audio_recognize res text filter <asr_text>

src/models/common/generate.rs

@@ -140,6 +140,7 @@ pub fn generate_generic_text<M: InferenceModel>(
         }
         input_ids = ctx.prepare_for_next_token(next_token)?;
     }
+    model.clear_cache();
     let text = tokenizer.token_decode(generated)?;
     Ok(text)
 }

src/models/fire_red_vad/vad.rs

@@ -102,7 +102,7 @@ impl FireRedVad {
             Ok(Some(VadFrameResult {
                 is_speech: true,
                 is_i16: true,
-                is_speech_start: false, // TODO: is start speech, asr to clear cache
+                is_speech_start: true, // TODO: is start speech, asr to clear cache
                 orig_audio: Some(audio_frame.clone()),
                 kaldi_audio: Some(feats),
                 model_name: self.model_name.clone(),

src/models/qwen3_asr/generate.rs

@@ -89,9 +89,9 @@ impl<'a> Qwen3AsrGenerateModel<'a> {
         if !vad_res.is_speech || vad_res.orig_audio.is_none() {
             return Ok(AsrResult::init_empty());
         }
-        if vad_res.is_speech_start {
+        // if vad_res.is_speech_start {
-            self.qwen3_asr.clear_kv_cache();
+        //     self.qwen3_asr.clear_kv_cache();
-        }
+        // }
         let audio_data =
             self.processor
                 .process_vad_res(&self.default_template, vad_res, &self.tokenizer)?;
@@ -110,13 +110,17 @@ impl<'a> Qwen3AsrGenerateModel<'a> {
         );
         let data_vec = vec![input_features];
         let data = MultiModalData::new(data_vec);
-        let text = generate_generic_text(
+        let mut text = generate_generic_text(
             &mut self.qwen3_asr,
             &self.tokenizer,
             input_ids,
             data,
             &mut ctx,
         )?;
+        if text.contains("<asr_text>") {
+            let mut split: Vec<&str> = text.split("<asr_text>").collect();
+            text = split.pop().unwrap_or(&text).to_string();
+        }
         Ok(AsrResult::init(text))
     }
 }