noyes 0.9.2 → 1.0.0

data/lib/ruby_impl/compression.rb

@@ -0,0 +1,214 @@
+module Noyes
+  
+  class ULaw
+    def << data
+      result = data.map do |cmn|
+        cmn.map do |x|
+          e2x = Math::E ** x.abs
+          Math.log(1 + 255 * e2x) / Math.log(1 + e2x)
+        end
+      end
+    end
+  end
+
+  class Compressor
+    def << data
+      @last = Array.new(data.first.size){0.0} unless @last
+      result = data.map do |signal|
+        signal.zip(@last).map do |both|
+          both[0] - both[1]
+        end
+      end
+      @last = data.last
+      result
+    end
+  end
+
+  class DeltaEncoder
+    def initialize dimensions=13
+      @dimensions = dimensions
+      @delta = Array.new dimensions, 0.0
+    end
+
+    def << data
+      data.each_slice(@dimensions).map do |array|
+        array.each_with_index.map do |element, index|
+          current_delta = @delta[index]
+          @delta[index] = element
+          element - current_delta
+        end
+      end
+    end
+  end
+
+  class DeltaDecoder
+    def initialize dimensions=13
+      @dimensions = dimensions
+      @delta = Array.new dimensions, 0.0
+    end
+
+    def << data
+      data.each_slice(@dimensions).map do |array|
+        array.each_with_index.map do |element, index|
+          @delta[index] += element
+        end
+      end
+    end
+  end
+
+  # Splits a float into an array containint its sign bit, its exponent, and its
+  # significand
+  class FloatSplitter
+    def << data
+      data.pack('g*').unpack('N*').map do |bits|
+        signbit = bits >> 31
+        exponent = (bits & 0x7F800000) >> 23
+        significand = bits & 0x007FFFFF
+        [signbit, exponent - 127, significand]
+    end
+    end
+  end
+
+  # Reassembles a float that has been dissassembled using FloatSplitter.
+  class FloatAssembler
+    def << data
+      data.map do |signbit, exponent, significand|
+        bits = signbit << 31 | 127 + exponent << 23 | significand
+      end.pack('N*').unpack('g*')
+    end
+  end
+
+  class BitArray
+    def initialize
+      @array = []
+      @end_bit = 0
+      @start_bit = 0
+    end
+
+    def size
+      @end_bit - @start_bit
+    end
+
+    def empty?
+      size <= 0
+    end
+
+    def to_s
+      @array.pack('N*').unpack('B*').join[@start_bit...@end_bit]
+    end
+
+    # Add a bit to the end of the bit array.  Bits may be anything that
+    # evaluates to either 1 or 0. Anything else is undefined.  If you can't
+    # afford zeros and only have the letter O, amazingly, that works too.
+    def push bit
+      @array.push 0 if @array.size <= @end_bit / 32
+      @array[-1] = set_bit(@array.last, @end_bit % 32) if bit == 1
+      @end_bit +=1
+    end
+
+    # Our bit array is packed into an array of 32 bit integers.  This
+    # function sets the ith bit of an integer.
+    def set_bit integer, i
+      integer | 0x80000000 >> i
+    end
+
+    # Returns the first bit and removes it, shifting all bits by one.
+    def shift
+      return if @array.empty?
+      bit = @array.first & 0x80000000 >> @start_bit == 0 ? 0 : 1
+      if @start_bit == 31
+        @start_bit = 0
+        @end_bit -= 32
+        @array.shift
+      else
+        @start_bit += 1
+      end
+      bit
+    end
+
+    # Returns the ith bit of the array.
+    def [] i
+      @array[i/32] & (0x80000000 >> (i + @start_bit)) != 0 ? 1 : 0
+    end
+  end
+
+  # Takes disassembled floats and compresses them.  We want them dissassembled
+  # because compressing exponents and signs have you unique properties and can
+  # be efficiently compressed with rice coding.  The same is not true of the
+  # significand.
+  class GolombRiceEncoder
+    def initialize m = 8
+      @M = m
+      @b = Math.log2(m).to_i
+    end
+
+    # data is a list of disassembled floats
+    def << data
+      data.map do |b,e,s|
+        exp_sign_combo = b |(e << 1)
+        [interleave(exp_sign_combo), s]
+      end
+    end
+
+    # Map negative nubers to odd postive numbers and postive numbers to even
+    # positive numbers.  We need to do this because negative numbers don't
+    # compress well with rice encoding.
+    def interleave x
+      x < 0 ? 2 * x.abs - 1 : 2 * x.abs
+    end
+
+    # Rice encoding returned as a BitArray.
+    def encode integer_array
+      integer_array = integer_array.clone
+      bits = BitArray.new
+      integer_array.each do |x|
+        q = x/@M
+        q.times {bits.push 1}
+        bits.push 0
+        r = x % @M
+        (@b-1).downto(0){|i| bits.push r[i]}
+      end
+      bits
+    end
+  end
+
+  class GolombRiceDecoder
+    def initialize m = 8
+      @M = m
+      @b = Math.log2(m).to_i
+    end
+    def << data
+      data.map do |exp_sign_combo, significand|
+        exp_sign_combo = deinterleave exp_sign_combo
+        sign = exp_sign_combo & 0x00000001
+        exp =  exp_sign_combo >> 1
+        [sign, exp, significand]
+      end
+    end
+    def deinterleave x
+      x.odd? ? (x + 1)/-2 : x/2
+    end
+
+    def decode bits
+      int_array = []
+      while !bits.empty?
+        q = 0
+        nr = 0
+        q+=1 while bits.shift == 1
+        (@b-1).downto(0) do |a|
+          break if bits.empty?
+          nr += (1 << a) if bits.shift == 1
+        end
+        nr += q * @M
+        int_array.push nr
+      end
+      int_array
+    end
+  end
+
+  class NullCompressor
+    def << data
+      data
+    end
+  end
+end

