Class: Diarize::Speaker

Inherits:
Object
  • Object
show all
Includes:
ToRdf
Defined in:
lib/diarize/speaker.rb

Constant Summary collapse

@@log_likelihood_threshold = 
-33
@@detection_threshold = 
0.2
@@speakers = 
{}

Instance Attribute Summary collapse

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(uri = nil, gender = nil, model_file = nil) ⇒ Speaker

Returns a new instance of Speaker.



11
12
13
14
15
16
 
# File 'lib/diarize/speaker.rb', line 11

def initialize(uri = nil, gender = nil, model_file = nil)
  @model = Speaker.load_model(model_file) if model_file
  @uri = uri
  @gender = gender
  @normalized = false
end

Instance Attribute Details

#gender ⇒ Object (readonly)

Returns the value of attribute gender.



9
10
11
 
# File 'lib/diarize/speaker.rb', line 9

def gender
  @gender
end

#model ⇒ Object

Returns the value of attribute model.



8
9
10
 
# File 'lib/diarize/speaker.rb', line 8

def model
  @model
end

#model_uri ⇒ Object

Returns the value of attribute model_uri.



8
9
10
 
# File 'lib/diarize/speaker.rb', line 8

def model_uri
  @model_uri
end

#normalized ⇒ Object

Returns the value of attribute normalized.



8
9
10
 
# File 'lib/diarize/speaker.rb', line 8

def normalized
  @normalized
end

Class Method Details

.detection_threshold ⇒ Object 



42
43
44
 
# File 'lib/diarize/speaker.rb', line 42

def self.detection_threshold
  @@detection_threshold
end

.detection_threshold=(threshold) ⇒ Object 



38
39
40
 
# File 'lib/diarize/speaker.rb', line 38

def self.detection_threshold=(threshold)
  @@detection_threshold = threshold
end

.divergence(speaker1, speaker2) ⇒ Object 



55
56
57
58
59
60
61
 
# File 'lib/diarize/speaker.rb', line 55

def self.divergence(speaker1, speaker2)
  # TODO bundle in mean_log_likelihood to weight down unlikely models?
  return unless speaker1.model and speaker2.model
  # MAP Gaussian divergence
  # See "A model space framework for efficient speaker detection", Interspeech'05
  divergence_lium(speaker1, speaker2)
end

.divergence_lium(speaker1, speaker2) ⇒ Object 



63
64
65
66
 
# File 'lib/diarize/speaker.rb', line 63

def self.divergence_lium(speaker1, speaker2)
  # fr.lium.spkDiarization.libModel.Distance.GDMAP(speaker1.model, speaker2.model)
  Rjb::import('fr.lium.spkDiarization.libModel.Distance').GDMAP(speaker1.model, speaker2.model)
end

.divergence_ruby(speaker1, speaker2) ⇒ Object 



68
69
70
 
# File 'lib/diarize/speaker.rb', line 68

def self.divergence_ruby(speaker1, speaker2)
  SuperVector.divergence(speaker1.supervector, speaker2.supervector)
end

.find_or_create(uri, gender) ⇒ Object 



50
51
52
53
 
# File 'lib/diarize/speaker.rb', line 50

def self.find_or_create(uri, gender)
  return @@speakers[uri] if @@speakers[uri]
  @@speakers[uri] = Speaker.new(uri, gender)
end

.load_model(filename) ⇒ Object 



46
47
48
 
# File 'lib/diarize/speaker.rb', line 46

def self.load_model(filename)
  read_gmm(filename)
end

.match(speakers) ⇒ Object 



82
83
84
 
# File 'lib/diarize/speaker.rb', line 82

def self.match(speakers)
  speakers.combination(2).select { |s1, s2| s1.same_speaker_as(s2) }
end

.match_sets(speakers1, speakers2) ⇒ Object 



72
73
74
75
76
77
78
79
80
 
# File 'lib/diarize/speaker.rb', line 72

def self.match_sets(speakers1, speakers2)
  matches = []
  speakers1.each do |s1|
    speakers2.each do |s2|
      matches << [ s1, s2 ] if s1.same_speaker_as(s2)
    end
  end
  matches
end

