python-voiceio 0.9.4__tar.gz → 0.9.5__tar.gz

{python_voiceio-0.9.4/python_voiceio.egg-info → python_voiceio-0.9.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 0.9.4
+Version: 0.9.5
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-0.9.4 → python_voiceio-0.9.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "python-voiceio"
-version = "0.9.4"
+version = "0.9.5"
 description = "Voice dictation for Linux. Speak → text, locally, instantly."
 readme = "README.md"
 license = "MIT"

{python_voiceio-0.9.4 → python_voiceio-0.9.5/python_voiceio.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 0.9.4
+Version: 0.9.5
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-0.9.4 → python_voiceio-0.9.5}/tests/test_vad.py

@@ -51,16 +51,39 @@ class TestSileroVad:
         chunk = np.zeros(1024, dtype=np.float32)
         assert vad.is_speech(chunk) is False
 
-    def test_speech_detected(self, vad):
-        # Generate a sine wave at speech-like frequency
-        t = np.linspace(0, 0.1, 1600, dtype=np.float32)
-        chunk = 0.5 * np.sin(2 * np.pi * 300 * t)
-        # Run several chunks to build state
-        for _ in range(5):
-            result = vad.is_speech(chunk)
-        # At least one should detect speech (sine wave isn't silence)
-        # Note: Silero may not classify pure sine as speech, so we just test it runs
-        assert isinstance(result, bool)
+    def test_v5_feeds_context_window(self, vad):
+        """Regression: Silero v5 consumes 512+64=576 samples — the window plus
+        the previous window's 64-sample tail as context. Feeding a bare 512
+        (the pre-fix bug) makes the model score all real speech as silence
+        (measured: max prob 0.016 vs 1.0 with context)."""
+        if not vad._use_state:
+            pytest.skip("v4 model has no context requirement")
+        seen = []
+        real_run = vad._session.run
+
+        def spy(outs, inputs):
+            seen.append(inputs["input"])
+            return real_run(outs, inputs)
+
+        vad._session.run = spy
+        # Two distinct windows so we can check the context carries across.
+        w1 = np.arange(512, dtype=np.float32) / 512.0
+        w2 = np.arange(512, 1024, dtype=np.float32) / 512.0
+        vad.is_speech(w1)
+        vad.is_speech(w2)
+        assert seen[0].shape[-1] == 512 + 64        # 576, not a bare 512
+        # First call's context is zeros; second call's leading 64 samples must
+        # be the tail of the first window (state carried between steps).
+        assert np.allclose(seen[0][0, :64], 0.0)
+        assert np.allclose(seen[1][0, :64], w1[-64:])
+
+    def test_reset_clears_context(self, vad):
+        if not vad._use_state:
+            pytest.skip("v4 model has no context")
+        vad.is_speech(np.arange(512, dtype=np.float32) / 512.0)
+        assert not np.allclose(vad._context, 0.0)  # context now populated
+        vad.reset()
+        assert np.allclose(vad._context, 0.0)
 
     def test_reset_clears_state(self, vad):
         chunk = np.zeros(1024, dtype=np.float32)

python_voiceio-0.9.5/voiceio/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.9.5"

{python_voiceio-0.9.4 → python_voiceio-0.9.5}/voiceio/vad.py

@@ -16,6 +16,11 @@ log = logging.getLogger(__name__)
 _MODEL_PATH = Path(__file__).parent / "models" / "silero_vad.onnx"
 _WINDOW_SIZE = 512  # Silero expects 512 samples at 16kHz (~32ms)
 _SAMPLE_RATE = 16000
+# Silero v5 consumes 576 samples per step: the 512-sample window PLUS the 64
+# trailing samples of the previous window as context. Feeding a bare 512 makes
+# the model score everything as non-speech (measured: max prob 0.016 vs 1.0
+# with context). The official ONNX wrapper prepends this outside the graph.
+_CONTEXT_SIZE = 64
 
 
 @runtime_checkable
@@ -46,6 +51,8 @@ class SileroVad:
             state_meta = [inp for inp in self._session.get_inputs() if inp.name == "state"][0]
             state_dim = state_meta.shape[2]  # 128 for v5
             self._state = np.zeros((2, 1, state_dim), dtype=np.float32)
+            # v5 context: the previous window's trailing 64 samples.
+            self._context = np.zeros(_CONTEXT_SIZE, dtype=np.float32)
         else:
             self._h = np.zeros((2, 1, 64), dtype=np.float32)
             self._c = np.zeros((2, 1, 64), dtype=np.float32)
@@ -56,6 +63,7 @@ class SileroVad:
     def reset(self) -> None:
         if self._use_state:
             self._state = np.zeros_like(self._state)
+            self._context = np.zeros(_CONTEXT_SIZE, dtype=np.float32)
         else:
             self._h = np.zeros_like(self._h)
             self._c = np.zeros_like(self._c)
@@ -76,13 +84,15 @@ class SileroVad:
         return speech
 
     def _infer(self, window: np.ndarray) -> float:
-        input_data = window.reshape(1, -1)
         sr = np.array(_SAMPLE_RATE, dtype=np.int64)
         if self._use_state:
-            ort_inputs = {"input": input_data, "state": self._state, "sr": sr}
+            # Prepend the previous window's tail; the model needs 512+64=576.
+            x = np.concatenate([self._context, window])
+            self._context = window[-_CONTEXT_SIZE:].copy()
+            ort_inputs = {"input": x.reshape(1, -1), "state": self._state, "sr": sr}
             out, self._state = self._session.run(None, ort_inputs)
         else:
-            ort_inputs = {"input": input_data, "h": self._h, "c": self._c, "sr": sr}
+            ort_inputs = {"input": window.reshape(1, -1), "h": self._h, "c": self._c, "sr": sr}
             out, self._h, self._c = self._session.run(None, ort_inputs)
         return float(out.squeeze())
 

python_voiceio-0.9.4/voiceio/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.9.4"