data/lib/ruby_impl/dct.rb

@@ -1,8 +1,8 @@
 require 'common/noyes_math'
 
 module Noyes
-  # Takes the discrete cosing transform.  Converts a n x m matrix to an n x order
-  # matrix.  ncol should be set to m.
+  # Takes the discrete cosing transform.  Converts an n x m matrix to an n x
+  # order matrix.  ncol should be set to m.
   class DCT 
     include Math
     attr_accessor :melcos

data/lib/ruby_impl/live_cmn.rb

@@ -1,8 +1,8 @@
 module Noyes
+  # Normalizes cepstrum means and applies them.  Dimensionality remains
+  # unchanged.  NOTE:  This class resets itself automatically if bounds drift
+  # too much.  Possibly these bounds should be parameterized.
   class LiveCMN
-    # Normalizes cepstrum means and applies them.  Dimensionality remains
-    # unchanged.  NOTE:  This class resets itself automatically if bounds drift
-    # too much.  Possibly these bounds should be parameterized.
     def initialize dimensions=13, init_mean=45.0, window_size=100, shift=160
       @init_mean = init_mean; @shift = shift; @ws = window_size
       @sums = Array.new dimensions, 0
@@ -29,7 +29,7 @@ module Noyes
       
       if @means.first > 70 || @means.first < 5
         reset
-      elsif @frame_count >= @shift
+      else
         @sums = @sums.map {|x| x * per_frame * @ws}
         @frame_count = @ws
       end

data/lib/ruby_impl/mel_filter.rb

@@ -19,15 +19,15 @@ module Noyes
     def << power_spectra
       power_spectra.map do |spectrum|
         mel_bank = Array.new @indices.size
-	@indices.size.times do |i|
-	  initial_index, weights = @indices[i], @weights[i]
-	  output = 0.0
-	  weights.size.times do |j|
-	    index = initial_index + j
-	    output += spectrum[index] * weights[j] if index < spectrum.length 
-	  end
-	  mel_bank[i] = output
-	end
+        @indices.size.times do |i|
+          initial_index, weights = @indices[i], @weights[i]
+          output = 0.0
+          weights.size.times do |j|
+            index = initial_index + j
+            output += spectrum[index] * weights[j] if index < spectrum.length
+          end
+          mel_bank[i] = output
+        end
         mel_bank
       end
     end

data/lib/ruby_impl/segment.rb

@@ -1,14 +1,25 @@
 module Noyes
   # Segments an array of data into an array of arrays.  Inner arrays are the
-  # size of the window.
+  # size of the window.  The segmenter will not return less than three
+  # segments at a time.  Here is the forumla for determining the number
+  # of segments produced by an array of data:
+  #
+  # if combolen >= winsize:
+  #   nsegs = 1 + (arrlen - wsize - (arrlen - wise) % shift)/shift
+  # else
+  #   combolen = 0
+  #
+  # This formula is derived from the following equation:
+  # arrlen = wsize + shift (nsegs - 1) + (arrlen - wsize) % shift
   class Segmenter
+    @@MIN_SEGMENTS = 3
     def initialize window_size, shift
       @winsz = window_size; @winshift = shift
       @overflow = nil
     end    
     def << data
       data = @overflow + data if @overflow
-      if data.size < @winsz + @winshift * 5
+      if data.size < @winsz + @winshift * @@MIN_SEGMENTS
         @overflow = data
         return nil 
       else

data/lib/ruby_impl/speech_trimmer.rb

@@ -20,7 +20,9 @@ module Noyes
 
     def << pcm
       return if eos?
-      (@segmenter << pcm).inject [] do |memo, centisec|
+      segments = @segmenter << pcm
+      return unless segments
+      segments.inject [] do |memo, centisec|
         enqueue centisec unless eos?
         while x = dequeue
           memo << x

