def add_voices(part):
by_staff = partition(attrgetter('staff'), part.notes_tied)
max_voice = 0
for staff, notes in by_staff.items():
voices = estimate_voices(notes_to_notearray(notes))
assert len(voices) == len(notes)
for n, voice in zip(notes, voices):
assert voice > 0
n.voice = voice + max_voice
n_next = n
while n_next.tie_next:
n_next = n_next.tie_next
n_next.voice = voice + max_voice
max_voice = np.max(voices)
if any([n.voice is None for n in part.notes]):
# Hack to add voices to notes not included in a staff!
# not musically meaningful
ev = 1
for n in part.notes:
if n.voice is None:
n.voice = max_voice + ev
ev += 1
def test_vosa_chew_chordnotes(self):
# Example from Chew and Wu.
notearray = musicxml_to_notearray(VOSA_TESTFILES[0], beat_times=False)
voices = estimate_voices(notearray, monophonic_voices=False)
ground_truth_voices = np.array(
[1, 2, 2, 1, 1, 2, 1, 1, 2, 1, 1, 2, 2, 3, 2, 2, 1, 1, 2, 2])
self.assertTrue(np.all(voices == ground_truth_voices),
'Incorrect voice assignment.')
def test_vosa_chew(self):
# Example from Chew and Wu.
# notearray = musicxml_to_notearray(VOSA_TESTFILES[0], beat_times=False)
voices = estimate_voices(self.score, monophonic_voices=True)
# ground_truth_voices = np.array([3, 2, 1, 3, 3, 2, 3, 3, 2, 3,
# 3, 1, 1, 2, 1, 1, 3, 2, 1, 1])
ground_truth_voices = np.array(
[1, 2, 3, 1, 1, 2, 1, 1, 2, 1, 1, 3, 3, 2, 3, 3, 1, 2, 3, 3]
)
self.assertTrue(
np.all(voices == ground_truth_voices), "Incorrect voice assignment."
)
def add_staffs_v1(part):
# assign staffs by first estimating voices jointly, then assigning voices to staffs
notes = part.notes_tied
# estimate voices in strictly monophonic way
voices = estimate_voices(notes_to_notearray(notes), monophonic_voices=True)
# for v, note in zip(voices, notes):
# print(note.start.t, note.midi_pitch, v)
# group notes by voice
by_voice = partition(itemgetter(0), zip(voices, notes))
clefs = {}
for v, vnotes in by_voice.items():
# voice numbers may be recycled throughout the piece, so we split by
# time gap
t_diffs = np.diff([n.start.t for _, n in vnotes])
t_threshold = np.inf # np.median(t_diffs)+1
note_groups = np.split([note for _, note in vnotes],
np.where(t_diffs > t_threshold)[0] + 1)
# for each note group estimate the clef
for note_group in note_groups:
if len(note_group) > 0:
pitches = [n.midi_pitch for n in note_group]
clef = tuple(estimate_clef_properties(pitches).items())
staff = clefs.setdefault(clef, len(clefs))
# print((note_group[0].start.t, note_group[-1].end.t),
# (np.min(pitches), np.max(pitches), np.mean(pitches)), clef)
for n in note_group:
n.staff = staff
n_tied = n.tie_next
while n_tied:
n_tied.staff = staff
n_tied = n_tied.tie_next
# re-order the staffs to a fixed order (see CLEF_ORDER), rather than by
# first appearance
clef_list = list((dict(clef), i) for clef, i in clefs.items())
clef_list.sort(key=lambda x: x[0].get('octave_change', 0))
clef_list.sort(key=lambda x: CLEF_ORDER.index(x[0].get('sign', 'G')))
staff_map = dict((j, i + 1) for i, (_, j) in enumerate(clef_list))
for n in notes:
n.staff = staff_map[n.staff]
for i, (clef_properties, _) in enumerate(clef_list):
part.add(score.Clef(number=i + 1, **clef_properties), 0)
def load_score_midi(fn, part_voice_assign_mode=0, ensure_list=False,
quantization_unit=None, estimate_voice_info=True,
estimate_key=False, assign_note_ids=True):
"""Load a musical score from a MIDI file and return it as a Part
instance.
This function interprets MIDI information as describing a score.
Pitch names are estimated using Meredith's PS13 algorithm [1]_.
Assignment of notes to voices can either be done using Chew and
Wu's voice separation algorithm [2]_, or by choosing one of the
part/voice assignment modes that assign voices based on
track/channel information. Furthermore, the key signature can be
estimated based on Krumhansl's 1990 key profiles [3]_.
This function expects times to be metrical/quantized. Optionally a
quantization unit may be specified. If you wish to access the non-
quantized time of MIDI events you may wish to used the
`load_performance_midi` function instead.
