About half a yr in the past, this weblog featured a submit, written by Daniel Falbel, on the best way to use Keras to categorise items of spoken language. The article acquired quite a lot of consideration and never surprisingly, questions arose the best way to apply that code to completely different datasets. We’ll take this as a motivation to discover in additional depth the preprocessing achieved in that submit: If we all know why the enter to the community appears to be like the best way it appears to be like, we will modify the mannequin specification appropriately if want be.
In case you could have a background in speech recognition, and even normal sign processing, for you the introductory a part of this submit will most likely not include a lot information. Nonetheless, you may nonetheless have an interest within the code half, which exhibits the best way to do issues like creating spectrograms with present variations of TensorFlow. In case you don’t have that background, we’re inviting you on a (hopefully) fascinating journey, barely pertaining to one of many larger mysteries of this universe.
We’ll use the identical dataset as Daniel did in his submit, that’s, model 1 of the Google speech instructions dataset(Warden 2018) The dataset consists of ~ 65,000 WAV information, of size one second or much less. Every file is a recording of considered one of thirty phrases, uttered by completely different audio system.
The aim then is to coach a community to discriminate between spoken phrases. How ought to the enter to the community look? The WAV information include amplitudes of sound waves over time. Listed here are just a few examples, comparable to the phrases chicken, down, sheila, and visible:
A sound wave is a sign extending in time, analogously to how what enters our visible system extends in house. At every cutoff date, the present sign depends on its previous. The plain structure to make use of in modeling it thus appears to be a recurrent neural community.
Nonetheless, the data contained within the sound wave might be represented in another approach: particularly, utilizing the frequencies that make up the sign.
Right here we see a sound wave (high) and its frequency illustration (backside).
Within the time illustration (known as the time area), the sign consists of consecutive amplitudes over time. Within the frequency area, it’s represented as magnitudes of various frequencies. It could seem as one of many biggest mysteries on this world that you could convert between these two with out lack of data, that’s: Each representations are primarily equal!
Conversion from the time area to the frequency area is finished utilizing the Fourier rework; to transform again, the Inverse Fourier Rework is used. There exist various kinds of Fourier transforms relying on whether or not time is considered as steady or discrete, and whether or not the sign itself is steady or discrete. Within the “actual world,” the place often for us, actual means digital as we’re working with digitized indicators, the time area in addition to the sign are represented as discrete and so, the Discrete Fourier Rework (DFT) is used. The DFT itself is computed utilizing the FFT (Quick Fourier Rework) algorithm, leading to important speedup over a naive implementation.
Trying again on the above instance sound wave, it’s a compound of 4 sine waves, of frequencies 8Hz, 16Hz, 32Hz, and 64Hz, whose amplitudes are added and displayed over time. The compound wave right here is assumed to increase infinitely in time. In contrast to speech, which adjustments over time, it may be characterised by a single enumeration of the magnitudes of the frequencies it’s composed of. So right here the spectrogram, the characterization of a sign by magnitudes of constituent frequencies various over time, appears to be like primarily onedimensional.
Nonetheless, once we ask Praat to create a spectrogram of considered one of our instance sounds (a seven), it might appear to be this:
Right here we see a twodimensional picture of frequency magnitudes over time (greater magnitudes indicated by darker coloring). This twodimensional illustration could also be fed to a community, instead of the onedimensional amplitudes. Accordingly, if we resolve to take action we’ll use a convnet as a substitute of an RNN.
Spectrograms will look completely different relying on how we create them. We’ll check out the important choices in a minute. First although, let’s see what we can’t at all times do: ask for all frequencies that have been contained within the analog sign.
Above, we stated that each representations, time area and frequency area, have been primarily equal. In our digital actual world, that is solely true if the sign we’re working with has been digitized accurately, or as that is generally phrased, if it has been “correctly sampled.”
Take speech for instance: As an analog sign, speech per se is steady in time; for us to have the ability to work with it on a pc, it must be transformed to occur in discrete time. This conversion of the unbiased variable (time in our case, house in e.g. picture processing) from steady to discrete known as sampling.
On this technique of discretization, a vital determination to be made is the sampling fee to make use of. The sampling fee needs to be a minimum of double the very best frequency within the sign. If it’s not, lack of data will happen. The best way that is most frequently put is the opposite approach spherical: To protect all data, the analog sign might not include frequencies above onehalf the sampling fee. This frequency – half the sampling fee – known as the Nyquist fee.