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification 
 name: noyes
 version: !ruby/object:Gem::Version 
-  version: 0.9.2
+  version: 1.0.0
 platform: ruby
 authors: 
 - Joe Woelfel
@@ -9,7 +9,7 @@ autorequire:
 bindir: bin
 cert_chain: []
 
-date: 2010-06-30 00:00:00 -04:00
+date: 2010-08-03 00:00:00 -04:00
 default_executable: 
 dependencies: 
 - !ruby/object:Gem::Dependency 
@@ -32,43 +32,43 @@ executables:
 - noyes_dump8k
 - nrec
 extensions: 
-- lib/c_impl/extconf.rb
+- lib/cext/extconf.rb
 extra_rdoc_files: 
 - COPYING
 - FAQ
 files: 
 - VERSION
-- lib/c_impl/array_list.c
-- lib/c_impl/bent_cent_marker.c
-- lib/c_impl/discrete_cosine_transform.c
-- lib/c_impl/extconf.rb
-- lib/c_impl/fast_8k_mfcc.c
-- lib/c_impl/hamming_window.c
-- lib/c_impl/live_cmn.c
-- lib/c_impl/log_compressor.c
-- lib/c_impl/mel_filter.c
-- lib/c_impl/n_array_list.c
-- lib/c_impl/n_array_list.h
-- lib/c_impl/n_bent_cent_marker.c
-- lib/c_impl/n_dft.c
-- lib/c_impl/n_discrete_cosine_transform.c
-- lib/c_impl/n_fast_8k_mfcc.c
-- lib/c_impl/n_hamming_window.c
-- lib/c_impl/n_live_cmn.c
-- lib/c_impl/n_log_compressor.c
-- lib/c_impl/n_matrix.c
-- lib/c_impl/n_mel_filter.c
-- lib/c_impl/n_power_spec.c
-- lib/c_impl/n_preemphasis.c
-- lib/c_impl/n_segmenter.c
-- lib/c_impl/n_speech_trimmer.c
-- lib/c_impl/noyes.h
-- lib/c_impl/noyes_c.c
-- lib/c_impl/power_spectrum.c
-- lib/c_impl/preemphasis.c
-- lib/c_impl/rnoyes.h
-- lib/c_impl/segmenter.c
-- lib/c_impl/speech_trimmer.c
+- lib/cext/array_list.c
+- lib/cext/bent_cent_marker.c
+- lib/cext/c_array_list.c
+- lib/cext/c_array_list.h
+- lib/cext/c_bent_cent_marker.c
+- lib/cext/c_dft.c
+- lib/cext/c_discrete_cosine_transform.c
+- lib/cext/c_hamming_window.c
+- lib/cext/c_live_cmn.c
+- lib/cext/c_log_compressor.c
+- lib/cext/c_matrix.c
+- lib/cext/c_mel_filter.c
+- lib/cext/c_mfcc_16x8.c
+- lib/cext/c_noyes.h
+- lib/cext/c_power_spec.c
+- lib/cext/c_preemphasis.c
+- lib/cext/c_segmenter.c
+- lib/cext/c_speech_trimmer.c
+- lib/cext/discrete_cosine_transform.c
+- lib/cext/extconf.rb
+- lib/cext/hamming_window.c
+- lib/cext/live_cmn.c
+- lib/cext/log_compressor.c
+- lib/cext/mel_filter.c
+- lib/cext/mfcc_16x8.c
+- lib/cext/power_spectrum.c
+- lib/cext/preemphasis.c
+- lib/cext/r_noyes.c
+- lib/cext/r_noyes.h
+- lib/cext/segmenter.c
+- lib/cext/speech_trimmer.c
 - lib/common.rb
 - lib/common/file2pcm.rb
 - lib/common/mock_noyes_server.rb
@@ -97,6 +97,7 @@ files:
 - lib/noyes_c.rb
 - lib/noyes_java.rb
 - lib/ruby_impl/bent_cent_marker.rb
+- lib/ruby_impl/compression.rb
 - lib/ruby_impl/dct.rb
 - lib/ruby_impl/delta.rb
 - lib/ruby_impl/discrete_fourier_transform.rb
@@ -109,6 +110,7 @@ files:
 - lib/ruby_impl/preemphasis.rb
 - lib/ruby_impl/segment.rb
 - lib/ruby_impl/speech_trimmer.rb
+- ship/noyes.jar
 - COPYING
 - FAQ
 has_rdoc: true

