---
license: mit
task_categories:
  - audio-classification
modalities:
  - audio
language:
  - en
tags:
  - MNIST
  - Audio
  - Classification
  - Audio Classification
pretty_name: Audio MNIST
size_categories:
- 10K<n<100K
---

# Audio MNIST

Based on [`AudioMNIST`](https://github.com/soerenab/AudioMNIST).

## Generation of the Parquet File

Given the path to the `data` folder from the [source](https://github.com/soerenab/AudioMNIST) as `audioMNISTFolderPath`:

```python
import numpy as np
import pandas as pd

import json

# Load all wave files in AudioMNIST dataset
# Parse each file name as <digit>_<speaker>_<index>.wav

dfData = pd.DataFrame(columns = ['Digit', 'Speaker', 'Index', 'SampleRate', 'NumSamples', 'Accent', 'Age', 'Gender', 'NativeSpeaker', 'Continent', 'Country', 'City', 'Room'])

with open(os.path.join(audioMNISTFolderPath, 'audioMNIST_meta.txt'), 'r') as f:
    dMetadata = json.load(f)

lWaveFiles = []

lFolders = os.listdir(audioMNISTFolderPath)
lFolders = [fld for fld in lFolders if os.path.isdir(os.path.join(audioMNISTFolderPath, fld))]
lFolders.sort()

fileIdx = -1
for fld in lFolders:
    lFiles = os.listdir(os.path.join(audioMNISTFolderPath, fld))
    lFiles = [f for f in lFiles if f.endswith('.wav')]
    lFiles.sort()
    print(f'Folder {fld}: {len(lFiles)} files')
    for f in lFiles:
        fileIdx += 1
        # Parse File Name
        digitIdx, speakerIdx, recIdx = f[:-4].split('_')
        sampleRate, vAudioData = sp.io.wavfile.read(os.path.join(audioMNISTFolderPath, fld, f))
        lWaveFiles.append(vAudioData)

        dfData.loc[fileIdx, 'Digit']       = int(digitIdx)
        dfData.loc[fileIdx, 'Speaker']     = int(speakerIdx)
        dfData.loc[fileIdx, 'Index']       = int(recIdx)
        dfData.loc[fileIdx, 'SampleRate']  = int(sampleRate)
        dfData.loc[fileIdx, 'NumSamples']  = int(len(vAudioData))
        # Parse Metadata
        metaIdx = f'{int(speakerIdx):02d}'
        dfData.loc[fileIdx, 'Accent']      = dMetadata[metaIdx]['accent']
        dfData.loc[fileIdx, 'Age']         = int(dMetadata[metaIdx]['age'])
        dfData.loc[fileIdx, 'Gender']      = dMetadata[metaIdx]['gender']
        dfData.loc[fileIdx, 'NativeSpeaker'] = dMetadata[metaIdx]['native speaker']
        # Parse Continent, Country, City
        locationStr = dMetadata[metaIdx]['origin']
        # Remove spaces, split by ','
        locationStr = locationStr.replace(' ', '')
        contStr, countryStr, cityStr = locationStr.split(',')
        dfData.loc[fileIdx, 'Continent']   = contStr
        dfData.loc[fileIdx, 'Country']     = countryStr
        dfData.loc[fileIdx, 'City']        = cityStr
        dfData.loc[fileIdx, 'Room']        = dMetadata[metaIdx]['recordingroom']

# Generate DataFrame of the Audio Data
maxSignals = dfData.shape[0]
maxNumSamples = dfData['NumSamples'].max()

mA = np.zeros((maxSignals, maxNumSamples), dtype = np.int16)
for ii, vA in enumerate(lWaveFiles):
    mA[ii, :len(vA)] = vA

dfAudio = pd.DataFrame(data = mA, columns = [f'{sampleIdx:d}' for sampleIdx in range(maxNumSamples)])

# Generate the AudioMNIST Data Frame
dfAudioMnist = pd.concat([dfData, dfAudio], axis = 1)

# Set the Type per column
dfAudioMnist['Digit'] = dfAudioMnist['Digit'].astype(np.int8)
dfAudioMnist['Speaker'] = dfAudioMnist['Speaker'].astype(np.int8)
dfAudioMnist['Index'] = dfAudioMnist['Index'].astype(np.int32)
dfAudioMnist['SampleRate'] = dfAudioMnist['SampleRate'].astype(np.int32)
dfAudioMnist['NumSamples'] = dfAudioMnist['NumSamples'].astype(np.int32)
dfAudioMnist['Age'] = dfAudioMnist['Age'].astype(np.int32)
dfAudioMnist['Gender'] = dfAudioMnist['Gender'].map({'male': 'Male', 'female': 'Female'})
dfAudioMnist['NativeSpeaker'] = dfAudioMnist['NativeSpeaker'].map({'yes': True, 'no': False})
dfAudioMnist['NativeSpeaker'] = dfAudioMnist['NativeSpeaker'].astype(bool)

# Export to Parquet
dfAudioMnist.to_parquet(os.path.join(audioMNISTFolderPath, 'AudioMNIST.parquet'), index = False)
```