audio2midi 0.6.0__tar.gz → 0.8.0__tar.gz

{audio2midi-0.6.0 → audio2midi-0.8.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: audio2midi
-Version: 0.6.0
+Version: 0.8.0
 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'
@@ -91,7 +97,7 @@ Description-Content-Type: text/markdown
 ---
 
 ```bash
-pip install audio2midi[all] audio2midi[pop2piano] audio2midi[violin_pitch_detector] audio2midi[crepe_pitch_detector] audio2midi[crepe_pitch_detector_tf] audio2midi[melodia_pitch_detector] audio2midi[basic_pitch_pitch_detector] audio2midi[librosa_pitch_detector]
+pip install audio2midi[all] audio2midi[pop2piano] audio2midi[violin_pitch_detector] audio2midi[crepe_pitch_detector] audio2midi[crepe_pitch_detector_tf] audio2midi[melodia_pitch_detector] audio2midi[basic_pitch_pitch_detector] audio2midi[librosa_pitch_detector] audio2midi[magenta_music_transcription]
 ```
 ---
 
@@ -140,6 +146,7 @@ from audio2midi.basic_pitch_pitch_detector import BasicPitch
 from audio2midi.crepe_pitch_detector import Crepe
 from audio2midi.violin_pitch_detector import Violin_Pitch_Det
 from audio2midi.pop2piano import Pop2Piano
+from audio2midi.magenta_music_transcription import Magenta
 from torch import device as Device
 from torch.cuda import is_available as cuda_is_available
 device = Device("cuda" if cuda_is_available() else "cpu")
@@ -147,6 +154,7 @@ Crepe().predict(audio_path)
 Pop2Piano(device=device).predict(audio_path)
 Violin_Pitch_Det(device=device).predict(audio_path)
 BasicPitch(device=device).predict(audio_path)
+Magenta().predict(audio_path)
 ```
 
 ---
@@ -160,6 +168,7 @@ from audio2midi.melodia_pitch_detector import Melodia
 from audio2midi.pop2piano import Pop2Piano
 from audio2midi.violin_pitch_detector import Violin_Pitch_Det
 from audio2midi.mt3_music_transcription import MT3
+from audio2midi.magenta_music_transcription import Magenta
 from os import environ
 from huggingface_hub import hf_hub_download
 from shutil import unpack_archive
@@ -214,11 +223,12 @@ gr.TabbedInterface([
     gr.Interface(Guitar_Pitch_Det().predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Number(4,label="mag_exp"),gr.Number(-61,label="Threshold"),gr.Number(6,label="Pre_post_max"),gr.Checkbox(False,label="backtrack"),gr.Checkbox(False,label="round_to_sixteenth"),gr.Number(1024,label="hop_length"),gr.Number(72,label="n_bins"),gr.Number(12,label="bins_per_octave")],gr.File(label="Midi File")),
     gr.Interface(Melodia().predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Number(120,label="BPM",step=30),gr.Number(0.25,label="smoothness",step=0.05,info="Smooth the pitch sequence with a median filter of the provided duration (in seconds)."),gr.Number(0.1,label="minimum duration",step=0.1,info="Minimum allowed duration for note (in seconds). Shorter notes will be removed."),gr.Number(128,label="HOP")],gr.File(label="Midi File")),
     gr.Interface(BasicPitch(device=device).predict,[gr.Audio(type="filepath", label="Upload Audio"),gr.Number(0.5,label="onset_thresh",info="Minimum amplitude of an onset activation to be considered an onset."),gr.Number(0.3,label="frame_thresh",info="Minimum energy requirement for a frame to be considered present."),gr.Number(127.70,label="min_note_len",info="The minimum allowed note length in milliseconds."),gr.Number(120,label="midi_tempo"),gr.Checkbox(True,label="infer_onsets",info="add additional onsets when there are large differences in frame amplitudes."),gr.Checkbox(True,label="include_pitch_bends",info="include pitch bends."),gr.Checkbox(False,label="multiple_pitch_bends",info="allow overlapping notes in midi file to have pitch bends."),gr.Checkbox(True,label="melodia_trick",info="Use the melodia post-processing step.")],gr.File(label="Download Midi File")),
+    gr.Interface(Magenta().predict,[gr.Audio(type="filepath", label="Upload Audio"),gr.Number(0.5,label="onset_thresh",info="Minimum amplitude of an onset activation to be considered an onset."),gr.Number(3,label="min_note_len",info="The minimum allowed note length"),gr.Number(3,label="gap_tolerance_frames"),gr.Number(4,label="pitch_bend_steps"),gr.Number(1500,label="pitch_bend_depth"),gr.Checkbox(True,label="include_pitch_bends",info="include pitch bends.")],gr.File(label="Download Midi File")),
     gr.Interface(Violin_Pitch_Det(device=device,model_capacity=getenv("violin_model_capacity","full")).predict, [gr.Audio(label="Upload your Audio file",type="filepath"),gr.Number(32,label="Batch size"),gr.Radio(["spotify","tiktok"],value="spotify",label="Post Processing"),gr.Checkbox(True,label="include_pitch_bends")],gr.File(label="Download MIDI file")),
     gr.Interface(Crepe(getenv("crepe_model_capacity","full")).predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Checkbox(False,label="viterbi",info="Apply viterbi smoothing to the estimated pitch curve"),gr.Checkbox(True,label="center"),gr.Number(10,label="step size",info="The step size in milliseconds for running pitch estimation."),gr.Number(0.8,label="minimum confidence"),gr.Number(32,label="batch size")],gr.File(label="Midi File")),
     gr.Interface(CrepeTF(getenv("crepe_model_capacity","full")).predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Checkbox(False,label="viterbi",info="Apply viterbi smoothing to the estimated pitch curve"),gr.Checkbox(True,label="center"),gr.Number(10,label="step size",info="The step size in milliseconds for running pitch estimation."),gr.Number(0.8,label="minimum confidence"),gr.Number(32,label="batch size")],gr.File(label="Midi File")),
     gr.Interface(Pop2Piano(device).predict,[gr.Audio(label="Input Audio",type="filepath"),gr.Number(1, minimum=1, maximum=21, label="Composer"),gr.Number(2,label="Details in Piano"),gr.Number(1,label="Efficiency of Piano"),gr.Radio([1,2,4],label="steps per beat",value=2)],gr.File(label="MIDI File")),
     gr.Interface(MT3(str(Path("mt3_model").absolute())).predict,[gr.Audio(label="Input Audio",type="filepath"),gr.Number(0,label="seed")],gr.File(label="MIDI File")),
     midi_viz_ui
-],["Normal Pitch Detection","Guitar Based Pitch Detection","Melodia","Spotify Pitch Detection","Violin Based Pitch Detection","Crepe Pitch Detection","Crepe Pitch Detection TF","Pop2Piano","MT3","Midi Vizulizer"]).launch()
+],["Normal Pitch Detection","Guitar Based Pitch Detection","Melodia","Spotify Pitch Detection","Magenta Pitch Detection","Violin Based Pitch Detection","Crepe Pitch Detection","Crepe Pitch Detection TF","Pop2Piano","MT3","Midi Vizulizer"]).launch()
 ```