data/lib/c_impl/bent_cent_marker.c

@@ -1,36 +0,0 @@
-#include "ruby.h"
-#include "noyes.h"
-#include "rnoyes.h"
-
-static int id_push;
-
-VALUE cBentCentMarker;
-
-static void bent_cent_marker_free(void *p) {
-  free_bent_cent_marker(p);
-}
-
-static VALUE t_init(VALUE self) {
-  BentCentMarker *pre = new_bent_cent_marker();
-  VALUE prev = Data_Wrap_Struct(cBentCentMarker, 0, bent_cent_marker_free, pre);
-  rb_iv_set(self, "@bent_cent_marker", prev);
-  return self;
-}
-
-static VALUE t_left_shift(VALUE self, VALUE obj) {
-  NMatrix1 *M = v_2_nmatrix1(obj);
-  BentCentMarker *pre;
-  VALUE prev = rb_iv_get(self, "@bent_cent_marker");
-  Data_Get_Struct(prev, BentCentMarker, pre);
-  int res = bent_cent_marker_apply(pre, M);
-  free_nmatrix1(M);
-  return res ? Qtrue : Qfalse;
-}
-
-void Init_bent_cent_marker() {
-  VALUE m_noyes_c = rb_define_module("NoyesC");
-  cBentCentMarker = rb_define_class_under(m_noyes_c, "BentCentMarker", rb_cObject);
-  rb_define_method(cBentCentMarker, "initialize", t_init, 0);
-  rb_define_method(cBentCentMarker, "<<", t_left_shift, 1);
-  id_push = rb_intern("push");
-}

data/lib/c_impl/fast_8k_mfcc.c

@@ -1,39 +0,0 @@
-#include "ruby.h"
-#include "noyes.h"
-#include "rnoyes.h"
-
-static int id_push;
-
-VALUE cFast8kMfcc;
-
-static void fast_8k_mfcc_free(Fast8kMfcc *p) {
-  free_fast_8k_mfcc(p);
-}
-
-static VALUE t_init(VALUE self, VALUE args) {
-  Fast8kMfcc *s = new_fast_8k_mfcc();
-  VALUE fast_8k_mfcc = Data_Wrap_Struct(cFast8kMfcc, 0, fast_8k_mfcc_free, s);
-  rb_iv_set(self, "@fast_8k_mfcc", fast_8k_mfcc);
-
-  return self;
-}
-
-static VALUE t_left_shift(VALUE self, VALUE obj) {
-  NMatrix1 *M = v_2_nmatrix1(obj); 
-  VALUE fast_8k_mfcc = rb_iv_get(self, "@fast_8k_mfcc");
-  Fast8kMfcc *s;
-  Data_Get_Struct(fast_8k_mfcc, Fast8kMfcc, s);
-  NMatrix *N = fast_8k_mfcc_apply(s, M);
-  VALUE result = nmatrix_2_v(N);
-  free_nmatrix(N);
-  free_nmatrix1(M);
-  return result;
-}
-
-void Init_fast_8k_mfcc() {
-  VALUE m_noyes_c = rb_define_module("NoyesC");
-  cFast8kMfcc = rb_define_class_under(m_noyes_c, "Fast8kMfcc", rb_cObject);
-  rb_define_method(cFast8kMfcc, "initialize", t_init, -2);
-  rb_define_method(cFast8kMfcc, "<<", t_left_shift, 1);
-  id_push = rb_intern("push");
-}

data/lib/c_impl/n_array_list.h

@@ -1,18 +0,0 @@
-#ifndef _N_ARRAY_LIST_H_
-#define _N_ARRAY_LIST_H_
-
-typedef struct {
-  int capacity;
-  void **data;
-  int size;
-} NList;
-
-NList * n_list_new();
-void n_list_free(NList * self);
-int n_list_size(const NList * self);
-int n_list_add(NList * self, void * object);
-int n_list_remove(NList * self, int start, int finish);
-void * n_list_get(const NList * self, const int index);
-int n_list_is_empty(const NList * self);
-
-#endif

