Improving the features we use for classification

Pythonで学ぶMachine Learningによる時系列データ解析

Chris Holdgraf

Fellow, Berkeley Institute for Data Science

The auditory envelope

Smooth the data to calculate the auditory envelope
Related to the total amount of audio energy present at each moment of time

Smoothing over time

Instead of averaging over all time, we can do a local average
This is called smoothing your timeseries
It removes short-term noise, while retaining the general pattern

Smoothing your data

Calculating a rolling window statistic

# Audio is a Pandas DataFrame
print(audio.shape)  
# (n_times, n_audio_files)

(5000, 20)

# Smooth our data by taking the rolling mean in a window of 50 samples
window_size = 50
windowed = audio.rolling(window=window_size)
audio_smooth = windowed.mean()

Calculating the auditory envelope

First rectify your audio, then smooth it

  audio_rectified = audio.apply(np.abs)
  audio_envelope = audio_rectified.rolling(50).mean()

Feature engineering the envelope

# Calculate several features of the envelope, one per sound
envelope_mean = np.mean(audio_envelope, axis=0)
envelope_std = np.std(audio_envelope, axis=0)
envelope_max = np.max(audio_envelope, axis=0)

# Create our training data for a classifier
X = np.column_stack([envelope_mean, envelope_std, envelope_max])

Preparing our features for scikit-learn

X = np.column_stack([envelope_mean, envelope_std, envelope_max])
y = labels.reshape(-1, 1)

Cross validation for classification

cross_val_score automates the process of:
- Splitting data into training / validation sets
- Fitting the model on training data
- Scoring it on validation data
- Repeating this process

Using cross_val_score

from sklearn.model_selection import cross_val_score

model = LinearSVC()
scores = cross_val_score(model, X, y, cv=3) 
print(scores)

[0.60911642 0.59975305 0.61404035]

Auditory features: The Tempogram

We can summarize more complex temporal information with timeseries-specific functions
librosa is a great library for auditory and timeseries feature engineering
Here we'll calculate the tempogram, which estimates the tempo of a sound over time
We can calculate summary statistics of tempo in the same way that we can for the envelope

Computing the tempogram

# Import librosa and calculate the tempo of a 1-D sound array
import librosa as lr
audio_tempo = lr.beat.tempo(y=audio, sr=sfreq, 
                            hop_length=2**6)

Let's practice!

Pythonで学ぶMachine Learningによる時系列データ解析