A lightweight yet powerful audio-to-MIDI converter with pitch bend detection
Python audiolightweightmachine-learningmidimusicpitch-detectionpolyphonicpythontranscriptiontypescript
Fix intermediate frames trimming to make sure final notes output is not truncated (#179)
Pierre Tharye989e40ca5f9
diff --git a/basic_pitch/constants.py b/basic_pitch/constants.pyindex bf4f632..c2b717a 100644--- a/basic_pitch/constants.py+++ b/basic_pitch/constants.py@@ -23,7 +23,6 @@ from enum import EnumSEMITONES_PER_OCTAVE = 12 # for frequency bin calculationsFFT_HOP = 256-N_FFT = 8 * FFT_HOPNOTES_BINS_PER_SEMITONE = 1CONTOURS_BINS_PER_SEMITONE = 3diff --git a/basic_pitch/inference.py b/basic_pitch/inference.pyindex 24f02ea..06fa54f 100644--- a/basic_pitch/inference.py+++ b/basic_pitch/inference.py@@ -57,6 +57,7 @@ from basic_pitch.constants import (AUDIO_N_SAMPLES,ANNOTATIONS_FPS,FFT_HOP,+ AUDIO_WINDOW_LENGTH,)from basic_pitch.commandline_printing import (generating_file_message,@@ -247,6 +248,7 @@ def unwrap_output(output: npt.NDArray[np.float32],audio_original_length: int,n_overlapping_frames: int,+ hop_size: int,) -> np.array:"""Unwrap batched model predictions to a single matrix.@@ -254,6 +256,7 @@ def unwrap_output(output: array (n_batches, n_times_short, n_freqs)audio_original_length: length of original audio signal (in samples)n_overlapping_frames: number of overlapping frames in the output+ hop_size: size of the hop used when scanning the input audioReturns:array (n_times, n_freqs)@@ -266,10 +269,14 @@ def unwrap_output(# remove half of the overlapping frames from beginning and endoutput = output[:, n_olap:-n_olap, :]+ # Concatenate the frames outputs (overlapping frames removed) into a single dimensionoutput_shape = output.shape- n_output_frames_original = int(np.floor(audio_original_length * (ANNOTATIONS_FPS / AUDIO_SAMPLE_RATE)))unwrapped_output = output.reshape(output_shape[0] * output_shape[1], output_shape[2])- return unwrapped_output[:n_output_frames_original, :] # trim to original audio length++ # trim to number of expected windows in output+ n_expected_windows = audio_original_length / hop_size+ n_frames_per_window = (AUDIO_WINDOW_LENGTH * ANNOTATIONS_FPS) - n_overlapping_frames+ return unwrapped_output[: int(n_expected_windows * n_frames_per_window), :]def run_inference(@@ -303,7 +310,8 @@ def run_inference(output[k].append(v)unwrapped_output = {- k: unwrap_output(np.concatenate(output[k]), audio_original_length, n_overlapping_frames) for k in output+ k: unwrap_output(np.concatenate(output[k]), audio_original_length, n_overlapping_frames, hop_size)+ for k in output}if debug_file:diff --git a/tests/resources/vocadito_10/model_output.npz b/tests/resources/vocadito_10/model_output.npzindex 9b2a404..55e24a8 100644Binary files a/tests/resources/vocadito_10/model_output.npz and b/tests/resources/vocadito_10/model_output.npz differdiff --git a/tests/test_inference.py b/tests/test_inference.pyindex 3a71038..eb3a96d 100644--- a/tests/test_inference.py+++ b/tests/test_inference.py@@ -26,12 +26,13 @@ import pretty_midiimport numpy as npimport numpy.typing as npt-from basic_pitch import ICASSP_2022_MODEL_PATH, inference+from basic_pitch import ICASSP_2022_MODEL_PATH, inference, note_creationfrom basic_pitch.constants import (AUDIO_SAMPLE_RATE,AUDIO_N_SAMPLES,ANNOTATIONS_N_SEMITONES,FFT_HOP,+ ANNOTATION_HOP,)RESOURCES_PATH = pathlib.Path(__file__).parent / "resources"@@ -55,6 +56,13 @@ def test_predict() -> None:assert all(note_pitch_max)assert isinstance(note_events, list)+ # Check that model output has the expected length according to the last frame second computed downstream+ # (via model_frames_to_time) with to a few frames of tolerance+ audio_length_s = librosa.get_duration(filename=test_audio_path)+ n_model_output_frames = model_output["note"].shape[0]+ last_frame_s = note_creation.model_frames_to_time(n_model_output_frames)[-1]+ np.testing.assert_allclose(last_frame_s, audio_length_s, atol=2 * ANNOTATION_HOP)+expected_model_output = np.load(RESOURCES_PATH / "vocadito_10" / "model_output.npz", allow_pickle=True)["arr_0"].item()