Nightingale – open-source karaoke app that works with any song on your computer

Original Article ↗ Hacker News Discussion ↗

Overall reception & use cases

  • Many commenters are enthusiastic about a fully local, open-source karaoke app that works with arbitrary audio/video files and doesn’t require accounts or telemetry.
  • People see it as a promising alternative to YouTube karaoke and subscription services, especially for niche, avant‑garde, or local tracks, and as a party/family tool.

Transcription, stem separation & scoring quality

  • Stem separation generally works but struggles with:
    • Multiple singers / harmonies (often only one voice is removed).
    • Busy mixes, some electronic music, some non‑Western or niche tracks.
  • WhisperX transcription and alignment:
    • Works well on some songs (e.g., classic pop/country), but can drift, skip lyrics, or “slide off” in lyric-heavy tracks and with backing vocals.
    • Support for non‑English languages (Japanese, Mandarin, Cantonese) is described as possible but currently weaker and hit‑or‑miss.
  • Pitch scoring exists but is simple; there’s no clear “next note height” display yet.

Features & requested improvements

  • Frequently requested:
    • Pitch and tempo controls.
    • Duet/multi-singer support.
    • Better playback controls (seek forward/back), UI contrast fixes, confirmation before deleting models, and clearer model/settings behavior.
    • Ability to edit lyrics/timings, export to formats like UltraStar/Performous, and optionally show the original music video as background.
    • Remote/server preprocessing or a client–server mode for weak machines; potential plugins (e.g., Navidrome).

Performance & hardware

  • Runs best on NVIDIA GPUs (Maxwell+), Apple Silicon; Raspberry Pi–class devices and possibly Steam Deck are expected to struggle, especially during analysis.
  • Users report 10–15 minutes to process a ~3.5 minute song even on strong GPUs, though results can be good when it works.

Packaging, dependencies & security concerns

  • The single-binary design that downloads Python, FFmpeg, PyTorch, and models on first launch is contentious:
    • Critics argue runtime binary downloads are unsafe/unusual, especially on Linux where system packages exist, and report issues like mislocated interpreters.
    • Others defend vendoring dependencies for “grandma-proof” installation and to avoid distro packaging breakage, while agreeing they should be bundled rather than hot-downloaded.
  • VirusTotal and browser/AV warnings are reported; consensus in-thread is they are false positives, but they increase concern.

Comparisons & skepticism

  • Compared with tools like UltraStar, Karafun, YARG and others; this app’s advantage is auto-generating karaoke from any file instead of relying on pre-made tracks.
  • Some remain skeptical of claims like “works with any song” and robust auto-lyrics, noting current ML limitations, but still see strong potential.