Module notes

Expand description

The note list that every layer reads, the transcriber, the f0 track and the MIDI files that produce and store it, and the harmonic masks built from it.

A note has an onset, an offset, a pitch and, once a pitch track has run, an f0 contour. Three more fields stay unset until a later step fills them: the octave flag, the velocity and the overlap flag. The transcriber and the MIDI reader both produce this type, and every later step consumes it.

The transcriber runs the model once per stem (Transcriber::run) and decodes that run with DecodeSettings, as many times as needed (ModelRun::decode). A Profile gives fixed settings per kind of stem; detect() derives them from the stem instead, from a probe decode and the PitchTrack, the stem’s f0 track, which also fills the note contours. drop_quiet drops notes that sound only far below the stem’s loudest passage. write_midi stores a list as a format 0 file, and finish() rewrites a list with the layer report into the final notes.mid.

Structs§

CleanupSteps: The note counts of the cleanup, one per step.
Contour: The f0 of a note over time, so masks follow bends and glides. Points are sorted by time; between two points the pitch moves linearly in log-frequency, and outside them it holds the end value.
ContourPoint: One point of an f0 contour.
DecodeSettings: Everything the decoder needs besides its thresholds, which keep their defaults: the pitch range, the minimum note length and the melodia pass. Pitch bends stay off, because the notes file writes none.
Detection: What the detection found and the settings it chose.
DropSpan: A pitch drop of the pitch layer.
FinalNote: One note of the final list, in ticks.
FinalNotes: The final note list.
FinishInput: What finish reads.
FrameMask: The masks of one frame: for each bin the note that claims it and whether a second note claims it too.
GlideSpan: A glide of the pitch layer.
MaskParams: Settings of the harmonic masks.
Masker: Builds the masks of a note list on the bins of an STFT layout.
MidiError: A MIDI file that cannot give a note list.
ModelRun: The posteriorgrams of one model run over a stem.
Note: One note: onset, offset, pitch, and the fields later steps fill.
NoteFacts: What the layer report says of one note of the list.
NoteList: The notes of a stem, sorted by onset.
OctaveCheck: What the octave check did.
PitchRange: A range of whole MIDI pitches the decoder may create, both ends kept, within the model’s range A0 to C8.
PitchTrack: The f0 track of a stem.
StemLevels: The stem’s level in QUIET_FRAME frames at its own rate, with a quarter-frame hop, frame m centred at m hops. The quiet-note guard, the attack split and the split join read the same frames.
TrackFrame: One frame of the track.
Transcriber: The transcription model with its inference engine, built once and used for every stem, and the f0 tracker of the notes step.
Transcription: Everything the notes step of a stem found.
TranscriptionRecord: The transcriber’s model and settings, as the transcription group of layers.json records them.
Velocity: A MIDI velocity from 1 to 127. Velocity 0 means a note-off in MIDI, so it is not a velocity.
VibratoSpan: A vibrato of the pitch layer.

Enums§

EndRule: Which rule set the end of a final note.
MidiWriteError: A note list that cannot become a MIDI file.
NoteError: A note or contour that breaks a rule.
NotesError: A notes step that failed.
Profile: A fixed set of decode settings for one kind of stem.
RangeError: A pitch range that the decoder cannot use.
RegisterSource: Where the register came from.
TrackError: An f0 track that could not be built.
TranscribeError: A model run or a decode that failed.
Velocities: Which velocity each note gets.
Voicing: Whether a part plays one note at a time or chords.

