Merge pull request #1810 from mozilla/update-new-hparams

Update decoder hyperparameters

examples/ffmpeg_vad_streaming/index.js

@@ -11,11 +11,10 @@ const util = require('util');
 const BEAM_WIDTH = 1024;
 
 // The alpha hyperparameter of the CTC decoder. Language Model weight
-const LM_WEIGHT = 1.50;
+const LM_ALPHA = 0.75;
 
-// Valid word insertion weight. This is used to lessen the word insertion penalty
-// when the inserted word is part of the vocabulary
-const VALID_WORD_COUNT_WEIGHT = 2.25;
+// The beta hyperparameter of the CTC decoder. Word insertion bonus.
+const LM_BETA = 1.85;
 
 // These constants are tied to the shape of the graph used (changing them changes
 // the geometry of the first layer), so make sure you use the same constants that
@@ -63,7 +62,7 @@ if (args['lm'] && args['trie']) {
 	console.error('Loading language model from files %s %s', args['lm'], args['trie']);
 	const lm_load_start = process.hrtime();
 	model.enableDecoderWithLM(args['alphabet'], args['lm'], args['trie'],
-		LM_WEIGHT, VALID_WORD_COUNT_WEIGHT);
+		LM_ALPHA, LM_BETA);
 	const lm_load_end = process.hrtime(lm_load_start);
 	console.error('Loaded language model in %ds.', totalTime(lm_load_end));
 }

examples/mic_vad_streaming/README.md

@@ -26,7 +26,7 @@ brew install portaudio
 usage: mic_vad_streaming.py [-h] [-v VAD_AGGRESSIVENESS] [--nospinner]
                             [-w SAVEWAV] -m MODEL [-a ALPHABET] [-l LM]
                             [-t TRIE] [-nf N_FEATURES] [-nc N_CONTEXT]
-                            [-lw LM_WEIGHT] [-vwcw VALID_WORD_COUNT_WEIGHT]
+                            [-la LM_ALPHA] [-lb LM_BETA]
                             [-bw BEAM_WIDTH]
 
 Stream from microphone to DeepSpeech using VAD
@@ -56,13 +56,12 @@ optional arguments:
   -nc N_CONTEXT, --n_context N_CONTEXT
                         Size of the context window used for producing
                         timesteps in the input vector. Default: 9
-  -lw LM_WEIGHT, --lm_weight LM_WEIGHT
+  -la LM_ALPHA, --lm_alpha LM_ALPHA
                         The alpha hyperparameter of the CTC decoder. Language
-                        Model weight. Default: 1.5
-  -vwcw VALID_WORD_COUNT_WEIGHT, --valid_word_count_weight VALID_WORD_COUNT_WEIGHT
-                        Valid word insertion weight. This is used to lessen
-                        the word insertion penalty when the inserted word is
-                        part of the vocabulary. Default: 2.1
+                        Model weight. Default: 0.75
+  -lb LM_BETA, --lm_beta LM_BETA
+                        The beta hyperparameter of the CTC decoder. Word insertion
+                        bonus. Default: 1.85
   -bw BEAM_WIDTH, --beam_width BEAM_WIDTH
                         Beam width used in the CTC decoder when building
                         candidate transcriptions. Default: 500

examples/mic_vad_streaming/mic_vad_streaming.py

@@ -118,7 +118,7 @@ def main(ARGS):
     if ARGS.lm and ARGS.trie:
         logging.info("ARGS.lm: %s", ARGS.lm)
         logging.info("ARGS.trie: %s", ARGS.trie)
-        model.enableDecoderWithLM(ARGS.alphabet, ARGS.lm, ARGS.trie, ARGS.lm_weight, ARGS.valid_word_count_weight)
+        model.enableDecoderWithLM(ARGS.alphabet, ARGS.lm, ARGS.trie, ARGS.lm_alpha, ARGS.lm_beta)
 
