torchaudio 0.1.0 → 0.1.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: b527976494325cc12e81342c25d318204d2d7c75bfba7036be4296769cdb30a0
-  data.tar.gz: 2cfde7bd1b0e7a1628818d5bd74657cfbfba6dfa83ef42897f3ad0f98e77f739
+  metadata.gz: 354a8636b2a04fd0ed60db93e2667f630f1db04ab047a6eb6e0c3607190c58d5
+  data.tar.gz: cde6229f8ba416c5a72bbea6d47e26fd987b6f13073f1ca9a1e6239cd62b79c1
 SHA512:
-  metadata.gz: 8e6f34b014340b5ace3193ab589dae75ed0869ab7606402bd4b09de6042299e6f3a118d439dd381491f489ce9552bca4376a7d5b4693dddc3d1c5f5b26540900
-  data.tar.gz: d651c46f5185ceb70ae3d9c90154c77afe29a5c35854d1a9d98913096b7ab9ba39a745242dd268548ca87f9e109b56c96dee9dc5539cf066f9ad0f773eddbdcd
+  metadata.gz: 76c9c637e4f12700c16bbf343104bb46f5e4f04161a3556cd0600ea236222902ac0ba3b9504f4bdb28a9919317a1b791c8f7b4dbcb48b35ed4c674de39a83277
+  data.tar.gz: '019d89cd2231c386c3458447358058618a6de78b77ed5b4d6d279ee0e3879946bda03cf339051136069dd552d6eeaed33bf40defa9d6c90c9a07eb27d68fdce2'

data/CHANGELOG.md

@@ -1,3 +1,8 @@
+## 0.1.1 (2020-08-26)
+
+- Added `save` method
+- Added transforms
+
 ## 0.1.0 (2020-08-24)
 
 - First release

data/README.md

@@ -2,6 +2,8 @@
 
 :fire: An audio library for Torch.rb
 
+[![Build Status](https://travis-ci.org/ankane/torchaudio.svg?branch=master)](https://travis-ci.org/ankane/torchaudio)
+
 ## Installation
 
 First, [install SoX](#sox-installation). For Homebrew, use:
@@ -20,6 +22,40 @@ gem 'torchaudio'
 
 This library follows the [Python API](https://pytorch.org/audio/). Many methods and options are missing at the moment. PRs welcome!
 
+## Tutorial
+
+- [PyTorch tutorial](https://pytorch.org/tutorials/beginner/audio_preprocessing_tutorial.html)
+- [Ruby code](examples/tutorial.rb)
+
+Download the [audio file](https://github.com/pytorch/tutorials/raw/master/_static/img/steam-train-whistle-daniel_simon-converted-from-mp3.wav) and install the [matplotlib](https://github.com/mrkn/matplotlib.rb) gem first.
+
+## Basics
+
+Load a file
+
+```ruby
+waveform, sample_rate = TorchAudio.load("file.wav")
+```
+
+Save a file
+
+```ruby
+TorchAudio.save("new.wave", waveform, sample_rate)
+```
+
+## Transforms
+
+```ruby
+TorchAudio::Transforms::Spectrogram.new.call(waveform)
+```
+
+Supported transforms are:
+
+- MelSpectrogram
+- MuLawDecoding
+- MuLawEncoding
+- Spectrogram
+
 ## Datasets
 
 Load a dataset

data/ext/torchaudio/ext.cpp

@@ -1,33 +1,75 @@
 #include <torchaudio/csrc/sox.h>
 
+#include <rice/Constructor.hpp>
 #include <rice/Module.hpp>
 
 using namespace Rice;
 
+class SignalInfo {
+  sox_signalinfo_t* value = nullptr;
+  public:
+    SignalInfo(Object o) {
+      if (!o.is_nil()) {
+        value = from_ruby<sox_signalinfo_t*>(o);
+      }
+    }
+    operator sox_signalinfo_t*() {
+      return value;
+    }
+};
+
 template<>
 inline
-sox_signalinfo_t* from_ruby<sox_signalinfo_t*>(Object x)
+SignalInfo from_ruby<SignalInfo>(Object x)
 {
-  if (x.is_nil()) {
-    return nullptr;
-  }
-  throw std::runtime_error("Unsupported signalinfo");
+  return SignalInfo(x);
 }
 