Constants§

ATTACK_AFTER: How far after a note’s onset its loudest frame may lie: frames centred from the onset to 50 ms after it, and not past its end. The attack split’s piece after a cut lasts at least this long.
ATTACK_BEFORE: How far before a note’s onset the quiet frames of its attack lie: frames centred from 75 ms before the onset to the onset, and not before the earlier note starts.
ATTACK_DIP: How far the stem’s level, with any DC offset removed, must dip below the loudest frame in the quiet frames of an attack.
BEND_MIN: A note of a mono line gets bends when its contour leaves its key by this much outside every route’s span, over the tracker’s reach.
BEND_RANGE: The bend range the file assumes; a host must set the synth to it, and render reads no bend.
BPM_TOLERANCE: How far a tempo in the file may lie from --bpm. A tempo is stored in whole microseconds per beat, which rounds 140 BPM to 140.00003.
CHORD_OVERLAP: How much of a note’s span another pitch class must share with it for the note to sound in a chord.
CHORD_SHARE: The share of probe notes that sound with another pitch class from which a part counts as chords. It sits midway between the highest line (0.56) and the lowest chord part (0.79) of the generated set.
COPY_SLACK: How far a note may reach past the note one octave below it and still lie inside it: two model frames.
DEFAULT_MIN_NOTE_FRAMES: The decoder’s default minimum length in frames, 127.7 ms. A kept note lasts at least one frame more, 139.3 ms.
DEFAULT_VELOCITY: The velocity that the notes command passes to Velocities::Constant.
DROP_SLACK: A pitch drop takes in the notes that start from this long before its onset until it settles.
DRUM_CHANNEL: The General MIDI percussion channel, channel 10, counted from 0. Its keys name drums, not pitches.
FUNDAMENTAL_DIP: How far the level at the note’s fundamental alone must dip across a cut of the attack split.
GLIDE_OVERLAP: A glide join overlaps the two notes by at least this long, rounded up to whole ticks.
GLIDE_SLACK: A glide joins two notes when the later onset lies this close to the glide’s motion, from its start to its end.
HARMONICS: The harmonics of a note that StemLevels::harmonic_level sums.
HARMONIC_DIP: How far the level at the note’s own harmonics must dip in the same frames.
HIGHEST_MODEL_PITCH: The highest pitch the model can give, C8.
HOLD_BEFORE: How long the stem must hold above the first share of HOLD_SPAN before it last fell below the second share ahead of a rise.
HOLD_RATIO: How many times longer than its dip the stem must hold before it.
HOLD_SEARCH: How far back from the start of a rise the hold may end.
HOLD_SPAN: The share of the loudest reading the hold must stay above, and the share at which it ends.
HOLD_WINDOW: How far below the reference level a probe note’s median frame may lie and still count as a hold candidate. Half the register window, so bleed near -20 dB in a rest does not widen the range.
LOWEST_MODEL_PITCH: The lowest pitch the model can give, A0.
MIN_NOTE_LENGTH: The shortest note the file holds. A shorter note is lengthened to it, and to at least one tick.
MODEL_RATE: The input rate of the model.
MODEL_SHA256: SHA-256 of the official nmp.onnx (the bundled MODEL_ONNX), as a lowercase hex string.
NO_KNEE_CUT: A note without a note-off knee ends this long before its end in the list.
OCTAVE_AGREEMENT: How far a note may lie from the f0 track’s median and still agree.
OCTAVE_FIX_CEILING: The f0 median, A2 (110 Hz), at or below which, within OCTAVE_TOLERANCE, a note an octave above it moves down. The ghosts are measured there; above it a gap of an octave only sets the flag.
OCTAVE_MIN_FRAMES: The fewest voiced frames, 17 ms of track, that check a note.
OCTAVE_TOLERANCE: How close to an octave a gap must lie to count as one.
PAD_FALL: A pad note ends where its level falls this far below its plateau.
QUIET_FRAME: The length of a guard frame: 25 ms, short enough to follow a note.
QUIET_GATE: The gate of the guard: a frame is active within this many decibels of the loudest frame. It is the project’s active-frame gate.
REFERENCE_PERCENTILE: The percentile of the voiced frames’ levels that serves as the reference level. A short loud transient does not move it.
REGISTER_MARGIN: How far the range reaches beyond the register on each side.
REGISTER_PERCENTILES: The low and high percentiles of the voiced f0 frames that bound the register.
REGISTER_WINDOW: The register reads only voiced frames within this many decibels of the reference level, so quiet bleed in a rest cannot widen it.
RISE_LATEST: How far after the onset a rise may start: the decoder starts a note up to 15 ms before its attack.
RISE_READ: How many rise limits past RISE_LATEST the rise test reads.
RISE_SLACK: How much longer than a step read through the envelope window a rise may take: about the attack of a gate.
RISE_SPAN: The share of the loudest envelope reading a rise starts below, and the share it ends above.
RISE_STEP: The step between envelope readings of the rise test.
RISE_WINDOW: The shortest envelope window of the rise test; it spans one period of the note when that is longer.
SPLIT_GAP: The widest gap a split may leave: two model frames.
SPLIT_LOWEST: The lowest key whose notes split_at_attacks splits: A1, 55 Hz.
SPLIT_ONSET: The onset activation below which a note that starts where a note of the same key ends joins it.
TICKS_PER_BEAT: The resolution of the file in ticks per quarter note.
TRACK_FMAX: The highest f0 the track looks for, just above C8 (4186 Hz), the top of the transcription model. A lower ceiling reads every tone above it as a subharmonic.
TRACK_FMIN: The lowest f0 the track looks for, below B0 (30.87 Hz).
TRACK_FRAME: Samples per analysis frame at the model rate.
TRACK_HOP: Samples between frames at the model rate.
TRANSCRIBER_VERSION: The version of the transcription crate, pinned in Cargo.toml.
VELOCITY_LAW_DB: The level law: a note d dB under the loudest gets velocity 127 * 10^(-d / 40).

