Add LODR support to online and offline recognizers (#2026)

This PR integrates LODR (Level-Ordered Deterministic Rescoring) support from Icefall into both online and offline recognizers, enabling LODR for LM shallow fusion and LM rescore.

- Extended OnlineLMConfig and OfflineLMConfig to include lodr_fst, lodr_scale, and lodr_backoff_id.
- Implemented LodrFst and LodrStateCost classes and wired them into RNN LM scoring in both online and offline code paths.
- Updated Python bindings, CLI entry points, examples, and CI test scripts to accept and exercise the new LODR options.

sherpa-onnx/python/csrc/offline-lm-config.cc

@@ -13,13 +13,19 @@ namespace sherpa_onnx {
 void PybindOfflineLMConfig(py::module *m) {
   using PyClass = OfflineLMConfig;
   py::class_<PyClass>(*m, "OfflineLMConfig")
-      .def(py::init<const std::string &, float, int32_t, const std::string &>(),
+      .def(py::init<const std::string &, float, int32_t, const std::string &,
+           const std::string &, float, int32_t>(),
            py::arg("model"), py::arg("scale") = 0.5f,
-           py::arg("lm_num_threads") = 1, py::arg("lm_provider") = "cpu")
+           py::arg("lm_num_threads") = 1, py::arg("lm_provider") = "cpu",
+           py::arg("lodr_fst") = "", py::arg("lodr_scale") = 0.0f,
+           py::arg("lodr_backoff_id") = -1)
       .def_readwrite("model", &PyClass::model)
       .def_readwrite("scale", &PyClass::scale)
       .def_readwrite("lm_provider", &PyClass::lm_provider)
       .def_readwrite("lm_num_threads", &PyClass::lm_num_threads)
+      .def_readwrite("lodr_fst", &PyClass::lodr_fst)
+      .def_readwrite("lodr_scale", &PyClass::lodr_scale)
+      .def_readwrite("lodr_backoff_id", &PyClass::lodr_backoff_id)
       .def("__str__", &PyClass::ToString);
 }
 

sherpa-onnx/python/csrc/online-lm-config.cc

@@ -14,15 +14,21 @@ void PybindOnlineLMConfig(py::module *m) {
   using PyClass = OnlineLMConfig;
   py::class_<PyClass>(*m, "OnlineLMConfig")
       .def(py::init<const std::string &, float, int32_t,
-           const std::string &, bool>(),
+           const std::string &, bool, const std::string &,
+           float, int>(),
            py::arg("model") = "", py::arg("scale") = 0.5f,
            py::arg("lm_num_threads") = 1, py::arg("lm_provider") = "cpu",
-           py::arg("shallow_fusion") = true)
+           py::arg("shallow_fusion") = true, py::arg("lodr_fst") = "",
+           py::arg("lodr_scale") = 0.0f, py::arg("lodr_backoff_id") = -1)
       .def_readwrite("model", &PyClass::model)
       .def_readwrite("scale", &PyClass::scale)
       .def_readwrite("lm_provider", &PyClass::lm_provider)
       .def_readwrite("lm_num_threads", &PyClass::lm_num_threads)
       .def_readwrite("shallow_fusion", &PyClass::shallow_fusion)
+      .def_readwrite("lodr_fst", &PyClass::lodr_fst)
+      .def_readwrite("lodr_scale", &PyClass::lodr_scale)
+      .def_readwrite("lodr_backoff_id", &PyClass::lodr_backoff_id)
+
       .def("__str__", &PyClass::ToString);
 }
 

sherpa-onnx/python/sherpa_onnx/offline_recognizer.py

@@ -69,6 +69,8 @@ class OfflineRecognizer(object):
         hr_dict_dir: str = "",
         hr_rule_fsts: str = "",
         hr_lexicon: str = "",
+        lodr_fst: str = "",
+        lodr_scale: float = 0.0,
     ):
         """
         Please refer to
@@ -133,6 +135,10 @@ class OfflineRecognizer(object):
           rule_fars:
             If not empty, it specifies fst archives for inverse text normalization.
             If there are multiple archives, they are separated by a comma.
+          lodr_fst:
+            Path to the LODR FST file in binary format. If empty, LODR is disabled.
+          lodr_scale:
+            Scale factor for LODR rescoring. Only used when lodr_fst is provided.
         """
         self = cls.__new__(cls)
         model_config = OfflineModelConfig(
@@ -173,6 +179,8 @@ class OfflineRecognizer(object):
             scale=lm_scale,
             lm_num_threads=num_threads,
             lm_provider=provider,
+            lodr_fst=lodr_fst,
+            lodr_scale=lodr_scale,
         )
 
         recognizer_config = OfflineRecognizerConfig(

sherpa-onnx/python/sherpa_onnx/online_recognizer.py

@@ -89,6 +89,8 @@ class OnlineRecognizer(object):
         hr_dict_dir: str = "",
         hr_rule_fsts: str = "",
         hr_lexicon: str = "",
+        lodr_fst: str = "",
+        lodr_scale: float = 0.0,
     ):
         """
         Please refer to
@@ -216,6 +218,10 @@ class OnlineRecognizer(object):
             "Set path for storing timing cache." TensorRT EP
           trt_dump_subgraphs: bool = False,
             "Dump optimized subgraphs for debugging." TensorRT EP
+          lodr_fst:
+            Path to the LODR FST file in binary format. If empty, LODR is disabled.
+          lodr_scale:
+            Scale factor for LODR rescoring. Only used when lodr_fst is provided.
         """
         self = cls.__new__(cls)
         _assert_file_exists(tokens)
@@ -298,6 +304,8 @@ class OnlineRecognizer(object):
             model=lm,
             scale=lm_scale,
             shallow_fusion=lm_shallow_fusion,
+            lodr_fst=lodr_fst,
+            lodr_scale=lodr_scale,
         )
 
         recognizer_config = OnlineRecognizerConfig(