Address review comments: SPARC, SQUIM, PPG, SpeechBrain (#470)

* Address review comments: SPARC tempdir, SQUIM loop, public PPG API, SB revision

- SPARC decode/convert: create Audio from waveform tensor directly,
  no dangling filepath to deleted tempdir
- SQUIM: move model.to(device) before loop (was inside loop)
- PPG: rename _extract_ppg_segments → extract_ppg_segments (public),
  _to_frame_major_posteriorgram → to_frame_major_posteriorgram (public),
  export from __init__.py, update tutorials and tests
- SpeechBrain SER: add revision to cache key, pass revision to
  hf_hub_download, use yaml.safe_load instead of string splitting

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Fix SpeechBrain YAML parsing: use permissive loader for custom tags

SpeechBrain hyperparams.yaml uses custom YAML tags (!new:, !ref, etc.)
that yaml.safe_load cannot handle. Use a dedicated SafeLoader subclass
with a multi-constructor that ignores unknown tags.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Fix test function names: add missing underscore after 'test' prefix

7 test functions were named testXxx instead of test_Xxx, making
them invisible to pytest discovery. Now all 14 PPG/segment tests
are properly discovered and pass.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: satra

src/senselab/audio/tasks/classification/speech_emotion_recognition/api.py

@@ -271,7 +271,7 @@ def _classify_speechbrain_ser(
         user_preference=device, compatible_devices=[DeviceType.CUDA, DeviceType.CPU]
     )
 
-    key = f"{model.path_or_uri}-{device_type.value}"
+    key = f"{model.path_or_uri}-{model.revision or 'main'}-{device_type.value}"
     if key not in _speechbrain_ser_models:
         # Discover required modules from the model's hyperparams
         _speechbrain_ser_models[key] = _load_speechbrain_ser_model(model, device_type)
@@ -317,24 +317,23 @@ def _load_speechbrain_ser_model(model: SpeechBrainModel, device_type: DeviceType
     Returns:
         Tuple of (recognizer, label_list).
     """
+    import yaml  # type: ignore[import-untyped]
     from huggingface_hub import hf_hub_download
     from speechbrain.inference import Pretrained
 
     # Download hyperparams to discover MODULES_NEEDED
-    hp_path = hf_hub_download(str(model.path_or_uri), "hyperparams.yaml")
+    hp_path = hf_hub_download(str(model.path_or_uri), "hyperparams.yaml", revision=model.revision)
+
+    # SpeechBrain YAML uses custom tags (!new:, !ref, etc.) that yaml.safe_load
+    # cannot handle. Use a permissive loader subclass that ignores unknown tags.
+    class _PermissiveLoader(yaml.SafeLoader):
+        pass
+
+    _PermissiveLoader.add_multi_constructor("", lambda loader, suffix, node: None)
     with open(hp_path) as f:
-        content = f.read()
-
-    # Extract MODULES_NEEDED from the yaml
-    modules_needed = None
-    for line in content.split("\n"):
-        if "MODULES_NEEDED" in line:
-            # Parse: MODULES_NEEDED: ["wav2vec2", "avg_pool", "output_mlp"]
-            import ast
-
-            _, _, value = line.partition(":")
-            modules_needed = ast.literal_eval(value.strip())
-            break
+        hparams = yaml.load(f, Loader=_PermissiveLoader)  # noqa: S506
+
+    modules_needed = hparams.get("MODULES_NEEDED") if hparams else None
 
     if not modules_needed:
         raise ValueError(

src/senselab/audio/tasks/features_extraction/__init__.py

@@ -1,5 +1,11 @@
 """.. include:: ./doc.md"""  # noqa: D415
 
 from .api import extract_features_from_audios  # noqa: F401
-from .ppg import extract_mean_phoneme_durations, plot_ppg_phoneme_timeline  # noqa: F401
+from .ppg import (  # noqa: F401
+    extract_mean_phoneme_durations,
+    extract_ppg_segments,
+    extract_ppgs_from_audios,
+    plot_ppg_phoneme_timeline,
+    to_frame_major_posteriorgram,
+)
 from .sparc import SparcFeatureExtractor  # noqa: F401

src/senselab/audio/tasks/features_extraction/ppg.py

@@ -201,7 +201,7 @@ def extract_ppgs_from_audios(audios: List[Audio], device: Optional[DeviceType] =
 # ---------------------------------------------------------------------------
 
 
-def _to_frame_major_posteriorgram(posteriorgram: torch.Tensor) -> torch.Tensor:
+def to_frame_major_posteriorgram(posteriorgram: torch.Tensor) -> torch.Tensor:
     """Normalize a PPG tensor to frame-major layout ``(frames, phonemes)``.
 