{audio2midi-0.6.0 → audio2midi-0.8.0}/README.md

@@ -5,7 +5,7 @@
 ---
 
 ```bash
-pip install audio2midi[all] audio2midi[pop2piano] audio2midi[violin_pitch_detector] audio2midi[crepe_pitch_detector] audio2midi[crepe_pitch_detector_tf] audio2midi[melodia_pitch_detector] audio2midi[basic_pitch_pitch_detector] audio2midi[librosa_pitch_detector]
+pip install audio2midi[all] audio2midi[pop2piano] audio2midi[violin_pitch_detector] audio2midi[crepe_pitch_detector] audio2midi[crepe_pitch_detector_tf] audio2midi[melodia_pitch_detector] audio2midi[basic_pitch_pitch_detector] audio2midi[librosa_pitch_detector] audio2midi[magenta_music_transcription]
 ```
 ---
 
@@ -54,6 +54,7 @@ from audio2midi.basic_pitch_pitch_detector import BasicPitch
 from audio2midi.crepe_pitch_detector import Crepe
 from audio2midi.violin_pitch_detector import Violin_Pitch_Det
 from audio2midi.pop2piano import Pop2Piano
+from audio2midi.magenta_music_transcription import Magenta
 from torch import device as Device
 from torch.cuda import is_available as cuda_is_available
 device = Device("cuda" if cuda_is_available() else "cpu")
@@ -61,6 +62,7 @@ Crepe().predict(audio_path)
 Pop2Piano(device=device).predict(audio_path)
 Violin_Pitch_Det(device=device).predict(audio_path)
 BasicPitch(device=device).predict(audio_path)
+Magenta().predict(audio_path)
 ```
 
 ---
@@ -74,6 +76,7 @@ from audio2midi.melodia_pitch_detector import Melodia
 from audio2midi.pop2piano import Pop2Piano
 from audio2midi.violin_pitch_detector import Violin_Pitch_Det
 from audio2midi.mt3_music_transcription import MT3
+from audio2midi.magenta_music_transcription import Magenta
 from os import environ
 from huggingface_hub import hf_hub_download
 from shutil import unpack_archive
@@ -128,11 +131,12 @@ gr.TabbedInterface([
     gr.Interface(Guitar_Pitch_Det().predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Number(4,label="mag_exp"),gr.Number(-61,label="Threshold"),gr.Number(6,label="Pre_post_max"),gr.Checkbox(False,label="backtrack"),gr.Checkbox(False,label="round_to_sixteenth"),gr.Number(1024,label="hop_length"),gr.Number(72,label="n_bins"),gr.Number(12,label="bins_per_octave")],gr.File(label="Midi File")),
     gr.Interface(Melodia().predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Number(120,label="BPM",step=30),gr.Number(0.25,label="smoothness",step=0.05,info="Smooth the pitch sequence with a median filter of the provided duration (in seconds)."),gr.Number(0.1,label="minimum duration",step=0.1,info="Minimum allowed duration for note (in seconds). Shorter notes will be removed."),gr.Number(128,label="HOP")],gr.File(label="Midi File")),
     gr.Interface(BasicPitch(device=device).predict,[gr.Audio(type="filepath", label="Upload Audio"),gr.Number(0.5,label="onset_thresh",info="Minimum amplitude of an onset activation to be considered an onset."),gr.Number(0.3,label="frame_thresh",info="Minimum energy requirement for a frame to be considered present."),gr.Number(127.70,label="min_note_len",info="The minimum allowed note length in milliseconds."),gr.Number(120,label="midi_tempo"),gr.Checkbox(True,label="infer_onsets",info="add additional onsets when there are large differences in frame amplitudes."),gr.Checkbox(True,label="include_pitch_bends",info="include pitch bends."),gr.Checkbox(False,label="multiple_pitch_bends",info="allow overlapping notes in midi file to have pitch bends."),gr.Checkbox(True,label="melodia_trick",info="Use the melodia post-processing step.")],gr.File(label="Download Midi File")),
