offline transducer: treat unk as blank (#1005)

Co-authored-by: chungyi.li <chungyi.li@ailabs.tw>
2024-06-19 20:52:42 +08:00
parent a11c859971
commit 675fb1574f
5 changed files with 25 additions and 9 deletions
--- a/sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.cc
+++ b/sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.cc
@@ -57,7 +57,9 @@ OfflineTransducerGreedySearchDecoder::Decode(Ort::Value encoder_out,
          std::max_element(static_cast<const float *>(p_logit),
                           static_cast<const float *>(p_logit) + vocab_size)));
      p_logit += vocab_size;
-      if (y != 0) {
+      // blank id is hardcoded to 0
+      // also, it treats unk as blank
+      if (y != 0 && y != unk_id_) {
        ans[i].tokens.push_back(y);
        ans[i].timestamps.push_back(t);
        emitted = true;