Merged with 'master' (#3156)

* [scripts] Add fix regarding num-jobs for segment_long_utterances*.sh(#3130)

* [src] Enable allow_{upsample,downsample} with online features (#3139)

* [src] Fix bad assert in fstmakecontextsyms (#3142)

* [src] Fix to "Fixes to grammar-fst & LM-disambig symbols" (#3000) (#3143)

* [build] Make sure PaUtils exported from portaudio (#3144)

* [src] cudamatrix: fixing a synchronization bug in 'normalize-per-row' (#3145)

was only apparent using large matrices

* [src] Fix typo in comment (#3147)

* [src] Add binary that functions as a TCP server (#2938)

* [scripts] Fix bug in comment (#3152)

egs/wsj/s5/steps/cleanup/segment_long_utterances.sh

@@ -174,10 +174,17 @@ if [ $stage -le 3 ]; then
   cp $srcdir/phones.txt $dir 2>/dev/null || true
 
   mkdir -p $graph_dir
+
+  n_reco=$(cat $text | wc -l) || exit 1
+  nj_reco=$nj
+
+  if [ $nj -gt $n_reco ]; then
+    nj_reco=$n_reco
+  fi
 
   # Make graphs w.r.t. to the original text (usually recording-level)
   steps/cleanup/make_biased_lm_graphs.sh $graph_opts \
-    --nj $nj --cmd "$cmd" $text \
+    --nj $nj_reco --cmd "$cmd" $text \
     $lang $dir $dir/graphs
   if [ -z "$utt2text" ]; then
     # and then copy it to the sub-segments.

egs/wsj/s5/steps/cleanup/segment_long_utterances_nnet3.sh

@@ -235,10 +235,17 @@ if [ $stage -le 3 ]; then
 
   mkdir -p $graph_dir
 
+  n_reco=$(cat $text | wc -l) || exit 1
+  nj_reco=$nj
+
+  if [ $nj -gt $n_reco ]; then
+    nj_reco=$n_reco
+  fi
+
   # Make graphs w.r.t. to the original text (usually recording-level)
   steps/cleanup/make_biased_lm_graphs.sh $graph_opts \
     --scale-opts "$scale_opts" \
-    --nj $nj --cmd "$cmd" $text \
+    --nj $nj_reco --cmd "$cmd" $text \
     $lang $dir $dir/graphs
   if [ -z "$utt2text" ]; then
     # and then copy it to the sub-segments.

egs/wsj/s5/utils/parse_options.sh

@@ -42,7 +42,7 @@ done
 
 
 ###
-### No we process the command line options
+### Now we process the command line options
 ###
 while true; do
   [ -z "${1:-}" ] && break;  # break if there are no arguments

src/cudamatrix/cu-kernels.cu

@@ -2552,9 +2552,12 @@ static void _normalize_per_row(Real *y, int y_stride, const Real *x,
   const int i = blockIdx.x;
   const int tid = threadIdx.x;
   const Real* x_row = x + i * x_d.stride;
+
   typedef cub::BlockReduce<Real, CU1DBLOCK> BlockReduceT;
   __shared__ typename BlockReduceT::TempStorage temp_storage;
-  __shared__ Real ssum[CU1DBLOCK];
+
+  __shared__ Real stddev_div_target_rms;
+  __shared__ Real scale;
 
   // Reduce x_j^2 to CU1DBLOCK elements per row
   Real tsum = Real(0);
@@ -2563,14 +2566,14 @@ static void _normalize_per_row(Real *y, int y_stride, const Real *x,
   }
   tsum = BlockReduceT(temp_storage).Sum(tsum);
   __syncthreads();
-
 
-  const Real kSquaredNormFloor = 1.3552527156068805425e-20; // 2^-66
-  ssum[tid] = sqrt(
-    fmax(tsum / (target_rms * target_rms * x_d.cols), kSquaredNormFloor));
-
-  const Real stddev_div_target_rms = ssum[0];
-  const Real scale = Real(1) / stddev_div_target_rms;
+  if (tid == 0) {
+    const Real kSquaredNormFloor = 1.3552527156068805425e-20; // 2^-66
+    stddev_div_target_rms = sqrt(
+      fmax(tsum / (target_rms * target_rms * x_d.cols), kSquaredNormFloor));
+    scale = Real(1) / stddev_div_target_rms;
+  }
+  __syncthreads();
 
   // Store normalized input to output
   Real* y_row = y + i * y_stride;

src/cudamatrix/cu-math-test.cc

@@ -545,6 +545,50 @@ static void UnitTestCuMathNormalizePerRow() {
   }
 }
 
+
+template<typename Real>
+static void UnitTestCuMathNormalizePerRow_v2() {
+
+  int row = 128;
+  int col = 1024;
+
+  Matrix<Real> Hi(row,col);
+  Matrix<Real> Ho(row,col);
+  Hi.SetRandn();
+  Hi.Scale(5.0);
+  Hi.ApplyFloor(0.0); // like ReLU,
+
+  CuMatrix<Real> Di(row, col);
+  CuMatrix<Real> Do(row, col);
+  Di.CopyFromMat(Hi);
+
+  Real target_rms = 0.3456;
+  bool add_log_stddev = false;
+  const Real kSquaredNormFloor = 1.35525271560688e-20; // 2^-66
+
+  //gpu
+  cu::NormalizePerRow(Di, target_rms, add_log_stddev, &Do);
+
+  //cpu
+  {
+    MatrixBase<Real>& in(Hi);
+    MatrixBase<Real>& out(Ho);
+    Real target_rms=0.3456;
+    Vector<Real> in_norm(in.NumRows());
+    Real d_scaled = in.NumCols() * target_rms * target_rms;
+    in_norm.AddDiagMat2(1.0 / d_scaled, in, kNoTrans, 0.0);
+    in_norm.ApplyFloor(kSquaredNormFloor);
+    in_norm.ApplyPow(-0.5);
+    out.CopyFromMat(in);
+    out.MulRowsVec(in_norm);
+  }
+
+  Matrix<Real> Ho2(Do);
+  // here the BUG was detected (by processing big-enough matrix),
+  AssertEqual(Ho,Ho2,0.00001);
+}
+
+
 template<typename Real>
 static void UnitTestCuDiffNormalizePerRow() {
   for (int32 i = 0; i < 2; i++) {
@@ -660,6 +704,7 @@ template<typename Real> void CudaMathUnitTest() {
   UnitTestEnsureNonzero<Real>();
   UnitTestBackpropLstmNonlinearity<Real>();
   UnitTestCuMathNormalizePerRow<Real>();
+  UnitTestCuMathNormalizePerRow_v2<Real>();
   UnitTestCuDiffNormalizePerRow<Real>();
 }
 
