This is a small set of python scripts which will analyze an audio file and detect some specific sound...(like a snap, clap, etc)
Python 3+ should be installed
Pythonic packages
- matplotlib
- numpy
- scipy
- pyaudio
Note: These packages should be installed for python 3
- Record a sample
- Do some preliminary analysis on sample audio
- Record test audio which contains 'sample' sounds
- Split test into small audio clips (which "may" contain the sound we want to detect)
- Analyze above audio clips; identify if it contains the sound we are expecting
Note: These scripts don't have anything specific for recording. You may try audacity if you want to quickly get started. However, most operating systems would probably already have a sound recorder installed. You can use that too.
When analyzing the sample we make note of the following:
- Peak or dominant frequencies in the frequency spectogram (using FFT and some interpolation)
- We study the sample's wave form graph to find a spike in the audio output. (This is later used for splitting the test audio into smaller clips)
- We analyze the spectrograms of the test audio clips and compare it with our sample's spectrogram. This helps in identifying if the test clip has the desired sound.
- Peak Extraction:
The spectrogram is a XY scatter plot. For filtering peak amplitudes, we find the pairs (x and y values) whose y-value is greater than 50% of its maximum.
- Normalization:
In the dataset obtained above, we normalize both x and y values, relatively. Hence all x values are subtracted with its minimum. Similar process is done for y values.
Now we have a simpler discrete dataset where the values start from x=0. The "width" of this dataset (x range) is the maximum x value.
- Comparision:
We simply do a ratio comparision of the widths, with the constraint that the larger is always the numerator, and the other quantity is the denominator.
This ratio should approx be in the interval [1, 2)
Note: We are only comparing "widths". This approximately let us detect desired sound. A more rigorous approximation might require comparing the y-values via some approach.
For step 2 in the workflow above use the following scripts:
Plots the wave form graph of a wav file.
Usage:
./plotwav.py <wav_file> Int16
Plots the frequency spectrogram
Usage:
./fftplot.py <wav_file> Int16
Creates the frequency spectrogram discrete dataset as a CSV file. It gets stored in the samples folder
Usage:
./fftanalyze.py <wav_file> Int16
For step 4 in the workflow use the following scripts:
This script is usually run against a test audio file.
It splits the file into short audio clips which may contain our desired sound. They are stored in the processed folder
Usage:
./split_wav2.py <wav_file> Int16
Important: For each experiment you may want to edit this file to match the parameters you noted down in step 2. Manually change the values for the following variables: THRESHOLD and SAMPLE_LENGTH
For step 5 use the following scripts:
This will process wav files in the processed folder and generate fft datasets (in csv format). Output gets stored in the same folder.
Usage:
./batch_generate_wav_fft.py
Compare the fft datasets in the processed folder against a sample-fft dataset, and report the width ratio.
Usage:
./batch_analyze.py <sample_fft_csv_dataset_file>
The other scripts are here for any further diagnosis you may require. The ones that start with batch_ will mainly work with files in the processed folder.
Not mandatory to follow, but helps if you want to rigorously experiment.
- All sample audio files affix the word
samplein the filename E.g
data/clap_sample_472ms_16bit_mono.wav
- All test audio files affix the word
testin the filename - All audio files must be wav format with 16 bit data, mono channel. It is also good to incorporate the length of the audio clip, and, bit-depth for easily being able to distinguish.