Poor results

[johnbuts]

Hey everyone, thanks in advanced for the help.

So I wanted to use some of the instrument detection models, and was not impressed by the results. I fed it a wav file that just had saxophone playing for around a minute and 10 seconds. Here is the code and output I got:

`
from essentia.standard import MonoLoader, TensorflowPredictEffnetDiscogs, TensorflowPredict2D
import pandas as pd

audio = MonoLoader(filename="other_sax.wav", sampleRate=75000, resampleQuality=4)()
embedding_model = TensorflowPredictEffnetDiscogs(graphFilename="discogs-effnet-bs64-1.pb", output="PartitionedCall:1")
embeddings = embedding_model(audio)

model = TensorflowPredict2D(graphFilename="mtg_jamendo_instrument-discogs-effnet-1.pb")
predictions = model(embeddings)

instruments = [
'accordion', 'acousticbassguitar', 'acousticguitar', 'bass', 'beat', 'bell', 'bongo', 'brass',
'cello', 'clarinet', 'classicalguitar', 'computer', 'doublebass', 'drummachine', 'drums',
'electricguitar', 'electricpiano', 'flute', 'guitar', 'harmonica', 'harp', 'horn', 'keyboard',
'oboe', 'orchestra', 'organ', 'pad', 'percussion', 'piano', 'pipeorgan', 'rhodes', 'sampler',
'saxophone', 'strings', 'synthesizer', 'trombone', 'trumpet', 'viola', 'violin', 'voice'
]

df = pd.DataFrame(predictions, columns=instruments)
instrument_sums = df.sum()

top_5_instruments = instrument_sums.sort_values(ascending=False).head(5)

print(top_5_instruments)
`

output:
synthesizer 218.628510
piano 175.836365
drums 113.429985
cello 85.704750
flute 83.436066

Please tell me what I'm doing wrong, thanks.

[palonso]

Hi @johnbuts
The problem with your script is that MonoLoader's sampleRate parameter should match the model's expected sample rate (16000).

[johnbuts]

drums 53.502373
bass 42.751842
electricguitar 40.419640
piano 34.878933
guitar 32.601994

that didn't seem to help, been toying around with the sample rate, nothing really seems to help it. Is maybe my code wrong? like the order of the instrumens or something?

[johnbuts]

Its like no matter the instrument, its like always really high on sythesizer and piano, like ill put in violin and get this:
synthesizer 70.941162
piano 69.683495
drums 65.955116
electricguitar 63.488628
guitar 62.260086

and then ill put in a saxophone track and get this:
drums 73.532722
synthesizer 72.120262
piano 70.608063
bass 69.452019
electricguitar 66.740585

[csipapicsa]

Have your tried to set the resampleQuality=0

resampleQuality (integer ∈ [0, 4], default = 1) :
the resampling quality, 0 for best quality, 4 for fast linear approximation