fix(voice): normalize PCM16 float output correctly

src/agents/voice/result.py

@@ -109,7 +109,7 @@ def _transform_audio_buffer(
         if output_dtype == np.int16:
             return np_array
         elif output_dtype == np.float32:
-            return (np_array.astype(np.float32) / 32767.0).reshape(-1, 1)
+            return (np_array.astype(np.float32) / 32768.0).reshape(-1, 1)
         else:
             raise UserError("Invalid output dtype")
 

tests/voice/test_audio_result_pcm.py

@@ -0,0 +1,23 @@
+import numpy as np
+
+from agents.voice import StreamedAudioResult, TTSModelSettings, VoicePipelineConfig
+
+from .pipeline_test_models import ZeroPcmTTSModel
+
+
+def test_streamed_audio_result_normalizes_pcm16_full_scale_to_float32() -> None:
+    result = StreamedAudioResult(
+        ZeroPcmTTSModel(),
+        TTSModelSettings(dtype=np.float32),
+        VoicePipelineConfig(),
+    )
+    samples = np.array([-32768, 32767], dtype=np.int16)
+
+    transformed = result._transform_audio_buffer([samples.tobytes()], np.float32)
+
+    assert transformed.dtype == np.float32
+    assert transformed.shape == (2, 1)
+    assert transformed[0, 0] == -1.0
+    assert transformed[1, 0] == np.float32(32767 / 32768)
+    assert np.all(transformed >= -1.0)
+    assert np.all(transformed <= 1.0)