huri-module-vad 0.1.0.dev0__tar.gz → 0.2.0__tar.gz

{huri_module_vad-0.1.0.dev0 → huri_module_vad-0.2.0}/PKG-INFO

@@ -1,11 +1,11 @@
 Metadata-Version: 2.4
 Name: huri-module-vad
-Version: 0.1.0.dev0
+Version: 0.2.0
 Summary: HuRI module: Voice activity detection (WebRTC VAD): microphone frames in, speech segments out
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
-Requires-Dist: huri-core[server]<0.5,>=0.4.0.dev0
-Requires-Dist: huri-events<0.2,>=0.1.0.dev0
+Requires-Dist: huri-core[server]<0.5,>=0.4
+Requires-Dist: huri-events<0.3,>=0.1
 Requires-Dist: numpy<2,>=1.26
 Requires-Dist: webrtcvad==2.0.10
 Requires-Dist: setuptools<82

{huri_module_vad-0.1.0.dev0 → huri_module_vad-0.2.0}/huri_module_vad.egg-info/PKG-INFO

@@ -1,11 +1,11 @@
 Metadata-Version: 2.4
 Name: huri-module-vad
-Version: 0.1.0.dev0
+Version: 0.2.0
 Summary: HuRI module: Voice activity detection (WebRTC VAD): microphone frames in, speech segments out
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
-Requires-Dist: huri-core[server]<0.5,>=0.4.0.dev0
-Requires-Dist: huri-events<0.2,>=0.1.0.dev0
+Requires-Dist: huri-core[server]<0.5,>=0.4
+Requires-Dist: huri-events<0.3,>=0.1
 Requires-Dist: numpy<2,>=1.26
 Requires-Dist: webrtcvad==2.0.10
 Requires-Dist: setuptools<82

{huri_module_vad-0.1.0.dev0 → huri_module_vad-0.2.0}/huri_module_vad.egg-info/SOURCES.txt

@@ -9,4 +9,5 @@ huri_module_vad.egg-info/top_level.txt
 huri_vad/__init__.py
 huri_vad/huri_module.toml
 huri_vad/module.py
-tests/test_vad.py
+tests/test_vad.py
+tests/test_vad_detector.py

huri_module_vad-0.2.0/huri_module_vad.egg-info/requires.txt

@@ -0,0 +1,5 @@
+huri-core[server]<0.5,>=0.4
+huri-events<0.3,>=0.1
+numpy<2,>=1.26
+webrtcvad==2.0.10
+setuptools<82

{huri_module_vad-0.1.0.dev0 → huri_module_vad-0.2.0}/huri_vad/huri_module.toml

@@ -2,6 +2,6 @@ name = "vad"
 summary = "Detects speech in microphone frames (WebRTC VAD)"
 input = { topic = "huri.audio.in", event = "huri_core.core.events:BytesEvent" }
 output = { topic = "huri.voice", event = "huri_events.speech:Voice" }
-events = ["huri-events>=0.1,<0.2"]
+events = ["huri-events>=0.1,<0.3"]
 requires = { gpu = false, actuation = false }
-huri_core = ">=0.4.0.dev0,<0.5"
+huri_core = ">=0.4,<0.5"

{huri_module_vad-0.1.0.dev0 → huri_module_vad-0.2.0}/huri_vad/module.py

@@ -1,4 +1,4 @@
-from typing import Optional
+from typing import AsyncGenerator, Optional
 
 import numpy as np
 import webrtcvad
@@ -19,8 +19,8 @@ class VAD(Module):
     :vad_agressiveness: from 0 (low) to 3 (high, can distord audio).
     :silence_duration: how many seconds will a no voice be considered a silence.
     :sample_rate: size of received chunk of audio. Usually 8000, 16000 or 48000.
-    :block_duration: size of received chunk of audio (in s).
-        Can only be 0.010, 0.020 and 0.030.
+    :block_duration: size of the frames WebRTC analyses (in s), 0.010, 0.020
+        or 0.030. Received chunks of any size are re-cut to it.
     """
 
     # Inbound microphone frames travel on their own topic so the TTS-output
@@ -50,12 +50,26 @@ class VAD(Module):
         self.silence_frames_count: int = -1
 
         self.vad = webrtcvad.Vad(vad_agressiveness)
-
-    async def process(self, data: BytesEvent) -> Optional[Voice]:
-        if self.vad.is_speech(data.data, self.sample_rate) is True:
+        # WebRTC only takes frames of exactly block_size samples; a client may
+        # send any size, so the bytes are re-cut here and a partial frame waits
+        # for the next chunk.
+        self._pending = b""
+
+    async def process(self, data: BytesEvent) -> AsyncGenerator[Voice, None]:
+        self._pending += data.data
+        frame_bytes = self.block_size * 2  # int16
+        while len(self._pending) >= frame_bytes:
+            frame = self._pending[:frame_bytes]
+            self._pending = self._pending[frame_bytes:]
+            voice = self._frame(frame)
+            if voice is not None:
+                yield voice
+
+    def _frame(self, frame: bytes) -> Optional[Voice]:
+        if self.vad.is_speech(frame, self.sample_rate) is True:
             self.silence_frames_count = 0
 
-            audio_array = np.frombuffer(data.data, dtype=np.int16)
+            audio_array = np.frombuffer(frame, dtype=np.int16)
             audio_array_float = audio_array.astype(np.float32) / 32768.0
 
             return Voice(audio_array_float)

