A lightweight yet powerful audio-to-MIDI converter with pitch bend detection
Python audiolightweightmachine-learningmidimusicpitch-detectionpolyphonicpythontranscriptiontypescript
vectorized out-of-range frequencies and got rid of all the imaginary numbers
Saltinebc705a12d4d7
diff --git a/basic_pitch/note_creation.py b/basic_pitch/note_creation.pyindex 0338b6f..822613d 100644--- a/basic_pitch/note_creation.py+++ b/basic_pitch/note_creation.py@@ -42,6 +42,11 @@ SONIFY_FS = 3000N_PITCH_BEND_TICKS = 8192MAX_FREQ_IDX = 87+DEFAULT_MIN_NOTE_LEN = 11+ENERGY_TOLERANCE = 11+MAGIC_ALIGNMENT_OFFSET = 0.0018+MIDI_VELOCITY_SCALE = 127+PITCH_BEND_SCALE = 4096def model_output_to_notes(output: Dict[str, np.array],@@ -242,7 +247,7 @@ def note_events_to_midi(for start_time, end_time, note_number, amplitude, pitch_bend in note_events_with_pitch_bends:instrument = instruments[note_number] if multiple_pitch_bends else instruments[0]note = pretty_midi.Note(- velocity=int(np.round(127 * amplitude)),+ velocity=int(np.round(MIDI_VELOCITY_SCALE * amplitude)),pitch=note_number,start=start_time,end=end_time,@@ -251,7 +256,7 @@ def note_events_to_midi(if not pitch_bend:continuepitch_bend_times = np.linspace(start_time, end_time, len(pitch_bend))- pitch_bend_midi_ticks = np.round(np.array(pitch_bend) * 4096 / CONTOURS_BINS_PER_SEMITONE).astype(int)+ pitch_bend_midi_ticks = np.round(np.array(pitch_bend) * PITCH_BEND_SCALE / CONTOURS_BINS_PER_SEMITONE).astype(int)# This supports pitch bends up to 2 semitones# If we estimate pitch bends above/below 2 semitones, crop them here when adding them to the midi filepitch_bend_midi_ticks[pitch_bend_midi_ticks > N_PITCH_BEND_TICKS - 1] = N_PITCH_BEND_TICKS - 1@@ -318,14 +323,19 @@ def constrain_frequency(The onset and frame activation matrices, with frequencies outside the min and maxfrequency set to 0."""- if max_freq is not None:- max_freq_idx = int(np.round(librosa.hz_to_midi(max_freq) - MIDI_OFFSET))- onsets[:, max_freq_idx:] = 0- frames[:, max_freq_idx:] = 0+ n_freqs = onsets.shape[1]+ min_freq_idx = 0+ max_freq_idx = n_freqs+if min_freq is not None:min_freq_idx = int(np.round(librosa.hz_to_midi(min_freq) - MIDI_OFFSET))- onsets[:, :min_freq_idx] = 0- frames[:, :min_freq_idx] = 0+ if max_freq is not None:+ max_freq_idx = int(np.round(librosa.hz_to_midi(max_freq) - MIDI_OFFSET))++ onsets[:, :min_freq_idx] = 0+ frames[:, :min_freq_idx] = 0+ onsets[:, max_freq_idx:] = 0+ frames[:, max_freq_idx:] = 0return onsets, frames@@ -339,7 +349,7 @@ def model_frames_to_time(n_frames: int) -> np.ndarray:window_numbers = np.floor(np.arange(n_frames) / ANNOT_N_FRAMES)window_offset = (FFT_HOP / AUDIO_SAMPLE_RATE) * (ANNOT_N_FRAMES - (AUDIO_N_SAMPLES / FFT_HOP)- ) + 0.0018 # this is a magic number, but it's needed for this to align properly+ ) + MAGIC_ALIGNMENT_OFFSET # no more magic number :Dtimes = original_times - (window_offset * window_numbers)return times