Module: Cohere::Transcribe::Alignment::CTC

Defined in:
lib/cohere/transcribe/alignment/ctc.rb

Overview

Ruby port of TorchAudio's CPU forced_align kernel. The two-row score matrix remains float32 so its Viterbi choices follow the reference kernel's rounding and tie-breaking behavior, while the back-pointers use one byte per state and frame.

Class Method Summary collapse

Class Method Details

.forced_align(log_probs, targets, blank:) ⇒ Object 



17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
 
# File 'lib/cohere/transcribe/alignment/ctc.rb', line 17

def forced_align(log_probs, targets, blank:)
  frame_count, class_count = matrix_shape(log_probs)
  validate_targets!(targets, blank, class_count)

  target_count = targets.length
  repeat_count = targets.each_cons(2).count { |left, right| left == right }
  minimum_frames = target_count + repeat_count
  if frame_count < minimum_frames
    raise ArgumentError,
          "targets length is too long for CTC: " \
          "#{frame_count} frames, #{target_count} targets, #{repeat_count} repeats"
  end

  state_count = (2 * target_count) + 1
  alphas = Numo::SFloat.zeros(2, state_count).fill(-Float::INFINITY)
  back_pointers = "\xFF".b * (frame_count * state_count)

  first_state = frame_count > minimum_frames ? 0 : 1
  end_state = state_count == 1 ? 1 : 2
  (first_state...end_state).each do |state|
    alphas[0, state] = probability(log_probs, 0, label_for(state, targets, blank))
  end

  1.upto(frame_count - 1) do |frame|
    if frame_count - frame <= minimum_frames
      first_state += 1 if first_state.odd? && targets.fetch(first_state / 2) != targets.fetch((first_state / 2) + 1)
      first_state += 1
    end
    if frame <= minimum_frames
      if end_state.even? && end_state < 2 * target_count &&
         targets.fetch((end_state / 2) - 1) != targets.fetch(end_state / 2)
        end_state += 1
      end
      end_state += 1
    end

    current_row = frame % 2
    previous_row = (frame - 1) % 2
    alphas[current_row, true] = -Float::INFINITY
    loop_start = first_state
    if first_state.zero?
      alphas[current_row, 0] = alphas[previous_row, 0] + probability(log_probs, frame, blank)
      back_pointers.setbyte(frame * state_count, 0)
      loop_start += 1
    end

    (loop_start...end_state).each do |state|
      stay = alphas[previous_row, state]
      advance = alphas[previous_row, state - 1]
      skip = -Float::INFINITY
      if state.odd? && state != 1 && targets.fetch(state / 2) != targets.fetch((state / 2) - 1)
        skip = alphas[previous_row, state - 2]
      end

      score, step = if skip > advance && skip > stay
                      [skip, 2]
                    elsif advance > stay && advance > skip
                      [advance, 1]
                    else
                      [stay, 0]
                    end
      back_pointers.setbyte((frame * state_count) + state, step)
      alphas[current_row, state] =
        score + probability(log_probs, frame, label_for(state, targets, blank))
    end
  end

  final_row = (frame_count - 1) % 2
  state = if alphas[final_row, state_count - 1] > alphas[final_row, state_count - 2]
            state_count - 1
          else
            state_count - 2
          end
  path = Array.new(frame_count)
  (frame_count - 1).downto(0) do |frame|
    path[frame] = label_for(state, targets, blank)
    break if frame.zero?

    step = back_pointers.getbyte((frame * state_count) + state)
    raise "CTC back-pointer was not initialized" unless step && step <= 2

    state -= step
  end
  path.freeze
end

.merge_repeats(path, index_to_token) ⇒ Object 



103
104
105
106
107
108
109
110
111
112
113
114
 
# File 'lib/cohere/transcribe/alignment/ctc.rb', line 103

def merge_repeats(path, index_to_token)
  segments = []
  first = 0
  while first < path.length
    following = first + 1
    following += 1 while following < path.length && path.fetch(first) == path.fetch(following)
    label = index_to_token.fetch(path.fetch(first))
    segments << Segment.new(label: label, start: first, end: following - 1)
    first = following
  end
  segments.freeze
end

.spans(tokens, segments, blank) ⇒ Object 



116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
 
# File 'lib/cohere/transcribe/alignment/ctc.rb', line 116

def spans(tokens, segments, blank)
  letter_index = 0
  token_index = 0
  intervals = []
  interval_start = 0

  segments.each_with_index do |segment, segment_index|
    if token_index == tokens.length
      unless segment_index == segments.length - 1 && segment.label == blank
        raise "CTC path continued with a non-blank label after the transcript"
      end

      next
    end

    token_letters = tokens.fetch(token_index).split
    letter = token_letters.fetch(letter_index)
    next if segment.label == blank
    raise "#{segment.label} != #{letter}" unless segment.label == letter

    interval_start = segment_index if letter_index.zero?
    if letter_index == token_letters.length - 1
      letter_index = 0
      token_index += 1
      intervals << [interval_start, segment_index]
      while token_index < tokens.length && tokens.fetch(token_index).empty?
        intervals << [segment_index, segment_index]
        token_index += 1
      end
    else
      letter_index += 1
    end
  end
  raise "CTC path did not consume every transcript token" unless token_index == tokens.length

  intervals.each_with_index.map do |(first, last), interval_index|
    span = segments[first..last]
    if first.positive?
      previous = segments.fetch(first - 1)
      if previous.label == blank
        padded_start = if interval_index.zero?
                         previous.start
                       else
                         ((previous.start + previous.end) / 2.0).to_i
                       end
        span = [Segment.new(label: blank, start: padded_start, end: span.first.start), *span]
      end
    end
    if last + 1 < segments.length
      following = segments.fetch(last + 1)
      if following.label == blank
        padded_end = if interval_index == intervals.length - 1
                       following.end
                     else
                       ((following.start + following.end) / 2.0).floor
                     end
        span = [*span, Segment.new(label: blank, start: span.last.end, end: padded_end)]
      end
    end
    span.freeze
  end.freeze
end