Easy Audio Classification with Keras

Introduction

On this tutorial we’ll construct a deep studying mannequin to categorise phrases. We are going to use tfdatasets to deal with knowledge IO and pre-processing, and Keras to construct and practice the mannequin.

We are going to use the Speech Instructions dataset which consists of 65,000 one-second audio information of individuals saying 30 totally different phrases. Every file incorporates a single spoken English phrase. The dataset was launched by Google beneath CC License.

Our mannequin is a Keras port of the TensorFlow tutorial on Easy Audio Recognition which in flip was impressed by Convolutional Neural Networks for Small-footprint Key phrase Recognizing. There are different approaches to the speech recognition process, like recurrent neural networks, dilated (atrous) convolutions or Studying from Between-class Examples for Deep Sound Recognition.

The mannequin we’ll implement right here shouldn’t be the state-of-the-art for audio recognition methods, that are far more advanced, however is comparatively easy and quick to coach. Plus, we present methods to effectively use tfdatasets to preprocess and serve knowledge.

Audio illustration

Many deep studying fashions are end-to-end, i.e. we let the mannequin study helpful representations immediately from the uncooked knowledge. Nevertheless, audio knowledge grows very quick – 16,000 samples per second with a really wealthy construction at many time-scales. With a purpose to keep away from having to take care of uncooked wave sound knowledge, researchers often use some type of characteristic engineering.

Each sound wave could be represented by its spectrum, and digitally it may be computed utilizing the Quick Fourier Rework (FFT).

A standard option to symbolize audio knowledge is to interrupt it into small chunks, which often overlap. For every chunk we use the FFT to calculate the magnitude of the frequency spectrum. The spectra are then mixed, facet by facet, to type what we name a spectrogram.

It’s additionally frequent for speech recognition methods to additional remodel the spectrum and compute the Mel-Frequency Cepstral Coefficients. This transformation takes into consideration that the human ear can’t discern the distinction between two carefully spaced frequencies and neatly creates bins on the frequency axis. A fantastic tutorial on MFCCs could be discovered right here.

After this process, we now have a picture for every audio pattern and we are able to use convolutional neural networks, the usual structure sort in picture recognition fashions.

Downloading

First, let’s obtain knowledge to a listing in our mission. You possibly can both obtain from this hyperlink (~1GB) or from R with:

dir.create("knowledge")

obtain.file(
  url = "http://obtain.tensorflow.org/knowledge/speech_commands_v0.01.tar.gz", 
  destfile = "knowledge/speech_commands_v0.01.tar.gz"
)

untar("knowledge/speech_commands_v0.01.tar.gz", exdir = "knowledge/speech_commands_v0.01")

Contained in the knowledge listing we can have a folder referred to as speech_commands_v0.01. The WAV audio information inside this listing are organised in sub-folders with the label names. For instance, all one-second audio information of individuals talking the phrase “mattress” are contained in the mattress listing. There are 30 of them and a particular one referred to as _background_noise_ which incorporates varied patterns that might be blended in to simulate background noise.

Importing

On this step we’ll checklist all audio .wav information right into a tibble with 3 columns:

• fname: the file title;
• class: the label for every audio file;
• class_id: a novel integer quantity ranging from zero for every class – used to one-hot encode the courses.

This shall be helpful to the following step once we will create a generator utilizing the tfdatasets bundle.

Generator

We are going to now create our Dataset, which within the context of tfdatasets, provides operations to the TensorFlow graph with the intention to learn and pre-process knowledge. Since they’re TensorFlow ops, they’re executed in C++ and in parallel with mannequin coaching.

The generator we’ll create shall be answerable for studying the audio information from disk, creating the spectrogram for every one and batching the outputs.

Let’s begin by creating the dataset from slices of the knowledge.body with audio file names and courses we simply created.

Now, let’s outline the parameters for spectrogram creation. We have to outline window_size_ms which is the dimensions in milliseconds of every chunk we’ll break the audio wave into, and window_stride_ms, the gap between the facilities of adjoining chunks:

window_size_ms <- 30
window_stride_ms <- 10

Now we’ll convert the window dimension and stride from milliseconds to samples. We’re contemplating that our audio information have 16,000 samples per second (1000 ms).

window_size <- as.integer(16000*window_size_ms/1000)
stride <- as.integer(16000*window_stride_ms/1000)

We are going to acquire different portions that shall be helpful for spectrogram creation, just like the variety of chunks and the FFT dimension, i.e., the variety of bins on the frequency axis. The perform we’re going to use to compute the spectrogram doesn’t permit us to vary the FFT dimension and as an alternative by default makes use of the primary energy of two larger than the window dimension.