If the sampling fee is just too low, aliasing takes place: Greater frequencies alias themselves as decrease frequencies. Which means not solely can’t we get them, additionally they corrupt the magnitudes of corresponding decrease frequencies they’re being added to. Right here’s a schematic instance of how a highfrequency sign might alias itself as being lowerfrequency. Think about the highfrequency wave being sampled at integer factors (gray circles) solely:
Within the case of the speech instructions dataset, all sound waves have been sampled at 16 kHz. Which means once we ask Praat for a spectogram, we should always not ask for frequencies greater than 8kHz. Here’s what occurs if we ask for frequencies as much as 16kHz as a substitute – we simply don’t get them:
Now let’s see what choices we do have when creating spectrograms.
Within the above easy sine wave instance, the sign stayed fixed over time. Nonetheless in speech utterances, the magnitudes of constituent frequencies change over time. Ideally thus, we’d have an actual frequency illustration for each cutoff date. As an approximation to this superb, the sign is split into overlapping home windows, and the Fourier rework is computed for every time slice individually. That is referred to as the Quick Time Fourier Rework (STFT).
After we compute the spectrogram through the STFT, we have to inform it what dimension home windows to make use of, and the way huge to make the overlap. The longer the home windows we use, the higher the decision we get within the frequency area. Nonetheless, what we acquire in decision there, we lose within the time area, as we’ll have fewer home windows representing the sign. This can be a normal precept in sign processing: Decision within the time and frequency domains are inversely associated.
To make this extra concrete, let’s once more take a look at a easy instance. Right here is the spectrogram of an artificial sine wave, composed of two elements at 1000 Hz and 1200 Hz. The window size was left at its (Praat) default, 5 milliseconds:
We see that with a brief window like that, the 2 completely different frequencies are mangled into one within the spectrogram. Now enlarge the window to 30 milliseconds, and they’re clearly differentiated:
The above spectrogram of the phrase “seven” was produced utilizing Praats default of 5 milliseconds. What occurs if we use 30 milliseconds as a substitute?
We get higher frequency decision, however on the value of decrease decision within the time area. The window size used throughout preprocessing is a parameter we’d need to experiment with later, when coaching a community.
One other enter to the STFT to play with is the kind of window used to weight the samples in a time slice. Right here once more are three spectrograms of the above recording of seven, utilizing, respectively, a Hamming, a Hann, and a Gaussian window:
Whereas the spectrograms utilizing the Hann and Gaussian home windows don’t look a lot completely different, the Hamming window appears to have launched some artifacts.
Preprocessing choices don’t finish with the spectrogram. A preferred transformation utilized to the spectrogram is conversion to mel scale, a scale based mostly on how people really understand variations in pitch. We don’t elaborate additional on this right here, however we do briefly touch upon the respective TensorFlow code beneath, in case you’d wish to experiment with this. Up to now, coefficients remodeled to Mel scale have typically been additional processed to acquire the socalled MelFrequency Cepstral Coefficients (MFCCs). Once more, we simply present the code. For wonderful studying on Mel scale conversion and MFCCs (together with the explanation why MFCCs are much less usually used these days) see this submit by Haytham Fayek.
Again to our unique process of speech classification. Now that we’ve gained a little bit of perception in what’s concerned, let’s see the best way to carry out these transformations in TensorFlow.
Code might be represented in snippets based on the performance it supplies, so we might instantly map it to what was defined conceptually above. A whole instance is accessible right here. The entire instance builds on Daniel’s unique code as a lot as potential, with two exceptions:

The code runs in keen in addition to in static graph mode. In case you resolve you solely ever want keen mode, there are just a few locations that may be simplified. That is partly associated to the truth that in keen mode, TensorFlow operations instead of tensors return values, which we are able to instantly move on to TensorFlow features anticipating values, not tensors. As well as, much less conversion code is required when manipulating intermediate values in R.

With TensorFlow 1.13 being launched any day, and preparations for TF 2.0 operating at full velocity, we wish the code to necessitate as few modifications as potential to run on the subsequent main model of TF. One huge distinction is that there’ll now not be a
contrib
module. Within the unique submit,contrib
was used to learn within the.wav
information in addition to compute the spectrograms. Right here, we’ll use performance fromtf.audio
andtf.sign
as a substitute.