Parameters
----------
fn : str
Path to MIDI file
part_voice_assign_mode : {0, 1, 2, 3, 4, 5}, optional
This keyword controls how part and voice information is
associated to track and channel information in the MIDI file.
The semantics of the modes is as follows:
0
Return one Part per track, with voices assigned by channel
1
Return one PartGroup per track, with Parts assigned by channel
(no voices)
2
Return single Part with voices assigned by track (tracks are
combined, channel info is ignored)
3
Return one Part per track, without voices (channel info is
ignored)
4
Return single Part without voices (channel and track info is
ignored)
5
Return one Part per <track, channel> combination, without
voices Defaults to 0.
ensure_list : bool, optional
When True, return a list independent of how many part or partgroup
elements were created from the MIDI file. By default, when the
return value of `load_score_midi` produces a single
:class:`partitura.score.Part` or :class:`partitura.score.PartGroup`
element, the element itself is returned instead of a list
containing the element. Defaults to False.
quantization_unit : integer or None, optional
Quantize MIDI times to multiples of this unit. If None, the
quantization unit is chosen automatically as the smallest
division of the parts per quarter (MIDI "ticks") that can be
represented as a symbolic duration. Defaults to None.
estimate_key : bool, optional
When True use Krumhansl's 1990 key profiles [3]_ to determine
the most likely global key, discarding any key information in
the MIDI file.
estimate_voice_info : bool, optional
When True use Chew and Wu's voice separation algorithm [2]_ to
estimate voice information. This option is ignored for
part/voice assignment modes that infer voice information from
the track/channel info (i.e. `part_voice_assign_mode` equals
1, 3, 4, or 5). Defaults to True.
Returns
-------
:class:`partitura.score.Part`, :class:`partitura.score.PartGroup`, or a list of these
One or more part or partgroup objects
References
----------
.. [1] Meredith, D. (2006). "The ps13 Pitch Spelling Algorithm". Journal
of New Music Research, 35(2):121.
.. [2] Chew, E. and Wu, Xiaodan (2004) "Separating Voices in
Polyphonic Music: A Contig Mapping Approach". In Uffe Kock,
editor, Computer Music Modeling and Retrieval (CMMR), pp. 1–20,
Springer Berlin Heidelberg.
.. [3] Krumhansl, Carol L. (1990) "Cognitive foundations of musical pitch",
Oxford University Press, New York.
"""
mid = mido.MidiFile(fn)
divs = mid.ticks_per_beat
# these lists will contain information from dedicated tracks for meta
# information (i.e. without notes)
global_time_sigs = []
global_key_sigs = []
global_tempos = []
# these dictionaries will contain meta information indexed by track (only
# for tracks that contain notes)
time_sigs_by_track = {}
key_sigs_by_track = {}
tempos_by_track = {}
track_names_by_track = {}
# notes are indexed by (track, channel) tuples
notes_by_track_ch = {}
relevant = {'time_signature',
'key_signature',
'set_tempo',
'note_on',
'note_off'}
for track_nr, track in enumerate(mid.tracks):
time_sigs = []
key_sigs = []
# tempos = []
notes = defaultdict(list)
# dictionary for storing the last onset time and velocity for each
# individual note (i.e. same pitch and channel)
sounding_notes = {}
# current time (will be updated by delta times in messages)
t_raw = 0
for msg in track:
t_raw = t_raw + msg.time
if msg.type not in relevant:
continue
if quantization_unit:
t = quantize(t_raw, quantization_unit)
else:
t = t_raw
if msg.type == 'time_signature':
time_sigs.append((t, msg.numerator, msg.denominator))
if msg.type == 'key_signature':
key_sigs.append((t, msg.key))
if msg.type == 'set_tempo':
global_tempos.append((t, 60*10**6/msg.tempo))
else:
note_on = msg.type == 'note_on'
note_off = msg.type == 'note_off'
if not (note_on or note_off):
continue
# hash sounding note
note = note_hash(msg.channel, msg.note)
# start note if it's a 'note on' event with velocity > 0
if note_on and msg.velocity > 0:
# save the onset time and velocity
sounding_notes[note] = (t, msg.velocity)
# end note if it's a 'note off' event or 'note on' with velocity 0
elif note_off or (note_on and msg.velocity == 0):
if note not in sounding_notes:
warnings.warn('ignoring MIDI message %s' % msg)
continue
# append the note to the list associated with the channel
notes[msg.channel].append((sounding_notes[note][0], msg.note, t-sounding_notes[note][0]))
# sounding_notes[note][1]])
# remove hash from dict
del sounding_notes[note]