     # Start audio with VAD
     vad_audio = VADAudio(aggressiveness=ARGS.vad_aggressiveness)
@@ -148,8 +148,8 @@ def main(ARGS):
 
 if __name__ == '__main__':
     BEAM_WIDTH = 500
-    LM_WEIGHT = 1.50
-    VALID_WORD_COUNT_WEIGHT = 2.10
+    LM_ALPHA = 0.75
+    LM_BETA = 1.85
     N_FEATURES = 26
     N_CONTEXT = 9
 
@@ -175,10 +175,10 @@ def main(ARGS):
                         help=f"Number of MFCC features to use. Default: {N_FEATURES}")
     parser.add_argument('-nc', '--n_context', type=int, default=N_CONTEXT,
                         help=f"Size of the context window used for producing timesteps in the input vector. Default: {N_CONTEXT}")
-    parser.add_argument('-lw', '--lm_weight', type=float, default=LM_WEIGHT,
-                        help=f"The alpha hyperparameter of the CTC decoder. Language Model weight. Default: {LM_WEIGHT}")
-    parser.add_argument('-vwcw', '--valid_word_count_weight', type=float, default=VALID_WORD_COUNT_WEIGHT,
-                        help=f"Valid word insertion weight. This is used to lessen the word insertion penalty when the inserted word is part of the vocabulary. Default: {VALID_WORD_COUNT_WEIGHT}")
+    parser.add_argument('-la', '--lm_alpha', type=float, default=LM_ALPHA,
+                        help=f"The alpha hyperparameter of the CTC decoder. Language Model weight. Default: {LM_ALPHA}")
+    parser.add_argument('-lb', '--lm_beta', type=float, default=LM_BETA,
+                        help=f"The beta hyperparameter of the CTC decoder. Word insertion bonus. Default: {LM_BETA}")
     parser.add_argument('-bw', '--beam_width', type=int, default=BEAM_WIDTH,
                         help=f"Beam width used in the CTC decoder when building candidate transcriptions. Default: {BEAM_WIDTH}")
 

examples/net_framework/CSharpExamples/DeepSpeechConsole/Program.cs

@@ -38,8 +38,8 @@ static void Main(string[] args)
             const uint N_CEP = 26;
             const uint N_CONTEXT = 9;
             const uint BEAM_WIDTH = 200;
-            const float LM_WEIGHT = 1.50f;
-            const float VALID_WORD_COUNT_WEIGHT = 2.10f;
+            const float LM_ALPHA = 0.75f;
+            const float LM_BETA = 1.85f;
 
             Stopwatch stopwatch = new Stopwatch();
 
@@ -76,7 +76,7 @@ static void Main(string[] args)
                                 alphabet ?? "alphabet.txt",
                                 lm ?? "lm.binary",
                                 trie ?? "trie",
-                                LM_WEIGHT, VALID_WORD_COUNT_WEIGHT);
+                                LM_ALPHA, LM_BETA);
                         }
                         catch (IOException ex)
                         {

examples/net_framework/CSharpExamples/DeepSpeechWPF/MainWindow.xaml.cs

@@ -24,8 +24,8 @@ public partial class MainWindow : Window
         private const uint N_CEP = 26;
         private const uint N_CONTEXT = 9;
         private const uint BEAM_WIDTH = 500;
-        private const float LM_WEIGHT = 1.50f;
-        private const float VALID_WORD_COUNT_WEIGHT = 2.10f;
+        private const float LM_ALPHA = 0.75f;
+        private const float LM_BETA = 1.85f;
 
 
 