data/lib/c_impl/n_fast_8k_mfcc.c

@@ -1,55 +0,0 @@
-#include "noyes.h"
-
-Fast8kMfcc * new_fast_8k_mfcc() {
-  double factor = 0.97;
-  int nfilt = 32;
-  int min_freq = 200;
-  int max_freq = 3700;
-  int nfft = 256*2;
-  int freq = 8000*2;
-  int shift = 80*2;
-  int frame_size = 205*2;
-  double log_zero = -0.00001;
-  int dimensions=13;
-  int cmn_init_mean=45.0;
-  int cmn_window_size=100;
-  int cmn_shift=160;
-  
-  Fast8kMfcc *self = malloc(sizeof(Fast8kMfcc));
-  self->pre = new_preemphasizer(factor);
-  self->seg = new_segmenter(frame_size, shift);
-  self->ham = new_hamming_window(frame_size);
-  self->pow = new_power_spectrum(nfft);
-  self->mel = new_mel_filter(freq, nfft, nfilt, min_freq, max_freq);
-  self->log = new_log_compressor(log_zero);
-  self->dct = new_dct(dimensions, nfilt);
-  self->cmn = new_live_cmn(dimensions, cmn_init_mean, cmn_window_size, cmn_shift);
-  return self;
-}
-
-void free_fast_8k_mfcc(Fast8kMfcc *self) {
-  free(self->seg);
-  free(self->ham);
-  free(self->pow);
-  free(self->mel);
-  free(self->log);
-  free(self->dct);
-  free(self->cmn);
-  free(self);
-}
-
-NMatrix *fast_8k_mfcc_apply(Fast8kMfcc *self, NMatrix1 * data) {
-  NMatrix *M = NULL;
-  NMatrix *N = NULL;
-  NMatrix1 *data1 = preemphasizer_apply(self->pre, data);
-  M = segmenter_apply(self->seg, data1); free_nmatrix1(data1);
-  if (!M)
-    return NULL;
-  N = hamming_window_apply(self->ham, M); free_nmatrix(M);
-  M = power_spectrum_apply(self->pow, N); free_nmatrix(N);
-  N = mel_filter_apply(self->mel, M); free_nmatrix(M);
-  M = log_compressor_apply(self->log, N); free_nmatrix(N);
-  N = dct_apply(self->dct, M); free_nmatrix(M);
-  M = live_cmn_apply(self->cmn, N); free_nmatrix(N);
-  return M;
-}

data/lib/c_impl/n_matrix.c

@@ -1,72 +0,0 @@
-#include "noyes.h"
-
-// A 2 dimensional matrix "class".
-NMatrix *new_nmatrix(int rows, int cols) {
-  NMatrix *M = malloc(sizeof(NMatrix));
-  M->data = malloc(rows * sizeof(double*));
-  int i;
-  for (i=0;i<rows;++i) {
-    M->data[i] = malloc(cols * sizeof(double));
-  }
-  M->rows = rows;
-  M->cols = cols;
-  return M;
-}
-
-void free_nmatrix(NMatrix *M) {
-  if (M) {
-    int i;
-    for (i=0;i<M->rows;++i) {
-      free(M->data[i]);
-    }
-    free(M->data);
-    free(M);
-  }
-}
-
-// A 1 dimensional matrix "class".
-NMatrix1 *new_nmatrix1(int rows) {
-  NMatrix1 *M = malloc(sizeof(NMatrix1));
-  M->data = malloc(rows * sizeof(double));
-  M->rows = rows;
-  return M;
-}
-
-void free_nmatrix1(NMatrix1 *M) {
-  if (M) {
-    free(M->data);
-    free(M);
-  }
-}
-
-// Converts a square matrix to a list of one dimensional matrices.
-// Simultaneously frees the original square matrix.
-NMatrix1 ** nmatrix_2_nmatrix1s(NMatrix *M) {
-  NMatrix1 **single = malloc(sizeof(NMatrix1*) * M->rows);
-  int i;
-  for (i=0;i<M->rows;++i) {
-    single[i] = malloc(sizeof(NMatrix1));
-    single[i]->data = M->data[i];
-    single[i]->rows = M->cols;
-  }
-  free(M->data);
-  free(M);
-  return single;
-}
-
-// Converts an array of one dimensional arrays into a square matrix.  It frees
-// these arrays in the process.
-NMatrix * nmatrix1_2_nmatrix(NMatrix1 **array, int size) {
-  if (size ==0)
-    return NULL;
-  NMatrix *result = malloc(sizeof(NMatrix));
-  result->data = malloc(sizeof(double*) * size);
-  result->rows = size;
-  int i;
-  for (i=0; i<size; ++i) {
-    result->data[i] = array[i]->data;
-    free(array[i]);
-  }
-
-  return result;
-}