     The ppgs library typically returns tensors shaped ``(1, phonemes, frames)``
@@ -241,7 +241,7 @@ def _to_frame_major_posteriorgram(posteriorgram: torch.Tensor) -> torch.Tensor:
     return t
 
 
-def _extract_ppg_segments(
+def extract_ppg_segments(
     audio: Audio,
     frame_major_posteriorgram: torch.Tensor,
 ) -> List[Dict[str, Any]]:
@@ -358,8 +358,8 @@ def extract_mean_phoneme_durations(
         logger.warning("Posteriorgram is empty or contains NaN — returning empty result.")
         return {}
 
-    frame_major = _to_frame_major_posteriorgram(posteriorgram)
-    segments = _extract_ppg_segments(audio, frame_major)
+    frame_major = to_frame_major_posteriorgram(posteriorgram)
+    segments = extract_ppg_segments(audio, frame_major)
 
     if not segments:
         return {}
@@ -432,8 +432,8 @@ def plot_ppg_phoneme_timeline(
     if posteriorgram.numel() == 0 or torch.isnan(posteriorgram).any():
         raise ValueError("Cannot plot an empty or NaN posteriorgram.")
 
-    frame_major = _to_frame_major_posteriorgram(posteriorgram)
-    segments = _extract_ppg_segments(audio, frame_major)
+    frame_major = to_frame_major_posteriorgram(posteriorgram)
+    segments = extract_ppg_segments(audio, frame_major)
 
     if not segments:
         raise ValueError("No segments found in posteriorgram.")

src/senselab/audio/tasks/features_extraction/sparc.py

@@ -370,10 +370,11 @@ def decode_sparc_features(
             sample_rate = output["sample_rate"]
             logger.info("SPARC decode produced audio at %d Hz: %s", sample_rate, audio_path)
 
-            audio = Audio(filepath=audio_path)
-            # Force lazy-load while the temp file still exists
-            _ = audio.waveform
-            return audio
+            tmp_audio = Audio(filepath=audio_path)
+            # Force lazy-load while the temp file still exists, then
+            # create a clean Audio without a filepath pointing at the
+            # (about-to-be-deleted) temp directory.
+            return Audio(waveform=tmp_audio.waveform, sampling_rate=tmp_audio.sampling_rate)
 
     @classmethod
     def convert_voice(
@@ -448,7 +449,8 @@ def convert_voice(
             sample_rate = output["sample_rate"]
             logger.info("SPARC convert produced audio at %d Hz: %s", sample_rate, audio_path)
 
-            audio = Audio(filepath=audio_path)
-            # Force lazy-load while the temp file still exists
-            _ = audio.waveform
-            return audio
+            tmp_audio = Audio(filepath=audio_path)
+            # Force lazy-load while the temp file still exists, then
+            # create a clean Audio without a filepath pointing at the
+            # (about-to-be-deleted) temp directory.
+            return Audio(waveform=tmp_audio.waveform, sampling_rate=tmp_audio.sampling_rate)

src/senselab/audio/tasks/features_extraction/torchaudio_squim.py

@@ -66,11 +66,12 @@ def extract_objective_quality_features_from_audios(
         raise ValueError("Only 16000 Hz sampling rate is supported by Torchaudio-Squim model.")
 
