fix(audio): ignore no-speech ASR control token

src/granite_switch/vllm/audio/asr.py

@@ -199,6 +199,7 @@ def _load_chunking():
 
 # Sample rate every supported ASR front-end expects.
 _TARGET_SAMPLE_RATE = 16_000
+_NO_SPEECH_TEXT = "<|nospeech|>"
 
 # Audio item shapes vLLM may pass to a multimodal processor.
 AudioInput = Union[
@@ -320,7 +321,11 @@ def _run_pipeline(
             **call_kwargs,
         )
         text = result["text"] if isinstance(result, dict) else str(result)
-        return text.strip()
+        text = text.strip()
+        # Some multilingual CTC checkpoints emit this control token for
+        # silence or non-speech. Treat it as an empty transcript so it cannot
+        # be inserted into the LLM prompt as if it were user speech.
+        return "" if text == _NO_SPEECH_TEXT else text
 
 
 # ── Module-level cache + convenience function ────────────────────────────────

tests/unit/test_asr.py

@@ -113,6 +113,14 @@ def test_transcribe_strips_and_uses_target_rate(self):
         passed = fake_pipe.call_args_list[-1][0][0]
         assert passed["sampling_rate"] == 16000
 
+    def test_transcribe_drops_no_speech_control_token(self):
+        t = asr.ASRTranscriber(model_id="x", device="cpu")
+        t._pipeline = mock.Mock(return_value={"text": "  <|nospeech|>  "})
+
+        out = t.transcribe(np.zeros(1600, dtype=np.float32), sampling_rate=16000)
+
+        assert out == ""
+
     def test_load_is_idempotent_when_pipeline_set(self):
         # Once the pipeline is loaded, load() must early-return (no rebuild).
         t = asr.ASRTranscriber(model_id="x", device="cpu")