Support getting word IDs for CTC HLG decoding. (#978)

@@ -15,8 +15,16 @@ struct OfflineCtcDecoderResult {
   /// The decoded token IDs
   std::vector<int64_t> tokens;
 
+  /// The decoded word IDs
+  /// Note: tokens.size() is usually not equal to words.size()
+  /// words is empty for greedy search decoding.
+  /// it is not empty when an HLG graph or an HLG graph is used.
+  std::vector<int32_t> words;
+
   /// timestamps[i] contains the output frame index where tokens[i] is decoded.
   /// Note: The index is after subsampling
+  ///
+  /// tokens.size() == timestamps.size()
   std::vector<int32_t> timestamps;
 };
 

@@ -108,6 +108,9 @@ static OfflineCtcDecoderResult DecodeOne(kaldi_decoder::FasterDecoder *decoder,
     // -1 here since the input labels are incremented during graph
     // construction
     r.tokens.push_back(arc.ilabel - 1);
+    if (arc.olabel != 0) {
+      r.words.push_back(arc.olabel);
+    }
 
     r.timestamps.push_back(t);
     prev = arc.ilabel;

@@ -64,10 +64,6 @@ OfflineParaformerGreedySearchDecoder::Decode(
 
       if (timestamps.size() == results[i].tokens.size()) {
         results[i].timestamps = std::move(timestamps);
-      } else {
-        SHERPA_ONNX_LOGE("time stamp for batch: %d, %d vs %d", i,
-                         static_cast<int32_t>(results[i].tokens.size()),
-                         static_cast<int32_t>(timestamps.size()));
       }
     }
   }

@@ -65,6 +65,8 @@ static OfflineRecognitionResult Convert(const OfflineCtcDecoderResult &src,
     r.timestamps.push_back(time);
   }
 
+  r.words = std::move(src.words);
+
   return r;
 }
 

@@ -339,6 +339,20 @@ std::string OfflineRecognitionResult::AsJsonString() const {
     }
     sep = ", ";
   }
+  os << "], ";
+
+  sep = "";
+
+  os << "\""
+     << "words"
+     << "\""
+     << ": ";
+  os << "[";
+  for (int32_t w : words) {
+    os << sep << w;
+    sep = ", ";
+  }
+
   os << "]";
   os << "}";
 

@@ -30,6 +30,8 @@ struct OfflineRecognitionResult {
   /// timestamps[i] records the time in seconds when tokens[i] is decoded.
   std::vector<float> timestamps;
 
+  std::vector<int32_t> words;
+
   std::string AsJsonString() const;
 };
 

@@ -22,8 +22,16 @@ struct OnlineCtcDecoderResult {
   /// The decoded token IDs
   std::vector<int64_t> tokens;
 
+  /// The decoded word IDs
+  /// Note: tokens.size() is usually not equal to words.size()
+  /// words is empty for greedy search decoding.
+  /// it is not empty when an HLG graph or an HLG graph is used.
+  std::vector<int32_t> words;
+
   /// timestamps[i] contains the output frame index where tokens[i] is decoded.
   /// Note: The index is after subsampling
+  ///
+  /// tokens.size() == timestamps.size()
   std::vector<int32_t> timestamps;
 
   int32_t num_trailing_blanks = 0;

@@ -51,9 +51,9 @@ static void DecodeOne(const float *log_probs, int32_t num_rows,
     bool ok = decoder->GetBestPath(&fst_out);
     if (ok) {
       std::vector<int32_t> isymbols_out;
-      std::vector<int32_t> osymbols_out_unused;
-      ok = fst::GetLinearSymbolSequence(fst_out, &isymbols_out,
-                                        &osymbols_out_unused, nullptr);
+      std::vector<int32_t> osymbols_out;
+      ok = fst::GetLinearSymbolSequence(fst_out, &isymbols_out, &osymbols_out,
+                                        nullptr);
       std::vector<int64_t> tokens;
       tokens.reserve(isymbols_out.size());
 