We are going to now use dataset_map which permits us to specify a pre-processing perform for every commentary (line) of our dataset. It’s on this step that we learn the uncooked audio file from disk and create its spectrogram and the one-hot encoded response vector.

# shortcuts to used TensorFlow modules.
audio_ops <- tf$contrib$framework$python$ops$audio_ops

ds <- ds %>%
  dataset_map(perform(obs) {
    
    # a great way to debug when constructing tfdatsets pipelines is to make use of a print
    # assertion like this:
    # print(str(obs))
    
    # decoding wav information
    audio_binary <- tf$read_file(tf$reshape(obs$fname, form = checklist()))
    wav <- audio_ops$decode_wav(audio_binary, desired_channels = 1)
    
    # create the spectrogram
    spectrogram <- audio_ops$audio_spectrogram(
      wav$audio, 
      window_size = window_size, 
      stride = stride,
      magnitude_squared = TRUE
    )
    
    # normalization
    spectrogram <- tf$log(tf$abs(spectrogram) + 0.01)
    
    # transferring channels to final dim
    spectrogram <- tf$transpose(spectrogram, perm = c(1L, 2L, 0L))
    
    # remodel the class_id right into a one-hot encoded vector
    response <- tf$one_hot(obs$class_id, 30L)
    
    checklist(spectrogram, response)
  }) 

Now, we’ll specify how we would like batch observations from the dataset. We’re utilizing dataset_shuffle since we wish to shuffle observations from the dataset, in any other case it might comply with the order of the df object. Then we use dataset_repeat with the intention to inform TensorFlow that we wish to preserve taking observations from the dataset even when all observations have already been used. And most significantly right here, we use dataset_padded_batch to specify that we would like batches of dimension 32, however they need to be padded, ie. if some commentary has a special dimension we pad it with zeroes. The padded form is handed to dataset_padded_batch by way of the padded_shapes argument and we use NULL to state that this dimension doesn’t should be padded.

ds <- ds %>% 
  dataset_shuffle(buffer_size = 100) %>%
  dataset_repeat() %>%
  dataset_padded_batch(
    batch_size = 32, 
    padded_shapes = checklist(
      form(n_chunks, fft_size, NULL), 
      form(NULL)
    )
  )

That is our dataset specification, however we would wish to rewrite all of the code for the validation knowledge, so it’s good follow to wrap this right into a perform of the information and different essential parameters like window_size_ms and window_stride_ms. Beneath, we’ll outline a perform referred to as data_generator that may create the generator relying on these inputs.

data_generator <- perform(df, batch_size, shuffle = TRUE, 
                           window_size_ms = 30, window_stride_ms = 10) {
  
  window_size <- as.integer(16000*window_size_ms/1000)
  stride <- as.integer(16000*window_stride_ms/1000)
  fft_size <- as.integer(2^trunc(log(window_size, 2)) + 1)
  n_chunks <- size(seq(window_size/2, 16000 - window_size/2, stride))
  
  ds <- tensor_slices_dataset(df)
  
  if (shuffle) 
    ds <- ds %>% dataset_shuffle(buffer_size = 100)  
  
  ds <- ds %>%
    dataset_map(perform(obs) {
      
      # decoding wav information
      audio_binary <- tf$read_file(tf$reshape(obs$fname, form = checklist()))
      wav <- audio_ops$decode_wav(audio_binary, desired_channels = 1)
      
      # create the spectrogram
      spectrogram <- audio_ops$audio_spectrogram(
        wav$audio, 
        window_size = window_size, 
        stride = stride,
        magnitude_squared = TRUE
      )
      
      spectrogram <- tf$log(tf$abs(spectrogram) + 0.01)
      spectrogram <- tf$transpose(spectrogram, perm = c(1L, 2L, 0L))
      