+class EncodingInfo {
+  sox_encodinginfo_t* value = nullptr;
+  public:
+    EncodingInfo(Object o) {
+      if (!o.is_nil()) {
+        value = from_ruby<sox_encodinginfo_t*>(o);
+      }
+    }
+    operator sox_encodinginfo_t*() {
+      return value;
+    }
+};
+
 template<>
 inline
-sox_encodinginfo_t* from_ruby<sox_encodinginfo_t*>(Object x)
+EncodingInfo from_ruby<EncodingInfo>(Object x)
 {
-  if (x.is_nil()) {
-    return nullptr;
-  }
-  throw std::runtime_error("Unsupported encodinginfo");
+  return EncodingInfo(x);
 }
 
 extern "C"
 void Init_ext()
 {
   Module rb_mTorchAudio = define_module("TorchAudio");
-  Module rb_mNN = define_module_under(rb_mTorchAudio, "Ext")
-    .define_singleton_method("read_audio_file", &torch::audio::read_audio_file);
+
+  Module rb_mExt = define_module_under(rb_mTorchAudio, "Ext")
+    .define_singleton_method(
+      "read_audio_file",
+      *[](const std::string& file_name, at::Tensor output, bool ch_first, int64_t nframes, int64_t offset, SignalInfo si, EncodingInfo ei, const char* ft) {
+        return torch::audio::read_audio_file(file_name, output, ch_first, nframes, offset, si, ei, ft);
+      })
+    .define_singleton_method(
+      "write_audio_file",
+      *[](const std::string& file_name, const at::Tensor& tensor, SignalInfo si, EncodingInfo ei, const char* file_type) {
+        return torch::audio::write_audio_file(file_name, tensor, si, ei, file_type);
+      });
+
+  Class rb_cSignalInfo = define_class_under<sox_signalinfo_t>(rb_mExt, "SignalInfo")
+    .define_constructor(Constructor<sox_signalinfo_t>())
+    .define_method("rate", *[](sox_signalinfo_t self) { return self.rate; })
+    .define_method("channels", *[](sox_signalinfo_t self) { return self.channels; })
+    .define_method("precision", *[](sox_signalinfo_t self) { return self.precision; })
+    .define_method("length", *[](sox_signalinfo_t self) { return self.length; })
+    .define_method("rate=", *[](sox_signalinfo_t self, sox_rate_t rate) { self.rate = rate; })
+    .define_method("channels=", *[](sox_signalinfo_t self, unsigned channels) { self.channels = channels; })
+    .define_method("precision=", *[](sox_signalinfo_t self, unsigned precision) { self.precision = precision; })
+    .define_method("length=", *[](sox_signalinfo_t self, sox_uint64_t length) { self.length = length; });
 }

data/lib/torchaudio.rb

@@ -14,6 +14,12 @@ require "set"
 # modules
 require "torchaudio/datasets/utils"
 require "torchaudio/datasets/yesno"
+require "torchaudio/functional"
+require "torchaudio/transforms/mel_scale"
+require "torchaudio/transforms/mel_spectrogram"
+require "torchaudio/transforms/mu_law_encoding"
+require "torchaudio/transforms/mu_law_decoding"
+require "torchaudio/transforms/spectrogram"
 require "torchaudio/version"
 
 module TorchAudio
@@ -73,6 +79,70 @@ module TorchAudio
       load(filepath, **kwargs)
     end
 
