, and those scores are tracked per player profile. The app also routes your live mic into playback, with an adjustable monitor gain running from 0 to 200 percent, and it ships a beep based latency test in Settings so scoring lines up with your room. A gain range reaching twice unity means the microphone signal can be pushed harder than it was captured at, out of a speaker, into a room that also holds the microphone doing the pitch detection. The project treats that as setup work rather than something to design around, and puts the measurement tool in the same panel as the gain slider. The same living room scenario shows up elsewhere: gamepad navigation down to D-pad, sticks and face buttons, on-screen playback controls on touch devices, and a layout that scales from phones and tablets to 4K TVs, with 10 GPU shaders reacting to the microphone in real time.
Linux is excluded from the updater
On macOS and Windows, Nightingale checks for new releases once at launch, puts a green dot on the sidebar avatar, and opens a dialog with the release notes from the Update entry. Clicking Install & Restart downloads the signed bundle, installs it and relaunches. Linux gets none of that: the Update entry opens GitHub Releases and you fetch the build yourself. That split lines up with the self-hosted web mode, where you run Nightingale on a Linux box on your home network and open it from phones, laptops, tablets and TVs at `<hostname>.local`. The Docker image, documented for both CPU and CUDA or GPU hosts, follows the same pattern, since the first-launch downloads and any later upgrade are yours to redo. The update walkthrough itself stops partway through its Windows specific step, immediately after the Install & Restart dialog, so the Windows detail past that point is left to the release notes it points at. Release history is short and fast: v1.0.0 on 2026-07-25, v1.1.0 on 2026-08-14, v1.2.0 on 2026-09-02.
The Docker image path uses a different handle than the repository owner
The badge at the top of the README points at hub.docker.com/r/razzaru/nightingale, while the source lives under the owner rzru, so the container namespace and the repository namespace are spelled differently. The release link is relative, `../../releases`, which resolves only when you are already reading the file inside its own repository. Underneath those details the tree is a conventional Rust layout: .cargo, app-core, client, docker, scripts, site and xtask, with a CHANGELOG.md and CONTRIBUTING.md at the root, and a .rulesync directory plus rulesync.jsonc alongside them. The site directory holds the self-hosted, Docker and USDX pages the feature list links to. Licensing is GPL-3.0, and the same tree that carries that license bootstraps PyTorch, UVR, Demucs, WhisperX and NeMo onto the user's machine during setup, third party components that the Rust dependency policy file has no way to inspect.
Editorial conclusion
Nightingale earns a trial if your music already lives in Plex, Jellyfin, Navidrome or a folder and you want star ratings without a subscription karaoke service. Check three things before you commit: how much model weight the first launch pulls and which folders it lands in, whether your hardware has a CUDA, Apple Silicon or MPS path or will run everything through WhisperX on CPU, and whether you can live with a Linux build you update by hand. It asks for no credentials you do not already hold, and GPL-3.0 governs anything you redistribute.
Frequently asked questions
What does Nightingale need before it can score a song?
A local music folder, a Plex, Jellyfin or Navidrome server, or the self-hosted web library, plus a microphone. Stem separation uses the UVR Karaoke model by default or Demucs, and word level lyrics come from WhisperX alignment or from LRCLIB when a match exists.
Which platforms get in-app updates in Nightingale?
macOS and Windows. Both check for new releases once at launch, badge the sidebar avatar, and install a signed bundle from an Install & Restart dialog. On Linux the Update entry opens GitHub Releases and the build is fetched by hand.
Can Nightingale reuse analysis work done on another computer?
Yes, through a shared cache folder. Point several machines at the same cache and Rescan library picks up analyses completed elsewhere without re-running them. Cache, models, videos and vendor tools can each be redirected to their own folder from Settings.
Does Nightingale need Python or ffmpeg installed first?
No. Setup downloads ffmpeg, uv, Python, PyTorch and the ML packages automatically, and the video backgrounds arrive pre-downloaded, so the first session is ready once that finishes. The data folder is picked during that setup and can be changed later.
How does Nightingale handle UltraStar Deluxe songs?
It reads pitch and lyric data straight out of the files. A song folder holding a .txt or .usdx file plus sibling audio, vocals, instrumental and video can be dropped into your library, and no analyzer pass runs for it. The feature is labelled experimental.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/rzru-nightingale)Star history chart for your README
[](https://hysenlabs.com/projects/rzru-nightingale)Was this useful?