@@ -83,6 +83,7 @@ static void DecodeOne(const float *log_probs, int32_t num_rows,
       }
 
       result->tokens = std::move(tokens);
+      result->words = std::move(osymbols_out);
       result->timestamps = std::move(timestamps);
       // no need to set frame_offset
     }

@@ -59,6 +59,7 @@ static OnlineRecognizerResult Convert(const OnlineCtcDecoderResult &src,
   }
 
   r.segment = segment;
+  r.words = std::move(src.words);
   r.start_time = frames_since_start * frame_shift_ms / 1000.;
 
   return r;

@@ -22,14 +22,16 @@ namespace sherpa_onnx {
 template <typename T>
 std::string VecToString(const std::vector<T> &vec, int32_t precision = 6) {
   std::ostringstream oss;
+  if (precision != 0) {
     oss << std::fixed << std::setprecision(precision);
-  oss << "[ ";
+  }
+  oss << "[";
   std::string sep = "";
   for (const auto &item : vec) {
     oss << sep << item;
     sep = ", ";
   }
-  oss << " ]";
+  oss << "]";
   return oss.str();
 }
 
@@ -38,26 +40,29 @@ template <>  // explicit specialization for T = std::string
 std::string VecToString<std::string>(const std::vector<std::string> &vec,
                                      int32_t) {  // ignore 2nd arg
   std::ostringstream oss;
-  oss << "[ ";
+  oss << "[";
   std::string sep = "";
   for (const auto &item : vec) {
     oss << sep << "\"" << item << "\"";
     sep = ", ";
   }
-  oss << " ]";
+  oss << "]";
   return oss.str();
 }
 
 std::string OnlineRecognizerResult::AsJsonString() const {
   std::ostringstream os;
   os << "{ ";
-  os << "\"text\": " << "\"" << text << "\"" << ", ";
+  os << "\"text\": "
+     << "\"" << text << "\""
+     << ", ";
   os << "\"tokens\": " << VecToString(tokens) << ", ";
   os << "\"timestamps\": " << VecToString(timestamps, 2) << ", ";
   os << "\"ys_probs\": " << VecToString(ys_probs, 6) << ", ";
   os << "\"lm_probs\": " << VecToString(lm_probs, 6) << ", ";
   os << "\"context_scores\": " << VecToString(context_scores, 6) << ", ";
   os << "\"segment\": " << segment << ", ";
+  os << "\"words\": " << VecToString(words, 0) << ", ";
   os << "\"start_time\": " << std::fixed << std::setprecision(2) << start_time
      << ", ";
   os << "\"is_final\": " << (is_final ? "true" : "false");

@@ -47,6 +47,8 @@ struct OnlineRecognizerResult {
   /// log-domain scores from "hot-phrase" contextual boosting
   std::vector<float> context_scores;
 
+  std::vector<int32_t> words;
+
   /// ID of this segment
   /// When an endpoint is detected, it is incremented
   int32_t segment = 0;

@@ -34,6 +34,8 @@ static void PybindOfflineRecognitionResult(py::module *m) {  // NOLINT
           })
       .def_property_readonly("tokens",
                              [](const PyClass &self) { return self.tokens; })
+      .def_property_readonly("words",
+                             [](const PyClass &self) { return self.words; })
       .def_property_readonly(
           "timestamps", [](const PyClass &self) { return self.timestamps; });
 }

@@ -41,6 +41,9 @@ static void PybindOnlineRecognizerResult(py::module *m) {
       .def_property_readonly(
           "segment", [](PyClass &self) -> int32_t { return self.segment; })
       .def_property_readonly(
+          "words",
+          [](PyClass &self) -> std::vector<int32_t> { return self.words; })
+      .def_property_readonly(
           "is_final", [](PyClass &self) -> bool { return self.is_final; })
       .def("__str__", &PyClass::AsJsonString,
            py::call_guard<py::gil_scoped_release>())