+    def save(filepath, src, sample_rate, precision: 16, channels_first: true)
+      si = Ext::SignalInfo.new
+      ch_idx = channels_first ? 0 : 1
+      si.rate = sample_rate
+      si.channels = src.dim == 1 ? 1 : src.size(ch_idx)
+      si.length = src.numel
+      si.precision = precision
+      save_encinfo(filepath, src, channels_first: channels_first, signalinfo: si)
+    end
+
+    def save_encinfo(filepath, src, channels_first: true, signalinfo: nil, encodinginfo: nil, filetype: nil)
+      ch_idx, len_idx = channels_first ? [0, 1] : [1, 0]
+
+      # check if save directory exists
+      abs_dirpath = File.dirname(File.expand_path(filepath))
+      unless Dir.exist?(abs_dirpath)
+        raise "Directory does not exist: #{abs_dirpath}"
+      end
+      # check that src is a CPU tensor
+      check_input(src)
+      # Check/Fix shape of source data
+      if src.dim == 1
+        # 1d tensors as assumed to be mono signals
+        src.unsqueeze!(ch_idx)
+      elsif src.dim > 2 || src.size(ch_idx) > 16
+        # assumes num_channels < 16
+        raise ArgumentError, "Expected format where C < 16, but found #{src.size}"
+      end
+      # sox stores the sample rate as a float, though practically sample rates are almost always integers
+      # convert integers to floats
+      if signalinfo
+        if signalinfo.rate && !signalinfo.rate.is_a?(Float)
+          if signalinfo.rate.to_f == signalinfo.rate
+            signalinfo.rate = signalinfo.rate.to_f
+          else
+            raise ArgumentError, "Sample rate should be a float or int"
+          end
+        end
+        # check if the bit precision (i.e. bits per sample) is an integer
+        if signalinfo.precision && ! signalinfo.precision.is_a?(Integer)
+          if signalinfo.precision.to_i == signalinfo.precision
+            signalinfo.precision = signalinfo.precision.to_i
+          else
+            raise ArgumentError, "Bit precision should be an integer"
+          end
+        end
+      end
+      # programs such as librosa normalize the signal, unnormalize if detected
+      if src.min >= -1.0 && src.max <= 1.0
+        src = src * (1 << 31)
+        src = src.long
+      end
+      # set filetype and allow for files with no extensions
+      extension = File.extname(filepath)
+      filetype = extension.length > 0 ? extension[1..-1] : filetype
+      # transpose from C x L -> L x C
+      if channels_first
+        src = src.transpose(1, 0)
+      end
+      # save data to file
+      src = src.contiguous
+      Ext.write_audio_file(filepath, src, signalinfo, encodinginfo, filetype)
+    end
+
     private
 
     def check_input(src)