Traits§

Backend: An engine able to run the Basic Pitch network on one window of audio.

Functions§

check_octaves: Checks every note against the median of the voiced f0 frames inside it, and returns the notes with their octave flag set.
clean_up: Cleans decoded notes in five steps, and counts the notes after each.
detect: Detects the register and voicing of a stem from its model run and its f0 track, and builds the settings for the final decode.
detect_with_window: detect with another level window for the f0 register, for the evaluation of the window.
drop_quiet: Drops every note whose span holds no active frame of the stem, and returns the kept notes with the count of dropped ones.
drop_quiet_with: Drops the notes as drop_quiet does, over levels already measured.
finish: Rewrites the note list with the layer report.
join_splits: Joins the pieces of a held note that the decoder split, on the onset rule alone: join_splits_with with a stem of silence, which never gives an attack.
join_splits_with: Joins the pieces of a held note that the decoder split.
merge_overlaps: Merges every pair of notes of one key that sound at the same time into one note from the earlier onset to the later offset.
midi_track_name: Returns the first track name of a Standard MIDI File, or None when the bytes do not parse or the file names no track.
min_note_frames: The minimum note length for a BPM, in model frames.
min_note_length: The minimum note length for a BPM: min_note_frames model frames, the length the decoder drops a note at.
out_of_mask_fraction: Returns the share of the signal’s power that lies outside every mask.
probe_settings: The settings of the probe decode: the whole model range, the minimum length from the BPM and the melodia pass.
read_midi: Reads the notes of a MIDI file at path, checking its tempo against bpm.
read_midi_bytes: Reads the notes of a MIDI file at path as read_midi does, and returns the file’s bytes with them, so a caller can hash the very bytes it read.
read_midi_near: Reads the notes of a MIDI file at path as read_midi does, and also accepts a file whose every tempo matches refined, the trusted tempo refined from the stem. A refined equal to bpm counts as none.
split_at_attacks: Splits each note where the stem has an attack of its own inside it.
transcribe_stem: Runs the notes step on a mono stem at any rate: one model run, one f0 track, detection, the decode with the detected settings or those of profile, and clean_up.
voicing_of_notes: The voicing of a note list by the rule detection applies to its probe notes.
write_midi: Encodes the notes as a format 0 file at a BPM.
write_named_midi: Encodes the notes as write_midi does, with a track name event at tick 0 before the tempo. notes names the stem this way, so a later run can tell a file of another stem; midi_track_name reads it back.

Type Aliases§

LevelReader: Reads the full-band level of the stem from one time to another, with the RMS window of a frequency’s period.
Tracker: The f0 tracker that the notes step runs beside the model: PitchTrack::of unless Transcriber::with_tracker names another.

Module notes

Module notes Copy item path

Structs§

Enums§

Constants§

Traits§

Functions§

Type Aliases§

Module notes