Build Your Own Voice Recognition Model with Tensorflowby@FlippedCoding
12,292 reads
12,292 reads

Build Your Own Voice Recognition Model with Tensorflow

by MileciaMarch 7th, 2021
Read on Terminal Reader
Read this story w/o Javascript
tldt arrow

Too Long; Didn't Read

While I'm usually a JavaScript person, there are plenty of things that Python makes easier to do. Doing voice recognition with machine learning is one of those.

Company Mentioned

Mention Thumbnail
featured image - Build Your Own Voice Recognition Model with Tensorflow
Milecia HackerNoon profile picture

Voice recognition is a complex problem across a number of industries. Knowing some of the basics around handling audio data and how to classify sound samples is a good thing to have in your data science toolbox.

We're going to go through an example of classifying some sound clips using Tensorflow. By the time you get through this, you'll know enough to be able to build your own voice recognition models. With additional research, you can take these concepts and apply them to larger, more complex audio files.

You can find the full code in this Github repo.

Getting the Data

Gathering data is one of the hard problems in data science. There's so much data available, but not all of it is easy to use in machine learning problems. You have to make sure that the data is clean, labeled, and complete.

To do our example, we're going to use some audio files released by Google.

First, we'll create a new Conducto pipeline. This is where you'll be able to build, train, and test your model and share a link with anybody else interested:

# Main Pipeline
def main() -> co.Serial:
    path = "/conducto/data/pipeline"
    root = co.Serial(image = get_image())
    # Get data from keras for testing and training
    root["Get Data"] = co.Exec(run_whole_thing, f"{path}/raw")

    return root

Then we'll start writing the


def run_whole_thing(out_dir):
    os.makedirs(out_dir, exist_ok=True)
    # Set seed for experiment reproducibility
    seed = 55
    data_dir = pathlib.Path("data/mini_speech_commands")

Next, we need to set up the directory to hold the audio files:

if not data_dir.exists():
    # Get the files from external source and put them in an accessible directory

Pre-Processing the Data

Now that we have our data in the right directory, we can split it into training, test, and validation datasets.

First, we need to write a few functions to help pre-process the data so that it'll work in our model.

We need the data in a format our algorithm can understand. We'll be using a convolutional neural network, so the data needs to be transformed into images.

This first function will convert the binary audio file into a tensor:

# Convert the binary audio file to a tensor
def decode_audio(audio_binary):
    audio, _ =
    return tf.squeeze(audio, axis=-1)

Since we have a tensor we can work with that has the raw data, we need to get the labels to match them. That's what the following function does by getting the label for an audio file from the file path:

# Get the label (yes, no, up, down, etc) for an audio file.
def get_label(file_path):
   parts = tf.strings.split(file_path, os.path.sep)
   return parts[-2]

Next, we need to associate the audio files with the correct labels. We're doing this and returning a tuple that Tensorflow can work with:

# Create a tuple that has the labeled audio files
def get_waveform_and_label(file_path):
    label = get_label(file_path)
    audio_binary =
    waveform = decode_audio(audio_binary)
    return waveform, label
We briefly mentioned using the convolutional neural network (CNN) algorithm earlier. This is one of the ways we can handle a voice recognition model like this is. Typically CNNs work really well on image data and help decrease pre-processing time.

We're going to take advantage of that by converting our audio files into spectrograms. A spectrogram is an image of a spectrum of frequencies. If you take a look at an audio file, you'll see it's just frequency data. So we're going to write a function that converts our audio data into images:

# Convert audio files to images
def get_spectrogram(waveform):
    # Padding for files with less than 16000 samples
    zero_padding = tf.zeros([16000] - tf.shape(waveform), dtype=tf.float32)
    # Concatenate audio with padding so that all audio clips will be of the same length
    waveform = tf.cast(waveform, tf.float32)
    equal_length = tf.concat([waveform, zero_padding], 0)
    spectrogram = tf.signal.stft(
        equal_length, frame_length=255, frame_step=128)
    spectrogram = tf.abs(spectrogram)

    return spectrogram

Now that we have formatted our data as images, we need to apply the correct labels to those images. This is similar to what we did for the original audio files:

# Label the images created from the audio files and return a tuple
def get_spectrogram_and_label_id(audio, label):
    spectrogram = get_spectrogram(audio)
    spectrogram = tf.expand_dims(spectrogram, -1)
    label_id = tf.argmax(label == commands)
    return spectrogram, label_id

The last helper function we need is the one that will handle all of the above operations for any set of audio files we pass it:

# Preprocess any audio files
def preprocess_dataset(files, autotune, commands):
   # Creates the dataset
   files_ds =
   # Matches audio files with correct labels
   output_ds =,
   # Matches audio file images to the correct labels
   output_ds =
       get_spectrogram_and_label_id,  num_parallel_calls=autotune)
   return output_ds

Now that we have all of these helper functions, we get to split the data.

