Merge pull request #2358 from mozilla/rename-probability-confidence

Rename metadata probability field to confidence

native_client/client.cc

@@ -350,7 +350,7 @@ JSONOutput(Metadata* metadata)
   std::vector<meta_word> words = WordsFromMetadata(metadata);
 
   std::ostringstream out_string;
-  out_string << R"({"metadata":{"confidence":)" << metadata->probability << R"(},"words":[)";
+  out_string << R"({"metadata":{"confidence":)" << metadata->confidence << R"(},"words":[)";
 
   for (int i = 0; i < words.size(); i++) {
     meta_word w = words[i];

native_client/ctcdecode/__init__.py

@@ -51,14 +51,14 @@ def ctc_beam_search_decoder(probs_seq,
     :param scorer: External scorer for partially decoded sentence, e.g. word
                    count or language model.
     :type scorer: Scorer
-    :return: List of tuples of log probability and sentence as decoding
+    :return: List of tuples of confidence and sentence as decoding
-             results, in descending order of the probability.
+             results, in descending order of the confidence.
     :rtype: list
     """
     beam_results = swigwrapper.ctc_beam_search_decoder(
         probs_seq, alphabet.config_file(), beam_size, cutoff_prob, cutoff_top_n,
         scorer)
-    beam_results = [(res.probability, alphabet.decode(res.tokens)) for res in beam_results]
+    beam_results = [(res.confidence, alphabet.decode(res.tokens)) for res in beam_results]
     return beam_results
 
 
@@ -93,15 +93,15 @@ def ctc_beam_search_decoder_batch(probs_seq,
     :param scorer: External scorer for partially decoded sentence, e.g. word
                    count or language model.
     :type scorer: Scorer
-    :return: List of tuples of log probability and sentence as decoding
+    :return: List of tuples of confidence and sentence as decoding
-             results, in descending order of the probability.
+             results, in descending order of the confidence.
     :rtype: list
     """
     batch_beam_results = swigwrapper.ctc_beam_search_decoder_batch(
         probs_seq, seq_lengths, alphabet.config_file(), beam_size, num_processes,
         cutoff_prob, cutoff_top_n, scorer)
     batch_beam_results = [
-        [(res.probability, alphabet.decode(res.tokens)) for res in beam_results]
+        [(res.confidence, alphabet.decode(res.tokens)) for res in beam_results]
         for beam_results in batch_beam_results
     ]
     return batch_beam_results

native_client/ctcdecode/decoder_utils.cpp

@@ -46,7 +46,7 @@ std::vector<Output> get_beam_search_result(
   for (size_t i = 0; i < top_paths && i < prefixes.size(); ++i) {
     Output output;
     prefixes[i]->get_path_vec(output.tokens, output.timesteps);
-    output.probability = -prefixes[i]->approx_ctc;
+    output.confidence = -prefixes[i]->approx_ctc;
     output_vecs.push_back(output);
   }
 

native_client/ctcdecode/output.h

@@ -7,7 +7,7 @@
  * for each token in the beam search output
  */
 struct Output {
-    double probability;
+    double confidence;
     std::vector<int> tokens;
     std::vector<int> timesteps;
 };

native_client/deepspeech.h

@@ -30,8 +30,10 @@ typedef struct MetadataItem {
 typedef struct Metadata {
   MetadataItem* items;
   int num_items;
-  // Approximated probability (confidence value) for this transcription.
+  // Approximated confidence value for this transcription. This is roughly the
-  double probability;
+  // sum of the acoustic model logit values for each timestep/character that
+  // contributed to the creation of this transcription.
+  double confidence;
 } Metadata;
 
 enum DeepSpeech_Error_Codes

native_client/dotnet/DeepSpeechClient/Extensions/NativeExtensions.cs

@@ -36,7 +36,7 @@ namespace DeepSpeechClient.Extensions
             var metaData = (Metadata)Marshal.PtrToStructure(intPtr, typeof(Metadata));
 
             managedMetaObject.Items = new Models.MetadataItem[metaData.num_items];
-            managedMetaObject.Probability = metaData.probability;
+            managedMetaObject.Confidence = metaData.confidence;
 
 
             //we need to manually read each item from the native ptr using its size

native_client/dotnet/DeepSpeechClient/Models/Metadata.cs

@@ -6,9 +6,9 @@
     public class Metadata
     {
         /// <summary>
-        /// Approximated probability (confidence value) for this transcription.
+        /// Approximated confidence value for this transcription.
         /// </summary>
-        public double Probability { get; set; }
+        public double Confidence { get; set; }
         /// <summary>
         /// List of metada items containing char, timespet, and time offset.
         /// </summary>

native_client/dotnet/DeepSpeechClient/Structs/Metadata.cs

@@ -15,8 +15,8 @@ namespace DeepSpeechClient.Structs
         /// </summary>
         internal unsafe int num_items;
         /// <summary>
-        /// Approximated probability (confidence value) for this transcription.
+        /// Approximated confidence value for this transcription.
         /// </summary>
-        internal unsafe double probability;
+        internal unsafe double confidence;
     }
 }

native_client/dotnet/DeepSpeechConsole/Program.cs

@@ -26,7 +26,7 @@ namespace CSharpExamples
             var nl = Environment.NewLine;
             string retval =
              Environment.NewLine + $"Recognized text: {string.Join("", meta?.Items?.Select(x => x.Character))} {nl}"
-             + $"Prob: {meta?.Probability} {nl}"
+             + $"Confidence: {meta?.Confidence} {nl}"
              + $"Item count: {meta?.Items?.Length} {nl}"
              + string.Join(nl, meta?.Items?.Select(x => $"Timestep : {x.Timestep} TimeOffset: {x.StartTime} Char: {x.Character}"));
             return retval;

native_client/modelstate.cc

@@ -49,7 +49,7 @@ ModelState::decode_metadata(const DecoderState& state)
 
   std::unique_ptr<Metadata> metadata(new Metadata());
   metadata->num_items = out[0].tokens.size();
-  metadata->probability = out[0].probability;
+  metadata->confidence = out[0].confidence;
 
   std::unique_ptr<MetadataItem[]> items(new MetadataItem[metadata->num_items]());