data/lib/torchaudio/functional.rb

@@ -0,0 +1,285 @@
+module TorchAudio
+  module Functional
+    class << self
+      def spectrogram(waveform, pad, window, n_fft, hop_length, win_length, power, normalized)
+        if pad > 0
+          # TODO add "with torch.no_grad():" back when JIT supports it
+          waveform = Torch::NN::Functional.pad(waveform, [pad, pad], "constant")
+        end
+
+        # pack batch
+        shape = waveform.size
+        waveform = waveform.reshape(-1, shape[-1])
+
+        # default values are consistent with librosa.core.spectrum._spectrogram
+        spec_f = Torch.stft(
+          waveform, n_fft, hop_length: hop_length, win_length: win_length, window: window, center: true, pad_mode: "reflect", normalized: false, onesided: true
+        )
+
+        # unpack batch
+        spec_f = spec_f.reshape(shape[0..-2] + spec_f.shape[-3..-1])
+
+        if normalized
+          spec_f.div!(window.pow(2.0).sum.sqrt)
+        end
+        if power
+          spec_f = complex_norm(spec_f, power: power)
+        end
+
+        spec_f
+      end
+
+      def mu_law_encoding(x, quantization_channels)
+        mu = quantization_channels - 1.0
+        if !x.floating_point?
+          x = x.to(dtype: :float)
+        end
+        mu = Torch.tensor(mu, dtype: x.dtype)
+        x_mu = Torch.sign(x) * Torch.log1p(mu * Torch.abs(x)) / Torch.log1p(mu)
+        x_mu = ((x_mu + 1) / 2 * mu + 0.5).to(dtype: :int64)
+        x_mu
+      end
+
+      def mu_law_decoding(x_mu, quantization_channels)
+        mu = quantization_channels - 1.0
+        if !x_mu.floating_point?
+          x_mu = x_mu.to(dtype: :float)
+        end
+        mu = Torch.tensor(mu, dtype: x_mu.dtype)
+        x = ((x_mu) / mu) * 2 - 1.0
+        x = Torch.sign(x) * (Torch.exp(Torch.abs(x) * Torch.log1p(mu)) - 1.0) / mu
+        x
+      end
+
+      def complex_norm(complex_tensor, power: 1.0)
+        complex_tensor.pow(2.0).sum(-1).pow(0.5 * power)
+      end
+
+      def create_fb_matrix(n_freqs, f_min, f_max, n_mels, sample_rate, norm: nil)
+        if norm && norm != "slaney"
+          raise ArgumentError, "norm must be one of None or 'slaney'"
+        end
+
+        # freq bins
+        # Equivalent filterbank construction by Librosa
+        all_freqs = Torch.linspace(0, sample_rate.div(2), n_freqs)
+
+        # calculate mel freq bins
+        # hertz to mel(f) is 2595. * math.log10(1. + (f / 700.))
+        m_min = 2595.0 * Math.log10(1.0 + (f_min / 700.0))
+        m_max = 2595.0 * Math.log10(1.0 + (f_max / 700.0))
+        m_pts = Torch.linspace(m_min, m_max, n_mels + 2)
+        # mel to hertz(mel) is 700. * (10**(mel / 2595.) - 1.)
+        f_pts = (Torch.pow(10, m_pts / 2595.0) - 1.0) * 700.0
+        # calculate the difference between each mel point and each stft freq point in hertz
+        f_diff = f_pts[1..-1] - f_pts[0...-1]  # (n_mels + 1)
+        slopes = f_pts.unsqueeze(0) - all_freqs.unsqueeze(1)  # (n_freqs, n_mels + 2)
+        # create overlapping triangles
+        zero = Torch.zeros(1)
+        down_slopes = (slopes[0..-1, 0...-2] * -1.0) / f_diff[0...-1]  # (n_freqs, n_mels)
+        up_slopes = slopes[0..-1, 2..-1] / f_diff[1..-1]  # (n_freqs, n_mels)
+        fb = Torch.max(zero, Torch.min(down_slopes, up_slopes))
+
+        if norm && norm == "slaney"
+          # Slaney-style mel is scaled to be approx constant energy per channel
+          enorm = 2.0 / (f_pts[2...(n_mels + 2)] - f_pts[:n_mels])
+          fb *= enorm.unsqueeze(0)
+        end
+
+        fb
+      end
+
+      def compute_deltas(specgram, win_length: 5, mode: "replicate")
+        device = specgram.device
+        dtype = specgram.dtype
+
+        # pack batch
+        shape = specgram.size
+        specgram = specgram.reshape(1, -1, shape[-1])
+
+        raise ArgumentError, "win_length must be >= 3" unless win_length >= 3
+
+        n = (win_length - 1).div(2)
+
+        # twice sum of integer squared
+        denom = n * (n + 1) * (2 * n + 1) / 3
+
+        specgram = Torch::NN::Functional.pad(specgram, [n, n], mode: mode)
+
+        kernel = Torch.arange(-n, n + 1, 1, device: device, dtype: dtype).repeat([specgram.shape[1], 1, 1])
+
+        output = Torch::NN::Functional.conv1d(specgram, kernel, groups: specgram.shape[1]) / denom
+
+        # unpack batch
+        output = output.reshape(shape)
+      end
+
+      def gain(waveform, gain_db: 1.0)
+        return waveform if gain_db == 0
+
+        ratio = 10 ** (gain_db / 20)
+
+        waveform * ratio
+      end
+
+      def dither(waveform, density_function: "TPDF", noise_shaping: false)
+        dithered = _apply_probability_distribution(waveform, density_function: density_function)
+
+        if noise_shaping
+          raise "Not implemented yet"
+          # _add_noise_shaping(dithered, waveform)
+        else
+          dithered
+        end
+      end
+
+      def biquad(waveform, b0, b1, b2, a0, a1, a2)
+        device = waveform.device
+        dtype = waveform.dtype
+
+        output_waveform = lfilter(
+            waveform,
+            Torch.tensor([a0, a1, a2], dtype: dtype, device: device),
+            Torch.tensor([b0, b1, b2], dtype: dtype, device: device)
+        )
+        output_waveform
+      end
+
+      def highpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707)
+        w0 = 2 * Math::PI * cutoff_freq / sample_rate
