Release Date: May 25, 2026
AudioMuse AI v2.1.0 re-export the GTE Lyrics model to fix incorrect embeddings on older CPUs without VNNI support (avx512_vnni / avx_vnni). On previous AudioMuse-AI release, affected CPUs embeddings could overflow and produce inaccurate results.
To check whether your CPU supports VNNI:
grep -oE 'avx512_vnni|avx_vnni' /proc/cpuinfo | head -1If result is empty, VNNI is not supported by your cpu.
This release also improves lyrics processing performance by skipping unnecessary song downloads when other analyses are already completed and lyrics are retrieved directly from the music server or an external API.
Re-analysis (only for CPUs without VNNI support)
If your CPU does not support VNNI, a full lyrics re-analysis is required. To trigger a re-analysis you need to drop the lyrics tables.
Replace the database connection parameters as needed:
PGPASSWORD=audiomusepassword psql -h 192.168.3.213 -p 5432 -U audiomuse -d audiomusedb \
-c "DROP TABLE IF EXISTS lyrics_embedding; DROP TABLE IF EXISTS lyrics_index_data; DROP TABLE IF EXISTS lyrics_axes_index_data;"Docker Compose:
docker compose exec -e PGPASSWORD=audiomusepassword postgres \
psql -U audiomuse -d audiomusedb \
-c "DROP TABLE IF EXISTS lyrics_embedding; DROP TABLE IF EXISTS lyrics_index_data; DROP TABLE IF EXISTS lyrics_axes_index_data;"What's Changed
- Lyrics migration speed improvement - #558 by @NeptuneHub in #565
- tokenizer version pin and test improvement by @NeptuneHub in #566
- template and dependantbot fix by @NeptuneHub in #567
- Ram improvement + Deepseek LLM fix + dockerfile restructuring by @NeptuneHub in #568
- fix gte model - re export by @NeptuneHub in #569
Full Changelog: v2.0.2...v2.1.0