Fix Qwen3ASR hotwords handling

Author: csukuangfj

dotnet-examples/non-streaming-qwen3-asr-decode-files/Program.cs

@@ -1,33 +1,33 @@
-// Copyright (c)  2026  Xiaomi Corporation
-//
-// This file shows how to use a Qwen3 ASR model for speech recognition.
-//
-// You can find the model doc at
-// https://k2-fsa.github.io/sherpa/onnx/qwen3-asr.html
-using SherpaOnnx;
-
-class NonStreamingQwen3Asr
-{
-  static void Main(string[] args)
-  {
-    // please download model files from
-    // https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
-    var config = new OfflineRecognizerConfig();
-    config.ModelConfig.Qwen3Asr.ConvFrontend = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/conv_frontend.onnx";
-    config.ModelConfig.Qwen3Asr.Encoder = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/encoder.int8.onnx";
-    config.ModelConfig.Qwen3Asr.Decoder = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/decoder.int8.onnx";
-    config.ModelConfig.Qwen3Asr.Tokenizer = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/tokenizer";
-    config.ModelConfig.Qwen3Asr.Hotwords = "";
-    config.ModelConfig.Tokens = "";
-    config.ModelConfig.Debug = 1;
-    var recognizer = new OfflineRecognizer(config);
-
-    var testWaveFilename = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/test_wavs/raokouling.wav";
-    var reader = new WaveReader(testWaveFilename);
-    var stream = recognizer.CreateStream();
-    stream.AcceptWaveform(reader.SampleRate, reader.Samples);
-    recognizer.Decode(stream);
-    var text = stream.Result.Text;
-    Console.WriteLine("Text: {0}", text);
-  }
-}
+// Copyright (c)  2026  Xiaomi Corporation
+//
+// This file shows how to use a Qwen3 ASR model for speech recognition.
+//
+// You can find the model doc at
+// https://k2-fsa.github.io/sherpa/onnx/qwen3-asr
+using SherpaOnnx;
+
+class NonStreamingQwen3Asr
+{
+  static void Main(string[] args)
+  {
+    // please download model files from
+    // https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
+    var config = new OfflineRecognizerConfig();
+    config.ModelConfig.Qwen3Asr.ConvFrontend = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/conv_frontend.onnx";
+    config.ModelConfig.Qwen3Asr.Encoder = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/encoder.int8.onnx";
+    config.ModelConfig.Qwen3Asr.Decoder = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/decoder.int8.onnx";
+    config.ModelConfig.Qwen3Asr.Tokenizer = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/tokenizer";
+    config.ModelConfig.Qwen3Asr.Hotwords = "";
+    config.ModelConfig.Tokens = "";
+    config.ModelConfig.Debug = 1;
+    var recognizer = new OfflineRecognizer(config);
+
+    var testWaveFilename = "./sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/test_wavs/raokouling.wav";
+    var reader = new WaveReader(testWaveFilename);
+    var stream = recognizer.CreateStream();
+    stream.AcceptWaveform(reader.SampleRate, reader.Samples);
+    recognizer.Decode(stream);
+    var text = stream.Result.Text;
+    Console.WriteLine("Text: {0}", text);
+  }
+}

sherpa-onnx/csrc/offline-recognizer-qwen3-asr-impl.cc

@@ -709,7 +709,8 @@ OfflineRecognitionResult OfflineRecognizerQwen3ASRImpl::GenerateText(
   // Optional per-stream hotwords via SetOption("hotwords", comma-separated
   // CSV).
   const std::string hotwords = Qwen3FormatHotwordsForPrompt(
-      stream->HasOption("hotwords") ? stream->GetOption("hotwords") : "");
+      stream->HasOption("hotwords") ? stream->GetOption("hotwords")
+                                    : qwen3_config.hotwords);
 
   std::string language;
   if (stream->HasOption("language")) {
@@ -981,20 +982,18 @@ OfflineRecognitionResult OfflineRecognizerQwen3ASRImpl::GenerateText(
   std::vector<int64_t> cleaned_ids = generated_ids;
   if (!generated_ids.empty()) {
     const size_t prefix_window = std::min<size_t>(16, generated_ids.size());
-    auto asr_text_it = std::find(generated_ids.begin(),
-                                 generated_ids.begin() + prefix_window,
-                                 asr_text_token_id_);
+    auto asr_text_it =
+        std::find(generated_ids.begin(), generated_ids.begin() + prefix_window,
+                  asr_text_token_id_);
 
     // Only strip a leading scaffold prefix recognized by token ID.
     if (asr_text_it != generated_ids.begin() + prefix_window &&
         asr_text_it != generated_ids.begin()) {
       std::vector<int64_t> prefix_ids(generated_ids.begin(),
                                       std::next(asr_text_it));
       std::string prefix_text = tokenizer_->Decode(prefix_ids);
-      if (prefix_text.rfind("language ", 0) == 0 &&
-          prefix_text.size() >= 10 &&
-          prefix_text.compare(prefix_text.size() - 10, 10, "<asr_text>") ==
-              0) {
+      if (prefix_text.rfind("language ", 0) == 0 && prefix_text.size() >= 10 &&
+          prefix_text.compare(prefix_text.size() - 10, 10, "<asr_text>") == 0) {
         cleaned_ids.assign(std::next(asr_text_it), generated_ids.end());
       }
     }