Skip to content

CHANGES.md

Latest commit

 

History

History
84 lines (63 loc) · 4.27 KB

CHANGES.md

File metadata and controls

84 lines (63 loc) · 4.27 KB
Raw

Release Notes

Version 1.1.0

This version introduces the following changes:

Changed

  • in the frame-based majority metric mpteval.metrics.majority added an additional step to smoothen out spuriously (by majority) introduced note events by thresholding notes to have need a minimal duration of 50ms and merging fragmented notes (reonset within 30ms) into one long note. Also changed the majority voting to be based on dilated (+- 15 frames) roles to consider for timing jitter introduced by transcription.

Version 1.0.1

This version introduces the following changes:

Changed

Added

  • preprocessing: added chordification and time normalisation fns in mpteval.preprocessing.preprocess
  • alignment:
    • added alignment function and wrapper for parallelized processing to align chordified performances via DTW in mpteval.alignment.dtw_align
    • added custom distance metric in mpteval.alignment.dtw_dist
  • clustering:
    • added clustering fns in mpteval.clustering.cluster to cluster transcriptions of the same piece into different performed structural realisations
    • added cluster evaluation in mpteval.clustering.cluster_eval
    • added grid search for clustering parameters in mpteval.clustering.cluster_gs
  • metrics:
    • added majority based metric in mpteval.metrics.majority: given multiple transcriptions and the source audio file, selects a base transcription by audio duration proximity, removes notes found in only a minority of transcriptions (likely false positives), and inserts notes found in the majority but missing from the base (likely negatives).
  • utils:
    • added utility fns in mpteval.utils:
      • MusicEncoder: JSON encoder for numpy types
      • plot_piano_roll: Visualize piano rolls
      • create_note_list: Convert note arrays to structured lists with onset, offset, pitch, velocity
      • is_monophonic: Check if a note array is monophonic
      • midi_to_piano_note: Convert MIDI numbers to human-readable note names
      • alignment_list_to_df: Convert alignment lists to pandas DataFrames
      • group_by_structure: Group performance IDs by structural features
      • pairwise_distance_matrix: Compute pairwise distance matrices between sequences
      • accumulated_cost_matrix: Compute DTW accumulated cost matrix
      • optimal_warping_path: Extract optimal warping path from accumulated cost matrix
      • dynamic_time_warping: Vanilla dynamic time warping implementation
      • fast_dynamic_time_warping: Fast approximate DTW using fastdtw library
      • greedy_note_alignment: Extract note-level alignments from DTW warping paths
      • notewise_alignment: Complete note-wise music alignment combining DTW with greedy matching

Changed

  • moved all (global) configs to config.py
  • renamed mpteval.benchmarks to mpteval.metrics.objective
  • use consistent voice separation function in mpteval.metrics_feature.timing as in mpteval.metrics_feature.articulation

Version 0.1.4

This version was used to measure how variations and shifts in both sound characteristics and musical distribution impact the robustness of several state-of-the-art piano transcription models, as described here. It introduced the following changes:

Added

  • added benchmark metrics from mir_eval for standard frame and notewise transcription evaluation, with an added option to compute notewise metrics considering only onset and velocity (ignoring offsets)

Changed

  • updated partitura dependency to 1.7.0
  • mpteval.timing.timing_metrics_from_perf() returns only correlation metrics for melody, accompaniment and their ratio, and takes as additional argument include_distance to specify whether to compute DTW or histogram distance metrics (defaults to None)

Fixed

  • made timing, articulation and dynamics metrics robust to monophonic reference and polyphonic prediction.

Version 0.1.3

This version contains the initial implementation of musically informed evaluation metrics for transcriptions of piano performances as described here. The metrics are defined to measure similarity between a predicted performance and a reference performance in terms of timing, articulation, dynamics, and harmony.