Fix initial tokens to decoding (#246)

 cmake_minimum_required(VERSION 3.13 FATAL_ERROR)
 project(sherpa-onnx)
 
-set(SHERPA_ONNX_VERSION "1.6.1")
+set(SHERPA_ONNX_VERSION "1.6.2")
 
 # Disable warning about
 #

@@ -30,8 +30,9 @@ OfflineTransducerGreedySearchDecoder::Decode(Ort::Value encoder_out,
 
   std::vector<OfflineTransducerDecoderResult> ans(batch_size);
   for (auto &r : ans) {
+    r.tokens.resize(context_size, -1);
     // 0 is the ID of the blank token
-    r.tokens.resize(context_size, 0);
+    r.tokens.back() = 0;
   }
 
   auto decoder_input = model_->BuildDecoderInput(ans, ans.size());

@@ -32,7 +32,8 @@ OfflineTransducerModifiedBeamSearchDecoder::Decode(
   int32_t vocab_size = model_->VocabSize();
   int32_t context_size = model_->ContextSize();
 
-  std::vector<int64_t> blanks(context_size, 0);
+  std::vector<int64_t> blanks(context_size, -1);
+  blanks.back() = 0;
 
   std::deque<Hypotheses> finalized;
   std::vector<Hypotheses> cur;

@@ -55,7 +55,8 @@ OnlineTransducerGreedySearchDecoder::GetEmptyResult() const {
   int32_t context_size = model_->ContextSize();
   int32_t blank_id = 0;  // always 0
   OnlineTransducerDecoderResult r;
-  r.tokens.resize(context_size, blank_id);
+  r.tokens.resize(context_size, -1);
+  r.tokens.back() = blank_id;
 
   return r;
 }

@@ -42,7 +42,9 @@ OnlineTransducerModifiedBeamSearchDecoder::GetEmptyResult() const {
   int32_t context_size = model_->ContextSize();
   int32_t blank_id = 0;  // always 0
   OnlineTransducerDecoderResult r;
-  std::vector<int64_t> blanks(context_size, blank_id);
+  std::vector<int64_t> blanks(context_size, -1);
+  blanks.back() = blank_id;
+
   Hypotheses blank_hyp({{blanks, 0}});
   r.hyps = std::move(blank_hyp);
   r.tokens = std::move(blanks);