audio2midi 0.6.0__tar.gz → 0.7.0__tar.gz

{audio2midi-0.6.0 → audio2midi-0.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: audio2midi
-Version: 0.6.0
+Version: 0.7.0
 Summary: Audio To Midi
 Author-email: dummyjenil <dummyjenil@gmail.com>
 Provides-Extra: all
@@ -52,6 +52,12 @@ Provides-Extra: librosa-pitch-detector
 Requires-Dist: librosa; extra == 'librosa-pitch-detector'
 Requires-Dist: numpy; extra == 'librosa-pitch-detector'
 Requires-Dist: pretty-midi-fix; extra == 'librosa-pitch-detector'
+Provides-Extra: magenta-music-transcription
+Requires-Dist: huggingface-hub; extra == 'magenta-music-transcription'
+Requires-Dist: librosa; extra == 'magenta-music-transcription'
+Requires-Dist: numpy; extra == 'magenta-music-transcription'
+Requires-Dist: pretty-midi-fix; extra == 'magenta-music-transcription'
+Requires-Dist: tensorflow; extra == 'magenta-music-transcription'
 Provides-Extra: melodia-pitch-detector
 Requires-Dist: huggingface-hub; extra == 'melodia-pitch-detector'
 Requires-Dist: librosa; extra == 'melodia-pitch-detector'

{audio2midi-0.6.0 → audio2midi-0.7.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "audio2midi"
-version = "0.6.0"
+version = "0.7.0"
 description = "Audio To Midi"
 readme = "README.md"
 authors = [
@@ -18,6 +18,7 @@ crepe_pitch_detector = ["librosa", "numpy","pretty_midi_fix","hmmlearn","tensorf
 violin_pitch_detector = ["librosa", "numpy","pretty_midi_fix","scipy","torchaudio","torch","mir_eval","huggingface_hub"]
 pop2piano = ["librosa", "numpy==1.26.4","pretty_midi_fix","transformers","essentia","torch","scipy","resampy","pretty_midi","huggingface_hub"]
 mt3_music_transcription = ["nest_asyncio", "mt3-audio2midi"]
+magenta_music_transcription = ["librosa","numpy","pretty_midi_fix","tensorflow","huggingface_hub"]
 all = ["librosa", "numpy==1.26.4","pretty_midi_fix","transformers","essentia","torch","scipy","torchaudio","torch","mir_eval","hmmlearn","tensorflow","keras","vamp","nnAudio","resampy","pretty_midi","huggingface_hub","nest_asyncio", "mt3-audio2midi"]
 
 [build-system]

