fix: run longform VAD on preloaded waveform

.github/workflows/gigaam.yml

@@ -9,6 +9,8 @@ on:
 
 env:
   FORCE_COLOR: 1
+  # Repository secrets are unavailable to fork pull requests.
+  RUN_HF_TOKEN_TESTS: ${{ github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository }}
 
 jobs:
   test:
@@ -72,8 +74,13 @@ jobs:
     - name: Run batching tests
       run: |
         pytest -v tests/test_batching.py --tb=short
+
+    - name: Run VAD utility tests
+      run: |
+        pytest -v tests/test_vad_utils.py --tb=short
 
     - name: Run longform tests
+      if: env.RUN_HF_TOKEN_TESTS == 'true'
       env:
         HF_TOKEN: ${{ secrets.HF_TOKEN }}
       run: |
@@ -85,9 +92,19 @@ jobs:
 
     - name: Run timestamps tests
       run: |
-        pytest -v tests/test_timestamps.py --tb=short
+        pytest -v tests/test_timestamps.py -k "not transcribe_longform" --tb=short
+
+    - name: Run longform timestamps tests
+      if: env.RUN_HF_TOKEN_TESTS == 'true'
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      run: |
+        pytest -v tests/test_timestamps.py -k "transcribe_longform" --tb=short
 
     - name: Run training tests
+      if: env.RUN_HF_TOKEN_TESTS == 'true'
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
       run: |
         pytest -v tests/test_training.py --tb=short
 

gigaam/vad_utils.py

@@ -78,6 +78,19 @@ def get_pipeline(
     return _PIPELINE.to(device)
 
 
+def _apply_pipeline(pipeline: Pipeline, audio_file):
+    """
+    Applies PyAnnote pipeline to a preloaded waveform dict.
+
+    GigaAM already decodes the audio for ASR chunk slicing. Reusing the same
+    waveform for VAD avoids depending on pyannote/torchcodec path decoding and
+    keeps VAD boundaries aligned with the samples later sliced for ASR.
+    """
+    if hasattr(pipeline, "prepare_one") and hasattr(pipeline, "apply"):
+        return pipeline.apply(pipeline.prepare_one(audio_file, preload=False))
+    return pipeline(audio_file)
+
+
 def segment_audio_file(
     wav_file: str,
     sr: int,
@@ -92,9 +105,10 @@ def segment_audio_file(
     The segmentation is performed using a PyAnnote voice activity detection pipeline.
     """
 
-    audio = load_audio(wav_file)
+    audio = load_audio(wav_file, sample_rate=sr)
     pipeline = get_pipeline(device)
-    sad_segments = cast(Annotation, pipeline(wav_file))
+    sad_input = {"waveform": audio.unsqueeze(0), "sample_rate": sr}
+    sad_segments = cast(Annotation, _apply_pipeline(pipeline, sad_input))
 
     segments: List[torch.Tensor] = []
     curr_duration = 0.0

tests/test_vad_utils.py

@@ -0,0 +1,62 @@
+import torch
+
+
+class _FakeSegment:
+    def __init__(self, start, end):
+        self.start = start
+        self.end = end
+
+
+class _FakeTimeline:
+    def support(self):
+        return [_FakeSegment(1.0, 3.0), _FakeSegment(4.0, 6.0)]
+
+
+class _FakeSadSegments:
+    def get_timeline(self):
+        return _FakeTimeline()
+
+
+class _FakePyannote4Pipeline:
+    def __init__(self):
+        self.used_apply = False
+
+    def prepare_one(self, audio_file, preload=False):
+        assert set(audio_file) == {"waveform", "sample_rate"}
+        assert audio_file["waveform"].shape == (1, 80)
+        assert audio_file["sample_rate"] == 10
+        assert preload is False
+        return audio_file
+
+    def apply(self, prepared):
+        assert prepared["sample_rate"] == 10
+        self.used_apply = True
+        return _FakeSadSegments()
+
+    def __call__(self, audio_file):
+        raise AssertionError("segment_audio_file should use preloaded waveform input")
+
+
+def test_segmentation_uses_preloaded_waveform_for_pyannote_pipeline(monkeypatch):
+    """VAD should use the same decoded waveform later sliced for ASR chunks."""
+    from gigaam import vad_utils
+
+    pipeline = _FakePyannote4Pipeline()
+    monkeypatch.setattr(vad_utils, "get_pipeline", lambda device: pipeline)
+    monkeypatch.setattr(
+        vad_utils,
+        "load_audio",
+        lambda wav_file, sample_rate: torch.arange(80, dtype=torch.float32),
+    )
+
+    segments, boundaries = vad_utils.segment_audio_file(
+        "fake.wav",
+        sr=10,
+        max_duration=10.0,
+        min_duration=8.0,
+        strict_limit_duration=30.0,
+    )
+
+    assert pipeline.used_apply
+    assert boundaries == [(1.0, 6.0)]
+    assert torch.equal(segments[0], torch.arange(10, 60, dtype=torch.float32))