     features: List[Dict[str, Any]] = []
+    model = _get_objective_model().to(device.value)
 
     for audio in audios:
         audio_features = {}
         try:
-            stoi, pesq, si_sdr = _get_objective_model().to(device.value)(audio.waveform.to(device.value))
+            stoi, pesq, si_sdr = model(audio.waveform.to(device.value))
             audio_features["stoi"] = stoi.cpu().item()
             audio_features["pesq"] = pesq.cpu().item()
             audio_features["si_sdr"] = si_sdr.cpu().item()

src/tests/audio/tasks/features_extraction_test.py

@@ -9,11 +9,11 @@
 from senselab.audio.tasks.features_extraction import extract_features_from_audios
 from senselab.audio.tasks.features_extraction.opensmile import extract_opensmile_features_from_audios
 from senselab.audio.tasks.features_extraction.ppg import (
-    _extract_ppg_segments,
-    _to_frame_major_posteriorgram,
     extract_mean_phoneme_durations,
+    extract_ppg_segments,
     extract_ppgs_from_audios,
     plot_ppg_phoneme_timeline,
+    to_frame_major_posteriorgram,
 )
 from senselab.audio.tasks.features_extraction.praat_parselmouth import (
     extract_audio_duration,
@@ -419,31 +419,31 @@ def _make_synthetic_ppg(
 
 
 def test_to_frame_major_posteriorgram_3d() -> None:
-    """Test _to_frame_major_posteriorgram with a (1, phonemes, frames) input."""
+    """Test to_frame_major_posteriorgram with a (1, phonemes, frames) input."""
     ppg = torch.rand(1, 40, 100)
-    result = _to_frame_major_posteriorgram(ppg)
+    result = to_frame_major_posteriorgram(ppg)
     assert result.shape == (100, 40)
 
 
 def test_to_frame_major_posteriorgram_2d() -> None:
-    """Test _to_frame_major_posteriorgram with a (phonemes, frames) input."""
+    """Test to_frame_major_posteriorgram with a (phonemes, frames) input."""
     ppg = torch.rand(40, 100)
-    result = _to_frame_major_posteriorgram(ppg)
+    result = to_frame_major_posteriorgram(ppg)
     assert result.shape == (100, 40)
 
 
 def test_to_frame_major_posteriorgram_already_frame_major() -> None:
-    """Test _to_frame_major_posteriorgram when input is already (frames, phonemes)."""
+    """Test to_frame_major_posteriorgram when input is already (frames, phonemes)."""
     ppg = torch.rand(100, 40)
-    result = _to_frame_major_posteriorgram(ppg)
+    result = to_frame_major_posteriorgram(ppg)
     assert result.shape == (100, 40)
 
 
 def test_to_frame_major_posteriorgram_too_few_dims() -> None:
     """Test that a 1-D tensor raises ValueError."""
     ppg = torch.rand(40)
     with pytest.raises(ValueError, match="Expected at least a 2-D"):
-        _to_frame_major_posteriorgram(ppg)
+        to_frame_major_posteriorgram(ppg)
 
 
 def test_extract_ppg_segments_basic() -> None:
@@ -452,8 +452,8 @@ def test_extract_ppg_segments_basic() -> None:
     pattern = [0] * 10 + [1] * 5
     ppg = _make_synthetic_ppg(pattern)
     audio = Audio(waveform=torch.rand(1, 16000), sampling_rate=16000)
-    frame_major = _to_frame_major_posteriorgram(ppg)
-    segments = _extract_ppg_segments(audio, frame_major)
+    frame_major = to_frame_major_posteriorgram(ppg)
+    segments = extract_ppg_segments(audio, frame_major)
 
     assert len(segments) == 2
     assert segments[0]["phoneme_index"] == 0
@@ -473,8 +473,8 @@ def test_extract_ppg_segments_single_phoneme() -> None:
     pattern = [5] * 20
     ppg = _make_synthetic_ppg(pattern)
     audio = Audio(waveform=torch.rand(1, 32000), sampling_rate=16000)
-    frame_major = _to_frame_major_posteriorgram(ppg)
-    segments = _extract_ppg_segments(audio, frame_major)
+    frame_major = to_frame_major_posteriorgram(ppg)
+    segments = extract_ppg_segments(audio, frame_major)
 
     assert len(segments) == 1
     assert segments[0]["phoneme_index"] == 5
@@ -485,7 +485,7 @@ def test_extract_ppg_segments_empty() -> None:
     """Test segment extraction with zero frames."""
     ppg = torch.rand(40, 0)
     audio = Audio(waveform=torch.rand(1, 16000), sampling_rate=16000)
-    segments = _extract_ppg_segments(audio, ppg)
+    segments = extract_ppg_segments(audio, ppg)
     assert segments == []