SilentMetronome — pretrained checkpoints

Trained checkpoints for SilentMetronome, beat-phase conditioning and auxiliary supervision for beat-aligned streaming music accompaniment generation. The code, training configs, and full documentation live in the GitHub repository; this repo only hosts the weights.

All models are trained on Slakh2100 for 200k steps on a single A100 with the recipe in the paper. Checkpoints are weights-only Lightning checkpoints (optimizer states stripped), so they load for inference, evaluation, and warm-starting (--init_from_checkpoint) but cannot resume the original optimizer trajectory. Each folder pairs step=200000.ckpt with the exact args.yml the run was launched with, which is what load_lit_model reads to rebuild the architecture.

Models

The seven models correspond to Table I of the paper (Slakh2100 test split, 1024 windows, 1 s chunks, zero / 50 token look-ahead, oracle beat grid, mean over five sampling seeds).

Folder	Paper row	Beat-F	CoCoLa	FAD
pref_dec_online_fv0_k50_beat	Baseline	0.133	58.56	5.56
pref_dec_online_fv0_k50_aux_only_mp_cqt_tt_future	+ Aux only (all heads)	0.133	57.93	5.46
pref_dec_online_fv0_k50_beat_phase_dit_no_phase	+ Cond (tempo + time sig. only)	0.141	57.13	4.80
pref_dec_online_fv0_k50_beat_phase_dit	+ SiMe (full cond)	0.380	60.03	4.25
pref_dec_online_fv0_k50_beat_phase_dit_mp_cqt_aux	+ SiMe + Aux (pitch, spectrum)	0.411	60.36	3.96
pref_dec_online_fv0_k50_beat_phase_dit_mp_cqt_aux_tt_future	+ SiMe + Aux + future head (headline)	0.432	60.84	4.38
pref_dec_online_fv50_k50_beat	Non-causal ref. (fv = +50)	0.269	61.70	5.29

fv is the future-visibility offset in frames at the 50 Hz token rate (+50 = 1 s look-ahead, 0 = strictly causal). Full per-instrument, harmonic/percussive, and long-horizon drift breakdowns for every row are on the demo page.

Usage

# everything (~12.8 GB):
huggingface-cli download kevin-bretz/SilentMetronome --local-dir models/

# or a single model, e.g. the headline system:
huggingface-cli download kevin-bretz/SilentMetronome \
    --include "pref_dec_online_fv0_k50_beat_phase_dit_mp_cqt_aux_tt_future/*" --local-dir models/

Then follow the Evaluation section of the GitHub README, for example:

python scripts/gen_pred/gen_and_evaluate.py \
    --model_type prefix_decoder_online \
    --model_path models/pref_dec_online_fv0_k50_beat_phase_dit_mp_cqt_aux_tt_future/step=200000.ckpt \
    --split test \
    --num_samples 1024

The causal DAC codec weights are not included here, download them from lukewys/stream_music_gen as described in the GitHub README. The pinned x-transformers==2.16.0 is load-bearing, later 2.17.x releases silently break checkpoint compatibility.

Citation

A paper describing SilentMetronome is in preparation, and this section will carry its reference once it is public. The underlying system is stream-music-gen by Wu et al., see the code repository for the full acknowledgement and citation.