      # remodel the class_id right into a one-hot encoded vector
      response <- tf$one_hot(obs$class_id, 30L)
      
      checklist(spectrogram, response)
    }) %>%
    dataset_repeat()
  
  ds <- ds %>% 
    dataset_padded_batch(batch_size, checklist(form(n_chunks, fft_size, NULL), form(NULL)))
  
  ds
}

Now, we are able to outline coaching and validation knowledge turbines. It’s value noting that executing this received’t really compute any spectrogram or learn any file. It should solely outline within the TensorFlow graph the way it ought to learn and pre-process knowledge.

set.seed(6)
id_train <- pattern(nrow(df), dimension = 0.7*nrow(df))

ds_train <- data_generator(
  df[id_train,], 
  batch_size = 32, 
  window_size_ms = 30, 
  window_stride_ms = 10
)
ds_validation <- data_generator(
  df[-id_train,], 
  batch_size = 32, 
  shuffle = FALSE, 
  window_size_ms = 30, 
  window_stride_ms = 10
)

To truly get a batch from the generator we may create a TensorFlow session and ask it to run the generator. For instance:

sess <- tf$Session()
batch <- next_batch(ds_train)
str(sess$run(batch))

Checklist of two
 $ : num [1:32, 1:98, 1:257, 1] -4.6 -4.6 -4.61 -4.6 -4.6 ...
 $ : num [1:32, 1:30] 0 0 0 0 0 0 0 0 0 0 ...

Every time you run sess$run(batch) you must see a special batch of observations.

Mannequin definition

Now that we all know how we’ll feed our knowledge we are able to give attention to the mannequin definition. The spectrogram could be handled like a picture, so architectures which are generally utilized in picture recognition duties ought to work nicely with the spectrograms too.

We are going to construct a convolutional neural community just like what we now have constructed right here for the MNIST dataset.

The enter dimension is outlined by the variety of chunks and the FFT dimension. Like we defined earlier, they are often obtained from the window_size_ms and window_stride_ms used to generate the spectrogram.

We are going to now outline our mannequin utilizing the Keras sequential API:

mannequin <- keras_model_sequential()
mannequin %>%  
  layer_conv_2d(input_shape = c(n_chunks, fft_size, 1), 
                filters = 32, kernel_size = c(3,3), activation = 'relu') %>% 
  layer_max_pooling_2d(pool_size = c(2, 2)) %>% 
  layer_conv_2d(filters = 64, kernel_size = c(3,3), activation = 'relu') %>% 
  layer_max_pooling_2d(pool_size = c(2, 2)) %>% 
  layer_conv_2d(filters = 128, kernel_size = c(3,3), activation = 'relu') %>% 
  layer_max_pooling_2d(pool_size = c(2, 2)) %>% 
  layer_conv_2d(filters = 256, kernel_size = c(3,3), activation = 'relu') %>% 
  layer_max_pooling_2d(pool_size = c(2, 2)) %>% 
  layer_dropout(price = 0.25) %>% 
  layer_flatten() %>% 
  layer_dense(models = 128, activation = 'relu') %>% 
  layer_dropout(price = 0.5) %>% 
  layer_dense(models = 30, activation = 'softmax')

We used 4 layers of convolutions mixed with max pooling layers to extract options from the spectrogram pictures and a pair of dense layers on the prime. Our community is relatively easy when in comparison with extra superior architectures like ResNet or DenseNet that carry out very nicely on picture recognition duties.

Now let’s compile our mannequin. We are going to use categorical cross entropy because the loss perform and use the Adadelta optimizer. It’s additionally right here that we outline that we are going to take a look at the accuracy metric throughout coaching.

mannequin %>% compile(
  loss = loss_categorical_crossentropy,
  optimizer = optimizer_adadelta(),
  metrics = c('accuracy')
)

Mannequin becoming

Now, we’ll match our mannequin. In Keras we are able to use TensorFlow Datasets as inputs to the fit_generator perform and we’ll do it right here.

mannequin %>% fit_generator(
  generator = ds_train,
  steps_per_epoch = 0.7*nrow(df)/32,
  epochs = 10, 
  validation_data = ds_validation, 
  validation_steps = 0.3*nrow(df)/32
)