@@ -673,9 +718,9 @@ int main() {
   for (; loop < 2; loop++) {
     CuDevice::Instantiate().SetDebugStrideMode(true);
     if (loop == 0)
-      CuDevice::Instantiate().SelectGpuId("no"); // -1 means no GPU
+      CuDevice::Instantiate().SelectGpuId("no"); // 0 means no GPU
     else
-      CuDevice::Instantiate().SelectGpuId("yes"); // -2 .. automatic selection
+      CuDevice::Instantiate().SelectGpuId("yes"); // 1 .. automatic selection
 #endif
     srand(time(NULL));
     kaldi::CudaMathUnitTest<float>();

src/decoder/grammar-fst.cc

@@ -706,7 +706,7 @@ bool GrammarFstPreparer::IsEntryState(StateId s) const {
     // we check that at least one has label with nonterminal equal to #nonterm_begin...
     // in fact they will all have this value if at least one does, and this was checked
     // in NeedEpsilons().
-    if (nonterminal == kNontermBegin)
+    if (nonterminal == GetPhoneSymbolFor(kNontermBegin))
       return true;
   }
   return false;

src/doc/online_decoding.dox

@@ -438,6 +438,70 @@ and downloadable models that can be used with online nnet3 decoding, please
 see http://kaldi-asr.org/models.html (the first model there, the ASPIRE model,
 includes instructions in a README file).
 
+\subsection online_decoding_nnet3_tcp TCP server for nnet3 online decoding
+
+The program to run the TCP sever is online2-tcp-nnet3-decode-faster located in the
+~/src/online2bin folder. The usage is as follows:
+
+\verbatim
+online2-tcp-nnet3-decode-faster <nnet3-in> <fst-in> <word-symbol-table> <listen-port>
+\endverbatim
+
+For example:
+
+\verbatim
+online2-tcp-nnet3-decode-faster model/final.mdl graph/HCLG.fst graph/words.txt 5050
+\endverbatim
+
+The word symbol table is mandatory (unlike other nnet3 online decoding programs) because
+the server outputs word strings. Endpointing is mandatory to make the operation of the
+program reasonable. Other, non-standard options include:
+    - samp-freq - sampling frequency of audio (usually 8000 for telephony and 16000 for other uses)
+    - chunk-length - length of signal being processed by decoder at each step
+    - output-period - how often we check for changes in the decoding (ie. output refresh rate, default 1s)
+    - num-threads-startup - number of threads used when initializing iVector extractor
+
+The TCP protocol simply takes RAW signal on input (16-bit signed integer
+encoding at chosen sampling frequency) and outputs simple text using the following
+logic:
+    - each refresh period (output-freq argument) the current state of decoding is output
+    - each line is terminated by '\r'
+    - once an utterance boundary is detected due to endpointing a '\n' char is output
+
+Each output string (delimited by '\r') should be treated as uncertain and can change
+entirely until the utterance delimiter ('\n') is sent. The delimiter chars are chosen
+specifically in order to make the output look neat in the terminal. It is possible to
+use it with other interfaces and a web demo (HTML/JS AudioAPI+WebSockets) exists.
+
+To run the program from the terminal you can use one of the following commands. First,
+make sure the server is running and accepting connections. Using the Aspire models, the
+command should look like this:
+\verbatim
+online2-tcp-nnet3-decode-faster --samp-freq=8000 --frames-per-chunk=20 --extra-left-context-initial=0
+    --frame-subsampling-factor=3 --config=model/conf/online.conf --min-active=200 --max-active=7000
+    --beam=15.0 --lattice-beam=6.0 --acoustic-scale=1.0 model/final.mdl graph/HCLG.fst graph/words.txt 5050
+\endverbatim
+
+To send a WAV file into the server, it first needs to be decoded into raw audio, then it can be
+sent to the socket:
+\verbatim
+sox audio.wav -t raw -c 1 -b 16 -r 8k -e signed-integer - | nc -N localhost 5050
+\endverbatim
+
+It is possible to play audio (almost) simultaneously as decoding. It may require installing the
+'pv' program (used to throttle the signal into Kaldi at the same speed as the playback):
+
+\verbatim
+sox audio.wav -t raw -c 1 -b 16 -r 8k -e signed-integer - | \
+    tee >(play -t raw -r 8k -e signed-integer -b 16 -c 1 -q -) | \
+    pv -L 16000 -q | nc -N localhost 5050
+\endverbatim
+
+Finally, it is possible to send audio from the microphone directly into the server:
+
+\verbatim
+rec -r 8k -e signed-integer -c 1 -b 16 -t raw -q - | nc -N localhost 5050
+\endverbatim
 
 
 */