Splitting the Data into Datasets

Converting audio files to images helps make the data easier to process with a CNN and that's why we wrote all of those helper functions. We'll do a couple of things to make splitting the data more simple.

First, we'll get a list of all of the potential commands for the audio files that we'll use in a few other places in the code:

# Get all of the commands for the audio files
commands = np.array(
commands = commands[commands != '']

Then we'll get a list of all of the files in the data directory and shuffle them so we can assign random values to each of the datasets we need:

# Get a list of all the files in the directory
filenames = + '/*/*')

# Shuffle the file names so that random bunches can be used as the training, testing, and validation sets
filenames = tf.random.shuffle(filenames)

# Create the list of files for training data
train_files = filenames[:6400]

# Create the list of files for validation data
validation_files = filenames[6400: 6400 + 800]

# Create the list of files for test data
test_files = filenames[-800:]

Now we have our training, validation, and test files clearly separated so we can go ahead and pre-process these files to get them ready to build and test our model. We're using autotune here to tune the value of our parameters dynamically at runtime:

autotune =

This first example is just to show how the pre-processing works and it gives us the

value that we'll need in a bit:

# Get the converted audio files for training the model
files_ds =
waveform_ds =
    get_waveform_and_label, num_parallel_calls=autotune)
spectrogram_ds =
    get_spectrogram_and_label_id, num_parallel_calls=autotune)

Since you've seen what it's like to go through the pre-processing steps, we can go ahead and use the helper function to handle this for all of the datasets:

# Preprocess the training, test, and validation datasets
train_ds = preprocess_dataset(train_files, autotune, commands)
validation_ds = preprocess_dataset(
    validation_files, autotune, commands)
test_ds = preprocess_dataset(test_files, autotune, commands)

We want to set a number of training examples that run in each iteration of the epochs so we'll set a batch size:

# Batch datasets for training and validation
batch_size = 64
train_ds = train_ds.batch(batch_size)
validation_ds = validation_ds.batch(batch_size)

Lastly, we can reduce the amount of latency in training our model by taking advantage of caching:

# Reduce latency while training
train_ds = train_ds.cache().prefetch(autotune)
validation_ds = validation_ds.cache().prefetch(autotune)

Our datasets are finally in a form that we can train the model with.

Building the Model

Since our datasets are clearly defined, we can go ahead and build the model. We'll be using a CNN to create our model so we'll need to get the shape of the data to get the correct shape for our layers. Then we go ahead build the model sequentially:

# Build model
for spectrogram, _ in spectrogram_ds.take(1):
    input_shape = spectrogram.shape

num_labels = len(commands)

norm_layer = preprocessing.Normalization()
norm_layer.adapt( x, _: x))

model = models.Sequential([
    preprocessing.Resizing(32, 32),
    layers.Conv2D(32, 3, activation='relu'),
    layers.Conv2D(64, 3, activation='relu'),
    layers.Dense(128, activation='relu'),


We do some configuration on the model so that it gives us the best accuracy possible:

# Configure built model with losses and metrics

The model is built so now all that's left is training it.

Training the Model

After all of the work did pre-processing the data and building the model, training is relatively simple. We determine how many epochs we want to run with our training and validation datasets:

# Finally train the model and return info about each epoch
    callbacks=tf.keras.callbacks.EarlyStopping(verbose=1, patience=2),

That's it! The model has been trained and now we just need to test it.

Testing the Model

Now that we have a model with roughly 83% accuracy, it's time we test how well it performs on new data. So we take our test dataset and split the audio files from the labels:

# Test the model
test_audio = []
test_labels = []

for audio, label in test_ds:

test_audio = np.array(test_audio)
test_labels = np.array(test_labels)

Then we take the audio data and use it in our model to see if it predicts the correct label:

# See how accurate the model is when making predictions on the test dataset
y_pred = np.argmax(model.predict(test_audio), axis=1)
y_true = test_labels

test_acc = sum(y_pred == y_true) / len(y_true)

print(f'Test set accuracy: {test_acc:.0%}')

Finishing the Pipeline

There's just a tiny bit of code that you'll need to finish your pipeline and make it shareable with anyone. This defines the image that will be used in this Conducto pipeline and handles the file execution:

# Pipeline Helper functions
def get_image():
    return co.Image(
        reqs_py=["conducto", "tensorflow", "keras"],

if __name__ == "__main__":

Now you can run

python --local
in your terminal and it should spin up a link to a new Conducto pipeline. If you don't have an account, you can make one for free here.


This is one of the ways you can solve an audio processing problem, but it can be much more complex depending on what data you're trying to analyze. Building it in a pipeline makes it easy to share with coworkers and get help or feedback if you run into bugs.

Previously published here.