+    gr.Interface(Magenta().predict,[gr.Audio(type="filepath", label="Upload Audio"),gr.Number(0.5,label="onset_thresh",info="Minimum amplitude of an onset activation to be considered an onset."),gr.Number(3,label="min_note_len",info="The minimum allowed note length"),gr.Number(3,label="gap_tolerance_frames"),gr.Number(4,label="pitch_bend_steps"),gr.Number(1500,label="pitch_bend_depth"),gr.Checkbox(True,label="include_pitch_bends",info="include pitch bends.")],gr.File(label="Download Midi File")),
     gr.Interface(Violin_Pitch_Det(device=device,model_capacity=getenv("violin_model_capacity","full")).predict, [gr.Audio(label="Upload your Audio file",type="filepath"),gr.Number(32,label="Batch size"),gr.Radio(["spotify","tiktok"],value="spotify",label="Post Processing"),gr.Checkbox(True,label="include_pitch_bends")],gr.File(label="Download MIDI file")),
     gr.Interface(Crepe(getenv("crepe_model_capacity","full")).predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Checkbox(False,label="viterbi",info="Apply viterbi smoothing to the estimated pitch curve"),gr.Checkbox(True,label="center"),gr.Number(10,label="step size",info="The step size in milliseconds for running pitch estimation."),gr.Number(0.8,label="minimum confidence"),gr.Number(32,label="batch size")],gr.File(label="Midi File")),
     gr.Interface(CrepeTF(getenv("crepe_model_capacity","full")).predict,[gr.Audio(type="filepath",label="Input Audio"),gr.Checkbox(False,label="viterbi",info="Apply viterbi smoothing to the estimated pitch curve"),gr.Checkbox(True,label="center"),gr.Number(10,label="step size",info="The step size in milliseconds for running pitch estimation."),gr.Number(0.8,label="minimum confidence"),gr.Number(32,label="batch size")],gr.File(label="Midi File")),
     gr.Interface(Pop2Piano(device).predict,[gr.Audio(label="Input Audio",type="filepath"),gr.Number(1, minimum=1, maximum=21, label="Composer"),gr.Number(2,label="Details in Piano"),gr.Number(1,label="Efficiency of Piano"),gr.Radio([1,2,4],label="steps per beat",value=2)],gr.File(label="MIDI File")),
     gr.Interface(MT3(str(Path("mt3_model").absolute())).predict,[gr.Audio(label="Input Audio",type="filepath"),gr.Number(0,label="seed")],gr.File(label="MIDI File")),
     midi_viz_ui
-],["Normal Pitch Detection","Guitar Based Pitch Detection","Melodia","Spotify Pitch Detection","Violin Based Pitch Detection","Crepe Pitch Detection","Crepe Pitch Detection TF","Pop2Piano","MT3","Midi Vizulizer"]).launch()
+],["Normal Pitch Detection","Guitar Based Pitch Detection","Melodia","Spotify Pitch Detection","Magenta Pitch Detection","Violin Based Pitch Detection","Crepe Pitch Detection","Crepe Pitch Detection TF","Pop2Piano","MT3","Midi Vizulizer"]).launch()
 ```

{audio2midi-0.6.0 → audio2midi-0.8.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "audio2midi"
-version = "0.6.0"
+version = "0.8.0"
 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-0.8.0/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,gap_tolerance_frames = 3,pitch_bend_depth = 1500,pitch_bend_steps = 4,include_pitch_bends=True,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