Fix lm fusion (#157)

* share GetHypsRowSplits interface and fix getting Topk not taking logprob

* fix lm score of lm fusion and make padding len same with 'icefall/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/decode.py'

@@ -35,7 +35,7 @@ class OnlineRnnLM::Impl {
 
     // get lm score for cur token given the hyp->ys[:-1] and save to lm_log_prob
     const float *nn_lm_scores = hyp->nn_lm_scores.value.GetTensorData<float>();
-    hyp->lm_log_prob = nn_lm_scores[hyp->ys.back()] * scale;
+    hyp->lm_log_prob += nn_lm_scores[hyp->ys.back()] * scale;
 
     // get lm scores for next tokens given the hyp->ys[:] and save to
     // nn_lm_scores

@@ -152,7 +152,9 @@ void OnlineTransducerModifiedBeamSearchDecoder::Decode(
         } else {
           ++new_hyp.num_trailing_blanks;
         }
-        new_hyp.log_prob = p_logprob[k] - prev_lm_log_prob;
+        new_hyp.log_prob =
+            p_logprob[k] - prev_lm_log_prob;  // log_prob only includes the
+                                              // score of the transducer
         hyps.Add(std::move(new_hyp));
       }  // for (auto k : topk)
       cur.push_back(std::move(hyps));

@@ -94,7 +94,7 @@ for a list of pre-trained models to download.
   auto s = recognizer.CreateStream();
   s->AcceptWaveform(sampling_rate, samples.data(), samples.size());
 
-  std::vector<float> tail_paddings(static_cast<int>(0.5 * sampling_rate));
+  std::vector<float> tail_paddings(static_cast<int>(0.3 * sampling_rate));
   // Note: We can call AcceptWaveform() multiple times.
   s->AcceptWaveform(sampling_rate, tail_paddings.data(), tail_paddings.size());