@@ -160,7 +160,7 @@ await Task.Run(() =>
             {
                 try
                 {
-                    if (_sttClient.EnableDecoderWithLM("alphabet.txt", "lm.binary", "trie", LM_WEIGHT, VALID_WORD_COUNT_WEIGHT) != 0)
+                    if (_sttClient.EnableDecoderWithLM("alphabet.txt", "lm.binary", "trie", LM_ALPHA, LM_BETA) != 0)
                     {
                         MessageBox.Show("Error loading LM.");
                         Dispatcher.Invoke(() => btnEnableLM.IsEnabled = true);

examples/vad_transcriber/wavTranscriber.py

@@ -19,16 +19,16 @@ def load_model(models, alphabet, lm, trie):
     N_FEATURES = 26
     N_CONTEXT = 9
     BEAM_WIDTH = 500
-    LM_WEIGHT = 1.50
-    VALID_WORD_COUNT_WEIGHT = 2.10
+    LM_ALPHA = 0.75
+    LM_BETA = 1.85
 
     model_load_start = timer()
     ds = Model(models, N_FEATURES, N_CONTEXT, alphabet, BEAM_WIDTH)
     model_load_end = timer() - model_load_start
     logging.debug("Loaded model in %0.3fs." % (model_load_end))
 
     lm_load_start = timer()
-    ds.enableDecoderWithLM(alphabet, lm, trie, LM_WEIGHT, VALID_WORD_COUNT_WEIGHT)
+    ds.enableDecoderWithLM(alphabet, lm, trie, LM_ALPHA, LM_BETA)
     lm_load_end = timer() - lm_load_start
     logging.debug('Loaded language model in %0.3fs.' % (lm_load_end))
 

native_client/client.cc

@@ -24,8 +24,8 @@
 #define N_CEP 26
 #define N_CONTEXT 9
 #define BEAM_WIDTH 500
-#define LM_WEIGHT 1.50f
-#define VALID_WORD_COUNT_WEIGHT 2.10f
+#define LM_ALPHA 0.75f
+#define LM_BETA 1.85f
 
 typedef struct {
   const char* string;
@@ -253,8 +253,8 @@ main(int argc, char **argv)
                                         alphabet,
                                         lm,
                                         trie,
-                                        LM_WEIGHT,
-                                        VALID_WORD_COUNT_WEIGHT);
+                                        LM_ALPHA,
+                                        LM_BETA);
     if (status != 0) {
       fprintf(stderr, "Could not enable CTC decoder with LM.\n");
       return 1;

native_client/java/app/src/main/java/deepspeech/mozilla/org/deepspeech/DeepSpeechActivity.java

@@ -38,8 +38,8 @@ public class DeepSpeechActivity extends AppCompatActivity {
     final int N_CEP = 26;
     final int N_CONTEXT = 9;
     final int BEAM_WIDTH = 50;
-    final float LM_WEIGHT = 1.50f;
-    final float VALID_WORD_COUNT_WEIGHT = 2.10f;
+    final float LM_ALPHA = 0.75f;
+    final float LM_BETA = 1.85f;
 
     private char readLEChar(RandomAccessFile f) throws IOException {
         byte b1 = f.readByte();

native_client/java/app/src/main/java/deepspeech/mozilla/org/deepspeech/Model.java

@@ -16,8 +16,8 @@ public void destroyModel() {
         impl.DestroyModel(this._msp);
     }
 
-    public void enableDecoderWihLM(String alphabet, String lm, String trie, float lm_weight, float valid_word_count_weight) {
-        impl.EnableDecoderWithLM(this._msp, alphabet, lm, trie, lm_weight, valid_word_count_weight);
+    public void enableDecoderWihLM(String alphabet, String lm, String trie, float lm_alpha, float lm_beta) {
+        impl.EnableDecoderWithLM(this._msp, alphabet, lm, trie, lm_alpha, lm_beta);
     }
 
     public String stt(short[] buffer, int buffer_size, int sample_rate) {

native_client/javascript/client.js

@@ -15,11 +15,10 @@ const util = require('util');
 const BEAM_WIDTH = 500;
 
 // The alpha hyperparameter of the CTC decoder. Language Model weight
-const LM_WEIGHT = 1.50;
+const LM_ALPHA = 0.75;
 
