more refine

Author: pkufool

c-api-examples/decode-file-c-api.c

@@ -65,7 +65,8 @@ static struct cag_option options[] = {
      .access_name = "hotwords-score",
      .value_name = "hotwords-score",
      .description = "The bonus score for each token in hotwords. Used only "
-                    "when decoding_method is modified_beam_search"},
+                    "when decoding_method is modified_beam_search or "
+                    "prefix_beam_search"},
 };
 
 const char *kUsage =
@@ -80,7 +81,8 @@ const char *kUsage =
     "    /path/to/foo.wav\n"
     "\n\n"
     "Default num_threads is 1.\n"
-    "Valid decoding_method: greedy_search (default), modified_beam_search\n\n"
+    "Valid decoding_method: greedy_search (default), modified_beam_search,\n"
+    "  prefix_beam_search (for CTC models)\n\n"
     "Valid provider: cpu (default), cuda, coreml\n\n"
     "Please refer to \n"
     "https://k2-fsa.github.io/sherpa/onnx/pretrained_models/online-transducer/"

python-api-examples/online-decode-files.py

@@ -172,7 +172,10 @@ def get_args():
         "--decoding-method",
         type=str,
         default="greedy_search",
-        help="Valid values are greedy_search and modified_beam_search",
+        help=(
+            "Valid values are greedy_search, modified_beam_search "
+            "(for transducer), and prefix_beam_search (for CTC)"
+        ),
     )
 
     parser.add_argument(
@@ -368,7 +371,12 @@ def main():
             provider=args.provider,
             sample_rate=16000,
             feature_dim=80,
-            decoding_method="greedy_search",
+            decoding_method=args.decoding_method,
+            max_active_paths=args.max_active_paths,
+            hotwords_file=args.hotwords_file,
+            hotwords_score=args.hotwords_score,
+            modeling_unit=args.modeling_unit,
+            bpe_vocab=args.bpe_vocab,
         )
     elif args.paraformer_encoder:
         recognizer = sherpa_onnx.OnlineRecognizer.from_paraformer(
@@ -391,7 +399,12 @@ def main():
             provider=args.provider,
             sample_rate=16000,
             feature_dim=80,
-            decoding_method="greedy_search",
+            decoding_method=args.decoding_method,
+            max_active_paths=args.max_active_paths,
+            hotwords_file=args.hotwords_file,
+            hotwords_score=args.hotwords_score,
+            modeling_unit=args.modeling_unit,
+            bpe_vocab=args.bpe_vocab,
         )
     else:
         raise ValueError("Please provide a model")

sherpa-onnx/c-api/c-api.h

@@ -340,10 +340,11 @@ typedef struct SherpaOnnxOnlineRecognizerConfig {
   /** Streaming model configuration. */
   SherpaOnnxOnlineModelConfig model_config;
 
-  /** Decoding method, for example "greedy_search" or "modified_beam_search". */
+  /** Decoding method: "greedy_search", "modified_beam_search" (transducer),
+   *  or "prefix_beam_search" (CTC). */
   const char *decoding_method;
 
-  /** Number of active paths for modified beam search. */
+  /** Number of active paths for modified_beam_search / prefix_beam_search. */
   int32_t max_active_paths;
 
   /** Set to non-zero to enable endpoint detection. */
@@ -358,7 +359,8 @@ typedef struct SherpaOnnxOnlineRecognizerConfig {
   /** Endpoint rule 3 utterance-length threshold in seconds. */
   float rule3_min_utterance_length;
 
-  /** Path to a hotwords file. */
+  /** Path to a hotwords file. Used with modified_beam_search or
+   *  prefix_beam_search. */
   const char *hotwords_file;
 
   /** Bonus score added to each hotword token during decoding. */
@@ -1160,12 +1162,14 @@ typedef struct SherpaOnnxOfflineRecognizerConfig {
   /** Optional language model configuration. */
   SherpaOnnxOfflineLMConfig lm_config;
 
-  /** Decoding method, for example "greedy_search" or "modified_beam_search". */
+  /** Decoding method: "greedy_search", "modified_beam_search" (transducer),
+   *  or "prefix_beam_search" (CTC). */
   const char *decoding_method;
-  /** Number of active paths for modified beam search. */
+  /** Number of active paths for modified_beam_search / prefix_beam_search. */
   int32_t max_active_paths;
 
-  /** Path to a hotwords file. */
+  /** Path to a hotwords file. Used with modified_beam_search or
+   *  prefix_beam_search. */
   const char *hotwords_file;
 
   /** Bonus score added to each hotword token. */

sherpa-onnx/csrc/online-ctc-prefix-beam-search-decoder.cc

@@ -1,6 +1,6 @@
 // sherpa-onnx/csrc/online-ctc-prefix-beam-search-decoder.cc
 //
-// Copyright (c)  2024  Xiaomi Corporation
+// Copyright (c)  2026  Xiaomi Corporation
 
 #include "sherpa-onnx/csrc/online-ctc-prefix-beam-search-decoder.h"
 
@@ -34,12 +34,14 @@ static std::vector<Hypothesis> StepWorker(const float *p_log_probs,
         // Prefix does not change, update log_prob of blank
         new_hyp.log_prob_nb = -std::numeric_limits<float>::infinity();
         new_hyp.log_prob_b = hyp.LogProb(true) + log_prob;
+        new_hyp.num_trailing_blanks = hyp.num_trailing_blanks + 1;
         next_hyps.Add(std::move(new_hyp));
       } else if (hyp.ys.size() > 0 && hyp.ys.back() == new_token) {
         // Case 1: *a + a => *a
         // Prefix does not change, update log_prob of non_blank
         new_hyp.log_prob_nb = hyp.log_prob_nb + log_prob;
         new_hyp.log_prob_b = -std::numeric_limits<float>::infinity();
+        new_hyp.num_trailing_blanks = 0;
         next_hyps.Add(std::move(new_hyp));
 
         // Case 2: *aε + a => *aa
@@ -48,13 +50,15 @@ static std::vector<Hypothesis> StepWorker(const float *p_log_probs,
         new_hyp.ys.push_back(new_token);
         new_hyp.log_prob_nb = hyp.log_prob_b + log_prob;
         new_hyp.log_prob_b = -std::numeric_limits<float>::infinity();
+        new_hyp.num_trailing_blanks = 0;
         update_prefix = true;
       } else {
         // Case 3: *a + b => *ab, *aε + b => *ab
         // Prefix changes, update log_prob of non_blank
         new_hyp.ys.push_back(new_token);
         new_hyp.log_prob_nb = hyp.LogProb(true) + log_prob;
         new_hyp.log_prob_b = -std::numeric_limits<float>::infinity();
+        new_hyp.num_trailing_blanks = 0;
         update_prefix = true;
       }
 
@@ -133,11 +137,7 @@ void OnlineCtcPrefixBeamSearchDecoder::Decode(
 
     r.tokens = best_hyp.ys;
     r.frame_offset += num_frames;
-
-    // Count trailing blanks for endpointing
-    if (!best_hyp.ys.empty()) {
-      r.num_trailing_blanks = 0;
-    }
+    r.num_trailing_blanks = best_hyp.num_trailing_blanks;
   }
 }