Epoch 1/10
1415/1415 [==============================] - 87s 62ms/step - loss: 2.0225 - acc: 0.4184 - val_loss: 0.7855 - val_acc: 0.7907
Epoch 2/10
1415/1415 [==============================] - 75s 53ms/step - loss: 0.8781 - acc: 0.7432 - val_loss: 0.4522 - val_acc: 0.8704
Epoch 3/10
1415/1415 [==============================] - 75s 53ms/step - loss: 0.6196 - acc: 0.8190 - val_loss: 0.3513 - val_acc: 0.9006
Epoch 4/10
1415/1415 [==============================] - 75s 53ms/step - loss: 0.4958 - acc: 0.8543 - val_loss: 0.3130 - val_acc: 0.9117
Epoch 5/10
1415/1415 [==============================] - 75s 53ms/step - loss: 0.4282 - acc: 0.8754 - val_loss: 0.2866 - val_acc: 0.9213
Epoch 6/10
1415/1415 [==============================] - 76s 53ms/step - loss: 0.3852 - acc: 0.8885 - val_loss: 0.2732 - val_acc: 0.9252
Epoch 7/10
1415/1415 [==============================] - 75s 53ms/step - loss: 0.3566 - acc: 0.8991 - val_loss: 0.2700 - val_acc: 0.9269
Epoch 8/10
1415/1415 [==============================] - 76s 54ms/step - loss: 0.3364 - acc: 0.9045 - val_loss: 0.2573 - val_acc: 0.9284
Epoch 9/10
1415/1415 [==============================] - 76s 53ms/step - loss: 0.3220 - acc: 0.9087 - val_loss: 0.2537 - val_acc: 0.9323
Epoch 10/10
1415/1415 [==============================] - 76s 54ms/step - loss: 0.2997 - acc: 0.9150 - val_loss: 0.2582 - val_acc: 0.9323

The mannequin’s accuracy is 93.23%. Let’s discover ways to make predictions and try the confusion matrix.

Making predictions

We are able to use thepredict_generator perform to make predictions on a brand new dataset. Let’s make predictions for our validation dataset.
The predict_generator perform wants a step argument which is the variety of instances the generator shall be referred to as.

We are able to calculate the variety of steps by realizing the batch dimension, and the dimensions of the validation dataset.

df_validation <- df[-id_train,]
n_steps <- nrow(df_validation)/32 + 1

We are able to then use the predict_generator perform:

predictions <- predict_generator(
  mannequin, 
  ds_validation, 
  steps = n_steps
  )
str(predictions)

num [1:19424, 1:30] 1.22e-13 7.30e-19 5.29e-10 6.66e-22 1.12e-17 ...

It will output a matrix with 30 columns – one for every phrase and n_steps*batch_size variety of rows. Notice that it begins repeating the dataset on the finish to create a full batch.

We are able to compute the expected class by taking the column with the very best likelihood, for instance.

courses <- apply(predictions, 1, which.max) - 1

A pleasant visualization of the confusion matrix is to create an alluvial diagram:

library(dplyr)
library(alluvial)
x <- df_validation %>%
  mutate(pred_class_id = head(courses, nrow(df_validation))) %>%
  left_join(
    df_validation %>% distinct(class_id, class) %>% rename(pred_class = class),
    by = c("pred_class_id" = "class_id")
  ) %>%
  mutate(right = pred_class == class) %>%
  depend(pred_class, class, right)

alluvial(
  x %>% choose(class, pred_class),
  freq = x$n,
  col = ifelse(x$right, "lightblue", "crimson"),
  border = ifelse(x$right, "lightblue", "crimson"),
  alpha = 0.6,
  cover = x$n < 20
)

We are able to see from the diagram that essentially the most related mistake our mannequin makes is to categorise “tree” as “three”. There are different frequent errors like classifying “go” as “no”, “up” as “off”. At 93% accuracy for 30 courses, and contemplating the errors we are able to say that this mannequin is fairly cheap.

The saved mannequin occupies 25Mb of disk area, which is cheap for a desktop however is probably not on small gadgets. We may practice a smaller mannequin, with fewer layers, and see how a lot the efficiency decreases.

In speech recognition duties its additionally frequent to do some type of knowledge augmentation by mixing a background noise to the spoken audio, making it extra helpful for actual purposes the place it’s frequent to produce other irrelevant sounds taking place within the surroundings.

The total code to breed this tutorial is obtainable right here.