-// Valid word insertion weight. This is used to lessen the word insertion penalty
-// when the inserted word is part of the vocabulary
-const VALID_WORD_COUNT_WEIGHT = 2.10;
+// The beta hyperparameter of the CTC decoder. Word insertion bonus.
+const LM_BETA = 1.85;
 
 
 // These constants are tied to the shape of the graph used (changing them changes
@@ -102,7 +101,7 @@ audioStream.on('finish', () => {
     console.error('Loading language model from files %s %s', args['lm'], args['trie']);
     const lm_load_start = process.hrtime();
     model.enableDecoderWithLM(args['alphabet'], args['lm'], args['trie'],
-                              LM_WEIGHT, VALID_WORD_COUNT_WEIGHT);
+                              LM_ALPHA, LM_BETA);
     const lm_load_end = process.hrtime(lm_load_start);
     console.error('Loaded language model in %ds.', totalTime(lm_load_end));
   }

native_client/python/client.py

@@ -23,11 +23,10 @@
 BEAM_WIDTH = 500
 
 # The alpha hyperparameter of the CTC decoder. Language Model weight
-LM_WEIGHT = 1.50
+LM_ALPHA = 0.75
 
-# Valid word insertion weight. This is used to lessen the word insertion penalty
-# when the inserted word is part of the vocabulary
-VALID_WORD_COUNT_WEIGHT = 2.10
+# The beta hyperparameter of the CTC decoder. Word insertion bonus.
+LM_BETA = 1.85
 
 
 # These constants are tied to the shape of the graph used (changing them changes
@@ -85,8 +84,7 @@ def main():
     if args.lm and args.trie:
         print('Loading language model from files {} {}'.format(args.lm, args.trie), file=sys.stderr)
         lm_load_start = timer()
-        ds.enableDecoderWithLM(args.alphabet, args.lm, args.trie, LM_WEIGHT,
-                               VALID_WORD_COUNT_WEIGHT)
+        ds.enableDecoderWithLM(args.alphabet, args.lm, args.trie, LM_ALPHA, LM_BETA)
         lm_load_end = timer() - lm_load_start
         print('Loaded language model in {:.3}s.'.format(lm_load_end), file=sys.stderr)
 

tc-tests-utils.sh

@@ -151,12 +151,12 @@ assert_correct_multi_ldc93s1()
 
 assert_correct_ldc93s1_prodmodel()
 {
-  assert_correct_inference "$1" "she had a due in greasy wash water year"
+  assert_correct_inference "$1" "she had a due and greasy wash water year"
 }
 
 assert_correct_ldc93s1_prodmodel_stereo_44k()
 {
-  assert_correct_inference "$1" "she had a due in greasy wash water year"
+  assert_correct_inference "$1" "she had a due and greasy wash water year"
 }
 
 assert_correct_warning_upsampling()

util/flags.py

@@ -131,8 +131,8 @@ def create_flags():
     tf.app.flags.DEFINE_string  ('lm_binary_path',       'data/lm/lm.binary', 'path to the language model binary file created with KenLM')
     tf.app.flags.DEFINE_string  ('lm_trie_path',         'data/lm/trie', 'path to the language model trie file created with native_client/generate_trie')
     tf.app.flags.DEFINE_integer ('beam_width',       1024,        'beam width used in the CTC decoder when building candidate transcriptions')
-    tf.app.flags.DEFINE_float   ('lm_alpha',         1.50,        'the alpha hyperparameter of the CTC decoder. Language Model weight.')
-    tf.app.flags.DEFINE_float   ('lm_beta',          2.10,        'the beta hyperparameter of the CTC decoder. Word insertion weight.')
+    tf.app.flags.DEFINE_float   ('lm_alpha',         0.75,        'the alpha hyperparameter of the CTC decoder. Language Model weight.')
+    tf.app.flags.DEFINE_float   ('lm_beta',          1.85,        'the beta hyperparameter of the CTC decoder. Word insertion weight.')
 
     # Inference mode