.read_gmm(filename) ⇒ Object 



140
141
142
143
144
145
146
147
148
149
150
 
# File 'lib/diarize/speaker.rb', line 140

def self.read_gmm(filename)
  # gmmlist = java.util.ArrayList.new
  gmmlist = Rjb::JavaObjectWrapper.new("java.util.ArrayList")
  # input = fr.lium.spkDiarization.lib.IOFile.new(filename, 'rb')
  input = Rjb::import('fr.lium.spkDiarization.lib.IOFile').new(filename, 'rb')
  input.open
  # fr.lium.spkDiarization.libModel.ModelIO.readerGMMContainer(input, gmmlist)
  Rjb::import('fr.lium.spkDiarization.libModel.ModelIO').readerGMMContainer(input, gmmlist.java_object)
  input.close
  gmmlist.to_a.first.java_object
end

.ubm ⇒ Object 



18
19
20
21
22
23
 
# File 'lib/diarize/speaker.rb', line 18

def self.ubm
  speaker = Speaker.new
  speaker.normalized = true
  speaker.model = Speaker.load_model(File.join(File.expand_path(File.dirname(__FILE__)), 'ubm.gmm'))
  speaker
end

Instance Method Details

#mean_log_likelihood ⇒ Object 



25
26
27
 
# File 'lib/diarize/speaker.rb', line 25

def mean_log_likelihood
  @mean_log_likelihood ? @mean_log_likelihood : model.mean_log_likelihood # Will be NaN if model was loaded from somewhere
end

#mean_log_likelihood=(mll) ⇒ Object 



29
30
31
 
# File 'lib/diarize/speaker.rb', line 29

def mean_log_likelihood=(mll)
  @mean_log_likelihood = mll
end

#namespaces ⇒ Object 



122
123
124
 
# File 'lib/diarize/speaker.rb', line 122

def namespaces
  super.merge 'ws' => 'http://wsarchive.prototype0.net/ontology/'
end

#normalize! ⇒ Object 



86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
 
# File 'lib/diarize/speaker.rb', line 86

def normalize!
  unless @normalized
    # Applies M-Norm from "D-MAP: a Distance-Normalized MAP Estimation of Speaker Models for Automatic Speaker Verification"
    # to the associated GMM, placing it on a unit hyper-sphere with a UBM centre (model will be at distance one from the UBM
    # according to GDMAP)
    # Using supervectors: vector = (1.0 / distance_to_ubm) * vector + (1.0 - 1.0 / distance_to_ubm) * ubm_vector
    speaker_ubm = Speaker.ubm
    distance_to_ubm = Math.sqrt(Speaker.divergence(self, speaker_ubm))
    model.nb_of_components.times do |k|
      gaussian = model.components.get(k)
      gaussian.dim.times do |i|
        normalized_mean = (1.0 / distance_to_ubm) * gaussian.mean(i) + (1.0 - 1.0 / distance_to_ubm)  * speaker_ubm.model.components.get(k).mean(i)
        gaussian.set_mean(i, normalized_mean)
      end
    end
    @normalized = true
  end
  @normalized
end

#rdf_mapping ⇒ Object 



134
135
136
 
# File 'lib/diarize/speaker.rb', line 134

def rdf_mapping
  { 'ws:gender' => gender, 'ws:model' => model_uri, 'ws:mean_log_likelihood' => model.mean_log_likelihood, 'ws:supervector_hash' => supervector.hash.to_s }
end

#same_speaker_as(other) ⇒ Object 



106
107
108
109
110
111
112
113
 
# File 'lib/diarize/speaker.rb', line 106

def same_speaker_as(other)
  # Detection score defined in Ben2005
  return unless [ self.mean_log_likelihood, other.mean_log_likelihood ].min > @@log_likelihood_threshold
  self.normalize!
  other.normalize!
  detection_score = 1.0 - Speaker.divergence(other, self)
  detection_score > @@detection_threshold
end

#save_model(filename) ⇒ Object 



33
34
35
36
 
# File 'lib/diarize/speaker.rb', line 33

def save_model(filename)
  # TODO perhaps a warning if a normalised model is being saved?
  write_gmm(filename, @model)
end

#supervector ⇒ Object 



115
116
117
118
 
# File 'lib/diarize/speaker.rb', line 115

def supervector
  # TODO: cache only when normalized
  @supervector ||= SuperVector.generate_from_model(model)
end

#type_uri ⇒ Object 



130
131
132
 
# File 'lib/diarize/speaker.rb', line 130

def type_uri
  'ws:Speaker'
end

#uri ⇒ Object 



126
127
128
 
# File 'lib/diarize/speaker.rb', line 126

def uri
  @uri
end