Add vocab_log_probs to CTC and offline NeMo transducer decoders

Captures full vocabulary distribution for CTC (SenseVoice, MedASR,
paraformer, zipformer) and offline NeMo transducer (parakeet-tdt,
nemotron, nemo-fastconformer) decoders. Propagates through Convert
functions and exposes via Python pybind11 binding.

Addresses reviewer feedback to put log_probs inside the result struct.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Dokotela

sherpa-onnx/csrc/offline-ctc-decoder.h

@@ -31,6 +31,11 @@ struct OfflineCtcDecoderResult {
   /// May be empty if not provided by the decoder.
   /// If populated, token_log_probs.size() == tokens.size()
   std::vector<float> token_log_probs;
+
+  /// Full vocabulary log probabilities at each decoded token position.
+  /// vocab_log_probs[i] has vocab_size entries for tokens[i].
+  /// May be empty if not captured by the decoder.
+  std::vector<std::vector<float>> vocab_log_probs;
 };
 
 class OfflineCtcDecoder {

sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.cc

@@ -35,13 +35,15 @@ std::vector<OfflineCtcDecoderResult> OfflineCtcGreedySearchDecoder::Decode(
       int32_t idx = MaxElementIndex(p_log_probs, vocab_size);
       auto y = static_cast<int64_t>(idx);
       float log_prob = p_log_probs[idx];
-      p_log_probs += vocab_size;
 
       if (y != blank_id_ && y != prev_id) {
         r.tokens.push_back(y);
         r.timestamps.push_back(t);
         r.token_log_probs.push_back(log_prob);
+        r.vocab_log_probs.emplace_back(p_log_probs, p_log_probs + vocab_size);
       }
+
+      p_log_probs += vocab_size;
       prev_id = y;
     }  // for (int32_t t = 0; ...)
 

sherpa-onnx/csrc/offline-recognizer-ctc-impl.h

@@ -62,6 +62,9 @@ OfflineRecognitionResult Convert(const OfflineCtcDecoderResult &src,
     if (i < static_cast<int32_t>(src.token_log_probs.size())) {
       r.token_log_probs.push_back(src.token_log_probs[i]);
     }
+    if (i < static_cast<int32_t>(src.vocab_log_probs.size())) {
+      r.vocab_log_probs.push_back(src.vocab_log_probs[i]);
+    }
   }
 
   if (sym_table.IsByteBpe()) {

sherpa-onnx/csrc/offline-recognizer-sense-voice-impl.h

@@ -58,6 +58,9 @@ OfflineRecognitionResult ConvertSenseVoiceResult(
   for (int32_t i = start; i < static_cast<int32_t>(src.token_log_probs.size()); ++i) {
     r.token_log_probs.push_back(src.token_log_probs[i]);
   }
+  for (int32_t i = start; i < static_cast<int32_t>(src.vocab_log_probs.size()); ++i) {
+    r.vocab_log_probs.push_back(src.vocab_log_probs[i]);
+  }
 
   if (!is_funasr_nano) {
     // parse lang, emotion and event from tokens.

sherpa-onnx/csrc/offline-recognizer-transducer-impl.h

@@ -75,6 +75,9 @@ static OfflineRecognitionResult Convert(
   // Copy token log probabilities (confidence scores)
   r.ys_log_probs = src.ys_log_probs;
 
+  // Copy full vocabulary log probabilities
+  r.vocab_log_probs = src.vocab_log_probs;
+
   return r;
 }
 

sherpa-onnx/csrc/offline-transducer-decoder.h

@@ -27,6 +27,11 @@ struct OfflineTransducerDecoderResult {
 
   /// ys_log_probs[i] contains the log probability (confidence) for tokens[i].
   std::vector<float> ys_log_probs;
+
+  /// Full vocabulary log probabilities at each decoded token position.
+  /// vocab_log_probs[i] has vocab_size entries for tokens[i].
+  /// May be empty if not captured by the decoder.
+  std::vector<std::vector<float>> vocab_log_probs;
 };
 
 class OfflineTransducerDecoder {

sherpa-onnx/csrc/offline-transducer-greedy-search-nemo-decoder.cc

@@ -85,6 +85,7 @@ static OfflineTransducerDecoderResult DecodeOne(
         ans.tokens.push_back(y);
         ans.timestamps.push_back(t);
         ans.ys_log_probs.push_back(log_prob);
+        ans.vocab_log_probs.emplace_back(p_logit, p_logit + vocab_size);
 
         decoder_input_pair = BuildDecoderInput(y, model->Allocator());
 
@@ -169,6 +170,8 @@ static OfflineTransducerDecoderResult DecodeOneTDT(
       ans.timestamps.push_back(t);
       ans.durations.push_back(skip);
       ans.ys_log_probs.push_back(log_prob);
+      ans.vocab_log_probs.emplace_back(
+          token_logits_copy.begin(), token_logits_copy.end());
 
       decoder_input_pair = BuildDecoderInput(y, model->Allocator());
 

sherpa-onnx/python/csrc/offline-stream.cc

@@ -48,6 +48,8 @@ static void PybindOfflineRecognitionResult(py::module *m) {  // NOLINT
         [](const PyClass &self) { return self.durations; })
       .def_property_readonly("ys_log_probs",
         [](const PyClass &self) { return self.ys_log_probs; })
+      .def_property_readonly("vocab_log_probs",
+        [](const PyClass &self) { return self.vocab_log_probs; })
       .def_property_readonly("segment_timestamps",
         [](const PyClass &self) { return self.segment_timestamps; })
       .def_property_readonly("segment_durations",