All operations proven beneath will run inside tf.dataset
code, which on the R facet is achieved utilizing the tfdatasets
package deal. To elucidate the person operations, we take a look at a single file, however later we’ll additionally show the info generator as an entire.
For stepping via particular person strains, it’s at all times useful to have keen mode enabled, independently of whether or not finally we’ll execute in keen or graph mode:
We decide a random .wav
file and decode it utilizing tf$audio$decode_wav
.This may give us entry to 2 tensors: the samples themselves, and the sampling fee.
fname < "information/speech_commands_v0.01/chicken/00b01445_nohash_0.wav"
wav < tf$audio$decode_wav(tf$read_file(fname))
wav$sample_rate
incorporates the sampling fee. As anticipated, it’s 16000, or 16kHz:
sampling_rate < wav$sample_rate %>% as.numeric()
sampling_rate
16000
The samples themselves are accessible as wav$audio
, however their form is (16000, 1), so we’ve got to transpose the tensor to get the standard (batch_size, variety of samples) format we want for additional processing.
samples < wav$audio
samples < samples %>% tf$transpose(perm = c(1L, 0L))
samples
tf.Tensor(
[[0.00750732 0.04653931 0.02041626 ... 0.01004028 0.01300049
0.00250244]], form=(1, 16000), dtype=float32)
Computing the spectogram
To compute the spectrogram, we use tf$sign$stft
(the place stft stands for Quick Time Fourier Rework). stft
expects three nondefault arguments: Apart from the enter sign itself, there are the window dimension, frame_length
, and the stride to make use of when figuring out the overlapping home windows, frame_step
. Each are expressed in items of variety of samples
. So if we resolve on a window size of 30 milliseconds and a stride of 10 milliseconds …
window_size_ms < 30
window_stride_ms < 10
… we arrive on the following name:
samples_per_window < sampling_rate * window_size_ms/1000
stride_samples < sampling_rate * window_stride_ms/1000
stft_out < tf$sign$stft(
samples,
frame_length = as.integer(samples_per_window),
frame_step = as.integer(stride_samples)
)
Inspecting the tensor we acquired again, stft_out
, we see, for our single enter wave, a matrix of 98 x 257 advanced values:
tf.Tensor(
[[[ 1.03279948e04+0.00000000e+00j 1.95371482e046.41121820e04j
1.60833192e03+4.97534114e04j ... 3.61620914e051.07343149e04j
2.82576875e055.88812982e05j 2.66879797e05+0.00000000e+00j]
...
]],
form=(1, 98, 257), dtype=complex64)
Right here 98 is the variety of intervals, which we are able to compute prematurely, based mostly on the variety of samples in a window and the dimensions of the stride:
257 is the variety of frequencies we obtained magnitudes for. By default, stft
will apply a Quick Fourier Rework of dimension smallest energy of two larger or equal to the variety of samples in a window, after which return the fft_length / 2 + 1 distinctive elements of the FFT: the zerofrequency time period and the positivefrequency phrases.
In our case, the variety of samples in a window is 480. The closest enclosing energy of two being 512, we find yourself with 512/2 + 1 = 257 coefficients. This too we are able to compute prematurely:
Again to the output of the STFT. Taking the elementwise magnitude of the advanced values, we get hold of an power spectrogram:
magnitude_spectrograms < tf$abs(stft_out)
If we cease preprocessing right here, we’ll often need to log rework the values to raised match the sensitivity of the human auditory system:
log_magnitude_spectrograms = tf$log(magnitude_spectrograms + 1e6)
Mel spectrograms and MelFrequency Cepstral Coefficients (MFCCs)
If as a substitute we select to make use of Mel spectrograms, we are able to get hold of a metamorphosis matrix that can convert the unique spectrograms to Mel scale:
lower_edge_hertz < 0
upper_edge_hertz < 2595 * log10(1 + (sampling_rate/2)/700)
num_mel_bins < 64L
num_spectrogram_bins < magnitude_spectrograms$form[1]$worth
linear_to_mel_weight_matrix < tf$sign$linear_to_mel_weight_matrix(
num_mel_bins,
num_spectrogram_bins,
sampling_rate,
lower_edge_hertz,
upper_edge_hertz
)
Making use of that matrix, we get hold of a tensor of dimension (batch_size, variety of intervals, variety of Mel coefficients) which once more, we are able to logcompress if we wish:
mel_spectrograms < tf$tensordot(magnitude_spectrograms, linear_to_mel_weight_matrix, 1L)
log_mel_spectrograms < tf$log(mel_spectrograms + 1e6)
Only for completeness’ sake, lastly we present the TensorFlow code used to additional compute MFCCs. We don’t embrace this within the full instance as with MFCCs, we would want a distinct community structure.