+        alpha = Math.sin(w0) / 2.0 / q
+
+        b0 = (1 + Math.cos(w0)) / 2
+        b1 = -1 - Math.cos(w0)
+        b2 = b0
+        a0 = 1 + alpha
+        a1 = -2 * Math.cos(w0)
+        a2 = 1 - alpha
+        biquad(waveform, b0, b1, b2, a0, a1, a2)
+      end
+
+      def lowpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707)
+        w0 = 2 * Math::PI * cutoff_freq / sample_rate
+        alpha = Math.sin(w0) / 2 / q
+
+        b0 = (1 - Math.cos(w0)) / 2
+        b1 = 1 - Math.cos(w0)
+        b2 = b0
+        a0 = 1 + alpha
+        a1 = -2 * Math.cos(w0)
+        a2 = 1 - alpha
+        biquad(waveform, b0, b1, b2, a0, a1, a2)
+      end
+
+      def lfilter(waveform, a_coeffs, b_coeffs, clamp: true)
+        # pack batch
+        shape = waveform.size
+        waveform = waveform.reshape(-1, shape[-1])
+
+        raise ArgumentError unless (a_coeffs.size(0) == b_coeffs.size(0))
+        raise ArgumentError unless (waveform.size.length == 2)
+        raise ArgumentError unless (waveform.device == a_coeffs.device)
+        raise ArgumentError unless (b_coeffs.device == a_coeffs.device)
+
+        device = waveform.device
+        dtype = waveform.dtype
+        n_channel, n_sample = waveform.size
+        n_order = a_coeffs.size(0)
+        n_sample_padded = n_sample + n_order - 1
+        raise ArgumentError unless (n_order > 0)
+
+        # Pad the input and create output
+        padded_waveform = Torch.zeros(n_channel, n_sample_padded, dtype: dtype, device: device)
+        padded_waveform[0..-1, (n_order - 1)..-1] = waveform
+        padded_output_waveform = Torch.zeros(n_channel, n_sample_padded, dtype: dtype, device: device)
+
+        # Set up the coefficients matrix
+        # Flip coefficients' order
+        a_coeffs_flipped = a_coeffs.flip([0])
+        b_coeffs_flipped = b_coeffs.flip([0])
+
+        # calculate windowed_input_signal in parallel
+        # create indices of original with shape (n_channel, n_order, n_sample)
+        window_idxs = Torch.arange(n_sample, device: device).unsqueeze(0) + Torch.arange(n_order, device: device).unsqueeze(1)
+        window_idxs = window_idxs.repeat([n_channel, 1, 1])
+        window_idxs += (Torch.arange(n_channel, device: device).unsqueeze(-1).unsqueeze(-1) * n_sample_padded)
+        window_idxs = window_idxs.long
+        # (n_order, ) matmul (n_channel, n_order, n_sample) -> (n_channel, n_sample)
+        input_signal_windows = Torch.matmul(b_coeffs_flipped, Torch.take(padded_waveform, window_idxs))
+
+        input_signal_windows.div!(a_coeffs[0])
+        a_coeffs_flipped.div!(a_coeffs[0])
+        input_signal_windows.t.each_with_index do |o0, i_sample|
+          windowed_output_signal = padded_output_waveform[0..-1, i_sample...(i_sample + n_order)]
+          o0.addmv!(windowed_output_signal, a_coeffs_flipped, alpha: -1)
+          padded_output_waveform[0..-1, i_sample + n_order - 1] = o0
+        end
+
+        output = padded_output_waveform[0..-1, (n_order - 1)..-1]
+
+        if clamp
+          output = Torch.clamp(output, -1.0, 1.0)
+        end
+
+        # unpack batch
+        output = output.reshape(shape[0...-1] + output.shape[-1..-1])
+
+        output
+      end
+
+      private
+
+      def _apply_probability_distribution(waveform, density_function: "TPDF")
+        # pack batch
+        shape = waveform.size
+        waveform = waveform.reshape(-1, shape[-1])
+
+        channel_size = waveform.size[0] - 1
+        time_size = waveform.size[-1] - 1
+
+        random_channel = channel_size > 0 ? Torch.randint(channel_size, [1]).item.to_i : 0
+        random_time = time_size > 0 ? Torch.randint(time_size, [1]).item.to_i : 0
+
+        number_of_bits = 16
+        up_scaling = 2 ** (number_of_bits - 1) - 2
+        signal_scaled = waveform * up_scaling
+        down_scaling = 2 ** (number_of_bits - 1)
+
+        signal_scaled_dis = waveform
+        if density_function == "RPDF"
+          rpdf = waveform[random_channel][random_time] - 0.5
+
+          signal_scaled_dis = signal_scaled + rpdf
+        elsif density_function == "GPDF"
+          # TODO Replace by distribution code once
+          # https://github.com/pytorch/pytorch/issues/29843 is resolved
+          # gaussian = torch.distributions.normal.Normal(torch.mean(waveform, -1), 1).sample()
+
+          num_rand_variables = 6
+
+          gaussian = waveform[random_channel][random_time]
+          (num_rand_variables * [time_size]).each do |ws|
+            rand_chan = Torch.randint(channel_size, [1]).item.to_i
+            gaussian += waveform[rand_chan][Torch.randint(ws, [1]).item.to_i]
+          end
+
+          signal_scaled_dis = signal_scaled + gaussian
+        else
+          # dtype needed for https://github.com/pytorch/pytorch/issues/32358
+          # TODO add support for dtype and device to bartlett_window
+          tpdf = Torch.bartlett_window(time_size + 1).to(signal_scaled.device, dtype: signal_scaled.dtype)
+          tpdf = tpdf.repeat([channel_size + 1, 1])
+          signal_scaled_dis = signal_scaled + tpdf
+        end
+
+        quantised_signal_scaled = Torch.round(signal_scaled_dis)
+        quantised_signal = quantised_signal_scaled / down_scaling
+
+        # unpack batch
+        quantised_signal.reshape(shape[0...-1] + quantised_signal.shape[-1..-1])
+      end
+    end
+  end
+
+  F = Functional
+end

