fill lm_probs/context_scores only if LM/ContextGraph is present (make Result smaller)

Author: KarelVesely84

sherpa-onnx/csrc/hypothesis.h

@@ -36,10 +36,12 @@ struct Hypothesis {
 
   // lm_probs[i] contains the lm score for each token in ys.
   // Used only in transducer mofified beam-search.
+  // Elements filled only if LM is used.
   std::vector<float> lm_probs;
 
   // context_scores[i] contains the context-graph score for each token in ys.
   // Used only in transducer mofified beam-search.
+  // Elements filled only if `ContextGraph` is used.
   std::vector<float> context_scores;
 
   // The total score of ys in log space.

sherpa-onnx/csrc/online-recognizer.cc

@@ -20,7 +20,7 @@ namespace sherpa_onnx {
 
 /// Helper for `OnlineRecognizerResult::AsJsonString()`
 template<typename T>
-const std::string& VecToString(const std::vector<T>& vec, int32_t precision = 6) {
+std::string VecToString(const std::vector<T>& vec, int32_t precision = 6) {
   std::ostringstream oss;
   oss << std::fixed << std::setprecision(precision);
   oss << "[ ";
@@ -35,9 +35,8 @@ const std::string& VecToString(const std::vector<T>& vec, int32_t precision = 6)
 
 /// Helper for `OnlineRecognizerResult::AsJsonString()`
 template<>  // explicit specialization for T = std::string
-const std::string& VecToString<std::string>(const std::vector<std::string>& vec,
-                                            int32_t) // ignore 2nd arg
-{
+std::string VecToString<std::string>(const std::vector<std::string>& vec,
+                                     int32_t) {  // ignore 2nd arg
   std::ostringstream oss;
   oss << "[ ";
   std::string sep = "";
@@ -57,9 +56,10 @@ std::string OnlineRecognizerResult::AsJsonString() const {
   os << "\"timestamps\": " << VecToString(timestamps, 2) << ", ";
   os << "\"ys_probs\": " << VecToString(ys_probs, 6) << ", ";
   os << "\"lm_probs\": " << VecToString(lm_probs, 6) << ", ";
-  os << "\"constext_scores\": " << VecToString(context_scores, 6) << ", ";
+  os << "\"context_scores\": " << VecToString(context_scores, 6) << ", ";
   os << "\"segment\": " << segment << ", ";
-  os << "\"start_time\": " << std::fixed << std::setprecision(2) << start_time  << ", ";
+  os << "\"start_time\": " << std::fixed << std::setprecision(2)
+     << start_time  << ", ";
   os << "\"is_final\": " << (is_final ? "true" : "false");
   os << "}";
   return os.str();

sherpa-onnx/csrc/online-recognizer.h

@@ -42,7 +42,9 @@ struct OnlineRecognizerResult {
 
   std::vector<float> ys_probs;  //< log-prob scores from ASR model
   std::vector<float> lm_probs;  //< log-prob scores from language model
-  std::vector<float> context_scores;  //< log-domain scores from "hot-phrase" contextual boosting
+                                //
+  /// log-domain scores from "hot-phrase" contextual boosting
+  std::vector<float> context_scores;
 
   /// ID of this segment
   /// When an endpoint is detected, it is incremented

sherpa-onnx/csrc/online-transducer-greedy-search-decoder.cc

@@ -153,7 +153,6 @@ void OnlineTransducerGreedySearchDecoder::Decode(
                                      // probability
         r.ys_probs.push_back(p_logprob[y]);
       }
-
     }
     if (emitted) {
       Ort::Value decoder_input = model_->BuildDecoderInput(*result);

sherpa-onnx/csrc/online-transducer-modified-beam-search-decoder.cc

@@ -190,17 +190,22 @@ void OnlineTransducerModifiedBeamSearchDecoder::Decode(
         if (new_token != 0 && new_token != unk_id_) {
           const Hypothesis& prev_i = prev[hyp_index];
           // subtract 'prev[i]' path scores, which were added before
-          // for getting topk tokens
+          // getting topk tokens
           float y_prob = p_logprob[k] - prev_i.log_prob - prev_i.lm_log_prob;
           new_hyp.ys_probs.push_back(y_prob);
 
-          float lm_prob = new_hyp.lm_log_prob - prev_lm_log_prob;
-          if (lm_scale_ != 0.0) {
-            lm_prob /= lm_scale_;  // remove lm-scale
+          if (lm_) {  // export only when LM is used
+            float lm_prob = new_hyp.lm_log_prob - prev_lm_log_prob;
+            if (lm_scale_ != 0.0) {
+              lm_prob /= lm_scale_;  // remove lm-scale
+            }
+            new_hyp.lm_probs.push_back(lm_prob);
           }
-          new_hyp.lm_probs.push_back(lm_prob);
 
-          new_hyp.context_scores.push_back(context_score);
+          // export only when `ContextGraph` is used
+          if (ss != nullptr && ss[b]->GetContextGraph() != nullptr) {
+            new_hyp.context_scores.push_back(context_score);
+          }
         }
 
         hyps.Add(std::move(new_hyp));

sherpa-onnx/csrc/sherpa-onnx-alsa-offline-speaker-identification.cc

@@ -276,7 +276,7 @@ as the device_name.
       }
     }
 
-    using namespace std::chrono_literals;
+    using std::chrono_literals::20ms;
     std::this_thread::sleep_for(20ms);  // sleep for 20ms
   }
 

sherpa-onnx/python/csrc/online-recognizer.cc

@@ -37,17 +37,16 @@ static void PybindOnlineRecognizerResult(py::module *m) {
           [](PyClass &self) -> std::vector<float> { return self.lm_probs; })
       .def_property_readonly(
           "context_scores",
-          [](PyClass &self) -> std::vector<float> { return self.context_scores; })
-       .def_property_readonly(
+          [](PyClass &self) -> std::vector<float> {
+            return self.context_scores;
+      })
+      .def_property_readonly(
           "segment",
           [](PyClass &self) -> int32_t { return self.segment; })
-       .def_property_readonly(
-          "start_time",
-          [](PyClass &self) -> float { return self.start_time; })
-       .def_property_readonly(
+      .def_property_readonly(
           "is_final",
           [](PyClass &self) -> bool { return self.is_final; })
-       .def("as_json_string", &PyClass::AsJsonString,
+      .def("as_json_string", &PyClass::AsJsonString,
             py::call_guard<py::gil_scoped_release>());
 }