github NeptuneHub/AudioMuse-AI v1.1.0
AudioMuse AI v1.1.0: Lyrics Semantic Search

5 hours ago

Release Date: May 4, 2026

AudioMuse AI v1.1.0 introduces a new Lyrics Semantic Search ecosystem, enabling deep lyric-based discovery powered by semantic vectors.

When LYRICS_ENABLED=true (default enabled, configurable in setup wizard; not supported in -noavx2 builds), the system analyzes song lyrics and builds semantic embeddings for advanced search and similarity matching. Run a new analysis when enabled will analyze the lyrics for the already analyzed song, or do a full analysis (Musicnn + DCLAP + Lyrics) for the new one.

This release also introduce different bugfix.

New Search Capabilities

  • Axis-based search: Explore songs across 5 defined semantic axes, selecting one or more values that best describe the target mood or meaning.
  • Text search: Simple natural language queries (e.g., “love”, “run”) focused on lyrical meaning, not musical groove (distinct from DCLAP search).
  • Song similarity search: Use a reference track to find similar songs, weighted by default as 75% lyrical meaning and 25% audio similarity to preserve genre consistency.

Search quality depends on the completeness of your music library metadata and lyrics availability.

Lyrics Analysis Pipeline (Required)

To enable this feature, a full lyrics analysis is mandatory:

  • Retrieve lyrics from the local music server
    • Jellyfin: GET /Items/{id}/Lyrics (built-in since 10.8, no plugin needed)
    • Emby: GET /emby/Items/{id}/Lyrics (same structure as Jellyfin)
    • Navidrome: GET /rest/getLyricsBySongId?id={id} (requires Navidrome 0.49+ with OpenSubsonic)
    • Lyrion/LMS: POST /jsonrpc.js with songinfo + tags (standard LMS)
  • ELSE If configured, fetch lyrics via external API
  • ELSE If unavailable, fallback to Whisper-based transcription

Example API formats supported in Setup Wizard:

https://api.example.com/get?artist={artist}&title={title}
https://api.example.com/v1/{artist}/{title}

Performance Note

Lyrics analysis is computationally intensive and require (on top of the normal Musicnn/DCLAP analysis):

  • Raspberry Pi 5: ~3–4 minutes per track (Whisper fallback)
  • Intel i7 14th gen: ~30 seconds per track

On large music libraries and lower-end hardware, analysis can take considerable time if lyrics are not available locally or via external APIs. This is expected behavior rather than a performance issue.

Multilingual Lyrics Support

AudioMuse AI is optimized for English lyrics, as the embedding model is English-based. For non-English tracks, lyrics are first transcribed using Whisper (when not available from external sources) and then translated into English using MarianMT (Helsinki-NLP opus-mt-{lang}-en models) before embedding. Only supported language pairs (~60+ languages on HuggingFace) are processed; otherwise, lyrics are skipped to preserve embedding quality.

Whisper: https://github.com/openai/whisper#available-models-and-languages

MarianMT: https://huggingface.co/Helsinki-NLP

What's Changed

New Contributors

Full Changelog: v1.0.4...v1.1.0

Don't miss a new AudioMuse-AI release

NewReleases is sending notifications on new releases.