data/lib/torchaudio/transforms/mel_scale.rb

@@ -0,0 +1,39 @@
+module TorchAudio
+  module Transforms
+    class MelScale < Torch::NN::Module
+      def initialize(n_mels: 128, sample_rate: 16000, f_min: 0.0, f_max: nil, n_stft: nil)
+        super()
+        @n_mels = n_mels
+        @sample_rate = sample_rate
+        @f_max = f_max || sample_rate.div(2).to_f
+        @f_min = f_min
+
+        raise ArgumentError, "Require f_min: %f < f_max: %f" % [f_min, @f_max] unless f_min <= @f_max
+
+        fb = n_stft.nil? ? Torch.empty(0) : F.create_fb_matrix(n_stft, @f_min, @f_max, @n_mels, @sample_rate)
+        register_buffer("fb", fb)
+      end
+
+      def forward(specgram)
+        shape = specgram.size
+        specgram = specgram.reshape(-1, shape[-2], shape[-1])
+
+        if @fb.numel == 0
+          tmp_fb = F.create_fb_matrix(specgram.size(1), @f_min, @f_max, @n_mels, @sample_rate)
+          # Attributes cannot be reassigned outside __init__ so workaround
+          @fb.resize!(tmp_fb.size)
+          @fb.copy!(tmp_fb)
+        end
+
+        # (channel, frequency, time).transpose(...) dot (frequency, n_mels)
+        # -> (channel, time, n_mels).transpose(...)
+        mel_specgram = Torch.matmul(specgram.transpose(1, 2), @fb).transpose(1, 2)
+
+        # unpack batch
+        mel_specgram = mel_specgram.reshape(shape[0...-2] + mel_specgram.shape[-2..-1])
+
+        mel_specgram
+      end
+    end
+  end
+end