# if a track has no notes, we assume it may contain global time/key sigs
if not notes:
global_time_sigs.extend(time_sigs)
global_key_sigs.extend(key_sigs)
else:
# if there are note, we store the info under the track number
time_sigs_by_track[track_nr] = time_sigs
key_sigs_by_track[track_nr] = key_sigs
track_names_by_track[track_nr] = track.name
for ch, ch_notes in notes.items():
# if there are any notes, store the notes along with key sig / time
# sig / tempo information under the key (track_nr, ch_nr)
if len(ch_notes) > 0:
notes_by_track_ch[(track_nr, ch)] = ch_notes
tr_ch_keys = sorted(notes_by_track_ch.keys())
group_part_voice_keys, part_names, group_names = assign_group_part_voice(
part_voice_assign_mode,
tr_ch_keys,
track_names_by_track)
# for key and time sigs:
track_to_part_mapping = make_track_to_part_mapping(
tr_ch_keys,
group_part_voice_keys)
# pairs of (part, voice) for each note
part_voice_list = [[part, voice] for tr_ch, (_, part, voice)
in zip(tr_ch_keys, group_part_voice_keys)
for i in range(len(notes_by_track_ch[tr_ch]))]
# pitch spelling, voice estimation and key estimation are done on a
# structured array (onset, pitch, duration) of all notes in the piece
# jointly, so we concatenate all notes
# note_list = sorted(note for notes in (notes_by_track_ch[key] for key in tr_ch_keys) for note in notes)
note_list = [note for notes in (notes_by_track_ch[key]
for key in tr_ch_keys)
for note in notes]
note_array = np.array(note_list, dtype=[('onset', np.int),
('pitch', np.int),
('duration', np.int)])
LOGGER.debug('pitch spelling')
spelling_global = analysis.estimate_spelling(note_array)
if estimate_voice_info:
LOGGER.debug('voice estimation')
# TODO: deal with zero duration notes in note_array. Zero duration notes are currently deleted
estimated_voices = analysis.estimate_voices(note_array)
assert len(part_voice_list) == len(estimated_voices)
for part_voice, voice_est in zip(part_voice_list, estimated_voices):
if part_voice[1] is None:
part_voice[1] = voice_est
if estimate_key:
LOGGER.debug('key estimation')
_, mode, fifths = analysis.estimate_key(note_array)
key_sigs_by_track = {}
global_key_sigs = [(0, fifths_mode_to_key_name(fifths, mode))]
if assign_note_ids:
note_ids = ['n{}'.format(i) for i in range(len(note_array))]
else:
note_ids = [None for i in range(len(note_array))]
time_sigs_by_part = defaultdict(set)
for tr, ts_list in time_sigs_by_track.items():
for ts in ts_list:
for part in track_to_part_mapping[tr]:
time_sigs_by_part[part].add(ts)
for ts in global_time_sigs:
for part in set(part for _, part, _ in group_part_voice_keys):
time_sigs_by_part[part].add(ts)
key_sigs_by_part = defaultdict(set)
for tr, ks_list in key_sigs_by_track.items():
for ks in ks_list:
for part in track_to_part_mapping[tr]:
key_sigs_by_part[part].add(ks)
for ks in global_key_sigs:
for part in set(part for _, part, _ in group_part_voice_keys):
key_sigs_by_part[part].add(ks)
# names_by_part = defaultdict(set)
# for tr_ch, pg_p_v in zip(tr_ch_keys, group_part_voice_keys):
# print(tr_ch, pg_p_v)
# for tr, name in track_names_by_track.items():
# print(tr, track_to_part_mapping, name)
# for part in track_to_part_mapping[tr]:
# names_by_part[part] = name
notes_by_part = defaultdict(list)
for (part, voice), note, spelling, note_id in zip(part_voice_list,
note_list,
spelling_global,
note_ids):
notes_by_part[part].append((note, voice, spelling, note_id))
partlist = []
part_to_part_group = dict((p, pg) for pg, p, _ in group_part_voice_keys)
part_groups = {}
for part_nr, note_info in notes_by_part.items():
notes, voices, spellings, note_ids = zip(*note_info)
part = create_part(divs, notes, spellings, voices, note_ids,
sorted(time_sigs_by_part[part_nr]),
sorted(key_sigs_by_part[part_nr]),
part_id='P{}'.format(part_nr+1),
part_name=part_names.get(part_nr, None))
# print(part.pretty())
# if this part has an associated part_group number we create a PartGroup
# if necessary, and add the part to that. The newly created PartGroup is
# then added to the partlist.
pg_nr = part_to_part_group[part_nr]
if pg_nr is None:
partlist.append(part)
else:
if pg_nr not in part_groups:
part_groups[pg_nr] = score.PartGroup(group_name=group_names.get(pg_nr, None))
partlist.append(part_groups[pg_nr])
part_groups[pg_nr].children.append(part)
# add tempos to first part
part = next(score.iter_parts(partlist))
for t, qpm in global_tempos:
part.add(score.Tempo(qpm, unit='q'), t)
if not ensure_list and len(partlist) == 1:
return partlist[0]
else:
return partlist