audio2midi-0.7.0/src/audio2midi/magenta_music_transcription.py

@@ -0,0 +1,76 @@
+from typing import Callable
+import numpy as np
+import pretty_midi_fix
+import tensorflow
+import librosa.feature.rhythm
+from huggingface_hub import hf_hub_download
+import librosa
+
+
+def endpitch(pitch, endFrame,volProb,intervals,MIN_NOTE_FRAMES,frameLenSecs,PITCH_BEND_ENABLED,pitch_bend_steps,pitch_bend_depth,track):
+    startFrame = intervals[pitch]
+    if endFrame - startFrame < MIN_NOTE_FRAMES:
+        return
+    vol = volProb[startFrame, pitch]
+    if vol < 0 or vol > 1:
+        return
+    start_time = startFrame * frameLenSecs
+    track.notes.append(pretty_midi_fix.Note(velocity=int(max(0, min(1, vol)) * 80 + 10), pitch=pitch + 21, start=start_time, end=endFrame * frameLenSecs))
+    if PITCH_BEND_ENABLED:
+        for step in range(pitch_bend_steps):
+            track.pitch_bends.append(pretty_midi_fix.PitchBend(pitch=int(np.sin(np.pi * step / (pitch_bend_steps - 1)) * pitch_bend_depth), time=start_time + step * 0.01))
+        track.pitch_bends.append(pretty_midi_fix.PitchBend(pitch=0, time=start_time + 0.05))  # Reset
+    del intervals[pitch]
+
+def model_output_to_notes(model_output,onset_thresh,include_pitch_bends,min_note_len,gap_tolerance_frames,pitch_bend_depth,pitch_bend_steps):
+    actProb , onProb , offProb , volProb , tempo = model_output
+    midi = pretty_midi_fix.PrettyMIDI(initial_tempo=tempo)
+    track = pretty_midi_fix.Instrument(program=40)
+    frameLenSecs = librosa.frames_to_time(1, sr=16000)
+    intervals = {}
+    onsets = (onProb > onset_thresh).astype(np.int8)
+    frames = onsets | (actProb > onset_thresh).astype(np.int8)
+    for i, frame in enumerate(np.vstack([frames, np.zeros(frames.shape[1])])):
+        for pitch, active in enumerate(frame):
+            if active:
+                if pitch not in intervals:
+                    if onsets is None or onsets[i, pitch]:
+                        intervals[pitch] = i
+                elif onsets is not None and onsets[i, pitch] and (i - intervals[pitch] > 2):
+                    endpitch(pitch, i,volProb,intervals,min_note_len,frameLenSecs,include_pitch_bends,pitch_bend_steps,pitch_bend_depth,track)
+                    intervals[pitch] = i
+            elif pitch in intervals:
+                if i + gap_tolerance_frames < frames.shape[0] and np.any(frames[i:i + gap_tolerance_frames, pitch]):
+                    continue  # Don't end the note yet
+                endpitch(pitch, i,volProb,intervals,min_note_len,frameLenSecs,include_pitch_bends,pitch_bend_steps,pitch_bend_depth,track)
+    midi.instruments.append(track)
+    return midi
+
+class Magenta:
+    def __init__(self,model_path=hf_hub_download("shethjenil/Audio2Midi_Models","magenta.tflite")):
+        self.interp = tensorflow.lite.Interpreter(model_path=model_path)
+        self.interp.allocate_tensors()
+        self.inputLen = self.interp.get_input_details()[0]['shape'][0]
+        self.outputStep = self.interp.get_output_details()[0]['shape'][1] * 512
+
+    def run_inference(self,audio_path,progress_callback):
+        song = librosa.load(audio_path,sr=16000)[0]
+        actProb, onProb, offProb, volProb = np.empty((1, 88)), np.empty((1, 88)), np.empty((1, 88)), np.empty((1, 88))
+        paddedSong = np.append(song, np.zeros(-(song.size - self.inputLen) % self.outputStep, dtype=np.float32))
+        total_size = (paddedSong.size - self.inputLen) // self.outputStep + 1
+        tempo = librosa.feature.rhythm.tempo(y=song, sr=16000).mean()
+        for i in range(total_size):
+            self.interp.set_tensor(self.interp.get_input_details()[0]['index'], paddedSong[i * self.outputStep : i * self.outputStep + self.inputLen])
+            self.interp.invoke()
+            actProb = np.vstack((actProb, self.interp.get_tensor(self.interp.get_output_details()[0]['index'])[0]))
+            onProb  = np.vstack(( onProb, self.interp.get_tensor(self.interp.get_output_details()[1]['index'])[0]))
+            offProb = np.vstack((offProb, self.interp.get_tensor(self.interp.get_output_details()[2]['index'])[0]))
+            volProb = np.vstack((volProb, self.interp.get_tensor(self.interp.get_output_details()[3]['index'])[0]))
+            if progress_callback:
+                progress_callback(i,total_size)
+        return actProb , onProb , offProb , volProb , tempo
+
+
+    def predict(self,audio,onset_thresh=0,min_note_len=3,include_pitch_bends=True,gap_tolerance_frames = 3,pitch_bend_depth = 1500,pitch_bend_steps = 4,progress_callback: Callable[[int, int], None] = None,output_file="output.mid"):
+        model_output_to_notes(self.run_inference(audio,progress_callback),onset_thresh  = onset_thresh,min_note_len  = min_note_len,include_pitch_bends  = include_pitch_bends,pitch_bend_depth=pitch_bend_depth,pitch_bend_steps=pitch_bend_steps,gap_tolerance_frames=gap_tolerance_frames).write(output_file)
+        return output_file