num_mfccs < 13
mfccs < tf$sign$mfccs_from_log_mel_spectrograms(log_mel_spectrograms)[, , 1:num_mfccs]
Accommodating differentlength inputs
In our full instance, we decide the sampling fee from the primary file learn, thus assuming all recordings have been sampled on the similar fee. We do permit for various lengths although. For instance in our dataset, had we used this file, simply 0.65 seconds lengthy, for demonstration functions:
fname < "information/speech_commands_v0.01/chicken/1746d7b6_nohash_0.wav"
we’d have ended up with simply 63 intervals within the spectrogram. As we’ve got to outline a hard and fast input_size
for the primary conv layer, we have to pad the corresponding dimension to the utmost potential size, which is n_periods
computed above. The padding really takes place as a part of dataset definition. Let’s rapidly see dataset definition as an entire, leaving out the potential era of Mel spectrograms.
data_generator < perform(df,
window_size_ms,
window_stride_ms) {
# assume sampling fee is similar in all samples
sampling_rate <
tf$audio$decode_wav(tf$read_file(tf$reshape(df$fname[[1]], listing()))) %>% .$sample_rate
samples_per_window < (sampling_rate * window_size_ms) %/% 1000L
stride_samples < (sampling_rate * window_stride_ms) %/% 1000L
n_periods <
tf$form(
tf$vary(
samples_per_window %/% 2L,
16000L  samples_per_window %/% 2L,
stride_samples
)
)[1] + 1L
n_fft_coefs <
(2 ^ tf$ceil(tf$log(
tf$forged(samples_per_window, tf$float32)
) / tf$log(2)) /
2 + 1L) %>% tf$forged(tf$int32)
ds < tensor_slices_dataset(df) %>%
dataset_shuffle(buffer_size = buffer_size)
ds < ds %>%
dataset_map(perform(obs) {
wav <
tf$audio$decode_wav(tf$read_file(tf$reshape(obs$fname, listing())))
samples < wav$audio
samples < samples %>% tf$transpose(perm = c(1L, 0L))
stft_out < tf$sign$stft(samples,
frame_length = samples_per_window,
frame_step = stride_samples)
magnitude_spectrograms < tf$abs(stft_out)
log_magnitude_spectrograms < tf$log(magnitude_spectrograms + 1e6)
response < tf$one_hot(obs$class_id, 30L)
enter < tf$transpose(log_magnitude_spectrograms, perm = c(1L, 2L, 0L))
listing(enter, response)
})
ds < ds %>%
dataset_repeat()
ds %>%
dataset_padded_batch(
batch_size = batch_size,
padded_shapes = listing(tf$stack(listing(
n_periods, n_fft_coefs,1L
)),
tf$fixed(1L, form = form(1L))),
drop_remainder = TRUE
)
}
The logic is similar as described above, solely the code has been generalized to work in keen in addition to graph mode. The padding is taken care of by dataset_padded_batch(), which must be informed the utmost variety of intervals and the utmost variety of coefficients.
Time for experimentation
Constructing on the full instance, now’s the time for experimentation: How do completely different window sizes have an effect on classification accuracy? Does transformation to the mel scale yield improved outcomes? You may also need to attempt passing a nondefault window_fn
to stft
(the default being the Hann window) and see how that impacts the outcomes. And naturally, the easy definition of the community leaves quite a lot of room for enchancment.
Talking of the community: Now that we’ve gained extra perception into what’s contained in a spectrogram, we’d begin asking, is a convnet actually an enough resolution right here? Usually we use convnets on photos: twodimensional information the place each dimensions symbolize the identical type of data. Thus with photos, it’s pure to have sq. filter kernels. In a spectrogram although, the time axis and the frequency axis symbolize basically various kinds of data, and it’s not clear in any respect that we should always deal with them equally. Additionally, whereas in photos, the interpretation invariance of convnets is a desired characteristic, this isn’t the case for the frequency axis in a spectrogram.
Closing the circle, we uncover that as a result of deeper information concerning the topic area, we’re in a greater place to motive about (hopefully) profitable community architectures. We depart it to the creativity of our readers to proceed the search…