{huri_module_vad-0.1.0.dev0 → huri_module_vad-0.2.0}/pyproject.toml

@@ -4,13 +4,13 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "huri-module-vad"
-version = "0.1.0.dev0"
+version = "0.2.0"
 description = "HuRI module: Voice activity detection (WebRTC VAD): microphone frames in, speech segments out"
 readme = "README.md"
 requires-python = ">=3.10"
 dependencies = [
-  "huri-core[server]>=0.4.0.dev0,<0.5",
-  "huri-events>=0.1.0.dev0,<0.2",
+  "huri-core[server]>=0.4,<0.5",
+  "huri-events>=0.1,<0.3",
   "numpy>=1.26,<2",
   "webrtcvad==2.0.10",
   # webrtcvad imports pkg_resources, removed in setuptools 82

{huri_module_vad-0.1.0.dev0 → huri_module_vad-0.2.0}/tests/test_vad.py

@@ -49,3 +49,40 @@ def test_block_duration_must_be_one_webrtc_accepts():
 def test_real_detector_runs_on_a_frame():
     out = ModuleHarness(VAD).run(BytesEvent(bytes(640)))  # digital silence
     assert out == [] or isinstance(out[0], Voice)
+
+
+class Recording(Scripted):
+    def __init__(self, decisions):
+        super().__init__(decisions)
+        self.frames = []
+
+    def is_speech(self, frame, sample_rate):
+        self.frames.append(len(frame))
+        return super().is_speech(frame, sample_rate)
+
+
+def test_chunks_of_any_size_are_recut_into_frames_webrtc_accepts():
+    """WebRTC raises on any other frame size; a client may send any chunk
+    size, down to an odd number of bytes."""
+    harness = ModuleHarness(VAD, silence_duration=0.05)
+    harness.module.vad = detector = Recording([True] * 4)
+    audio = FRAME * 3 + FRAME[:100]
+
+    out = harness.run(
+        *(BytesEvent(audio[i : i + 333]) for i in range(0, len(audio), 333))
+    )
+
+    assert detector.frames == [640, 640, 640]  # the last 100 bytes wait
+    assert [v.data.size for v in out] == [320, 320, 320]
+    harness.run(BytesEvent(FRAME[100:]))  # completes the 4th frame
+    assert detector.frames == [640] * 4
+
+
+def test_one_chunk_holding_speech_then_silence_ends_the_utterance():
+    out = vad([True] + [False] * 6).run(BytesEvent(FRAME * 7))
+    assert [v.data is None for v in out] == [False, True]
+
+
+def test_real_detector_takes_an_odd_sized_chunk():
+    out = ModuleHarness(VAD).run(BytesEvent(bytes(1001)), BytesEvent(bytes(279)))
+    assert out == [] or all(isinstance(v, Voice) for v in out)

huri_module_vad-0.2.0/tests/test_vad_detector.py

@@ -0,0 +1,65 @@
+"""VAD with the real WebRTC detector, on synthetic speech (speech_signals.py).
+
+test_vad.py scripts the detector's decisions; these tests let it decide, so
+they also guard that the synthetic signal still reads as speech to it. What
+the rest of the pipeline relies on: ``Voice(None)`` exactly once per
+utterance (EMO, and through EAG/QAG the whole turn, depend on it) and nothing
+at all before anyone spoke.
+"""
+
+from huri_events.speech import Voice
+from huri_vad.module import VAD
+from speech_signals import SAMPLE_RATE, frames, silence, speech_like
+
+from huri_core.core.events import BytesEvent
+from huri_core.testing import ModuleHarness
+
+BLOCK_DURATION = 0.030
+BLOCK = int(BLOCK_DURATION * SAMPLE_RATE)
+
+
+def _vad(**kwargs):
+    params = dict(
+        vad_agressiveness=3,
+        silence_duration=0.3,
+        sample_rate=SAMPLE_RATE,
+        block_duration=BLOCK_DURATION,
+    )
+    params.update(kwargs)
+    return ModuleHarness(VAD, **params)
+
+
+def _feed(vad, *audios):
+    events = [BytesEvent(f.tobytes()) for audio in audios for f in frames(audio, BLOCK)]
+    return vad.run(*events)
+
+
+def _markers(out):
+    return [v for v in out if v.data is None]
+
+
+def test_synthetic_speech_reads_as_speech():
+    out = _feed(_vad(), speech_like(1.0))
+    voiced = [v for v in out if v.data is not None]
+    assert len(voiced) > len(frames(speech_like(1.0), BLOCK)) * 0.8, (
+        "the synthetic signal must read as speech to webrtcvad, or every test "
+        "in this file passes vacuously"
+    )
+    assert all(isinstance(v, Voice) and v.data.dtype.name == "float32" for v in voiced)
+
+
+def test_end_of_turn_is_emitted_exactly_once():
+    out = _feed(_vad(), speech_like(0.6), silence(2.0))
+    assert len(_markers(out)) == 1
+
+
+def test_no_end_of_turn_before_any_speech():
+    assert _feed(_vad(), silence(3.0)) == []
+
+
+def test_two_turns_each_get_their_own_marker():
+    out = _feed(_vad(), speech_like(0.6), silence(2.0), speech_like(0.6), silence(2.0))
+    assert len(_markers(out)) == 2
+    # Speech, marker, speech, marker: the markers close their own turn.
+    kinds = ["end" if v.data is None else "voice" for v in out]
+    assert kinds.index("end") < len(kinds) - 1 - kinds[::-1].index("voice")

huri_module_vad-0.1.0.dev0/huri_module_vad.egg-info/requires.txt

@@ -1,5 +0,0 @@
-huri-core[server]<0.5,>=0.4.0.dev0
-huri-events<0.2,>=0.1.0.dev0
-numpy<2,>=1.26
-webrtcvad==2.0.10
-setuptools<82