data/lib/torchaudio/transforms/mel_spectrogram.rb

@@ -0,0 +1,35 @@
+module TorchAudio
+  module Transforms
+    class MelSpectrogram < Torch::NN::Module
+      def initialize(
+        sample_rate: 16000, n_fft: 400, win_length: nil, hop_length: nil, f_min: 0.0,
+        f_max: nil, pad: 0, n_mels: 128, window_fn: Torch.method(:hann_window),
+        power: 2.0, normalized: false, wkwargs: nil
+      )
+
+        super()
+        @sample_rate = sample_rate
+        @n_fft = n_fft
+        @win_length = win_length || n_fft
+        @hop_length = hop_length || @win_length.div(2)
+        @pad = pad
+        @power = power
+        @normalized = normalized
+        @n_mels = n_mels  # number of mel frequency bins
+        @f_max = f_max
+        @f_min = f_min
+        @spectrogram =
+          Spectrogram.new(
+            n_fft: @n_fft, win_length: @win_length, hop_length: @hop_length, pad: @pad,
+            window_fn: window_fn, power: @power, normalized: @normalized, wkwargs: wkwargs
+          )
+        @mel_scale = MelScale.new(n_mels: @n_mels, sample_rate: @sample_rate, f_min: @f_min, f_max: @f_max, n_stft: @n_fft.div(2) + 1)
+      end
+
+      def forward(waveform)
+        specgram = @spectrogram.call(waveform)
+        @mel_scale.call(specgram)
+      end
+    end
+  end
+end

data/lib/torchaudio/transforms/mu_law_decoding.rb

@@ -0,0 +1,14 @@
+module TorchAudio
+  module Transforms
+    class MuLawDecoding < Torch::NN::Module
+      def initialize(quantization_channels: 256)
+        super()
+        @quantization_channels = quantization_channels
+      end
+
+      def forward(x_mu)
+        F.mu_law_decoding(x_mu, @quantization_channels)
+      end
+    end
+  end
+end

data/lib/torchaudio/transforms/mu_law_encoding.rb

@@ -0,0 +1,14 @@
+module TorchAudio
+  module Transforms
+    class MuLawEncoding < Torch::NN::Module
+      def initialize(quantization_channels: 256)
+        super()
+        @quantization_channels = quantization_channels
+      end
+
+      def forward(x)
+        F.mu_law_encoding(x, @quantization_channels)
+      end
+    end
+  end
+end

data/lib/torchaudio/transforms/spectrogram.rb

@@ -0,0 +1,27 @@
+module TorchAudio
+  module Transforms
+    class Spectrogram < Torch::NN::Module
+      def initialize(
+        n_fft: 400, win_length: nil, hop_length: nil, pad: 0,
+        window_fn: Torch.method(:hann_window), power: 2.0, normalized: false, wkwargs: nil
+      )
+
+        super()
+        @n_fft = n_fft
+        # number of FFT bins. the returned STFT result will have n_fft // 2 + 1
+        # number of frequecies due to onesided=True in torch.stft
+        @win_length = win_length || n_fft
+        @hop_length = hop_length || @win_length.div(2) # floor division
+        window = wkwargs.nil? ? window_fn.call(@win_length) : window_fn.call(@win_length, **wkwargs)
+        register_buffer("window", window)
+        @pad = pad
+        @power = power
+        @normalized = normalized
+      end
+
+      def forward(waveform)
+        F.spectrogram(waveform, @pad, @window, @n_fft, @hop_length, @win_length, @power, @normalized)
+      end
+    end
+  end
+end

data/lib/torchaudio/version.rb

@@ -1,3 +1,3 @@
 module TorchAudio
-  VERSION = "0.1.0"
+  VERSION = "0.1.1"
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: torchaudio
 version: !ruby/object:Gem::Version
-  version: 0.1.0
+  version: 0.1.1
 platform: ruby
 authors:
 - Andrew Kane
 autorequire: 
 bindir: bin
 cert_chain: []
-date: 2020-08-25 00:00:00.000000000 Z
+date: 2020-08-26 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: torch-rb
@@ -16,14 +16,14 @@ dependencies:
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 0.3.2
+        version: 0.3.4
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 0.3.2
+        version: 0.3.4
 - !ruby/object:Gem::Dependency
   name: rice
   requirement: !ruby/object:Gem::Requirement
@@ -118,6 +118,12 @@ files:
 - lib/torchaudio.rb
 - lib/torchaudio/datasets/utils.rb
 - lib/torchaudio/datasets/yesno.rb
+- lib/torchaudio/functional.rb
+- lib/torchaudio/transforms/mel_scale.rb
+- lib/torchaudio/transforms/mel_spectrogram.rb
+- lib/torchaudio/transforms/mu_law_decoding.rb
+- lib/torchaudio/transforms/mu_law_encoding.rb
+- lib/torchaudio/transforms/spectrogram.rb
 - lib/torchaudio/version.rb
 homepage: https://github.com/ankane/torchaudio
 licenses: