lfeats 0.2.0__tar.gz → 0.2.2__tar.gz

{lfeats-0.2.0 → lfeats-0.2.2}/.gitignore

@@ -11,6 +11,7 @@ build/
 
 # tests
 tests/outputs/
+*.wav
 
 # tools
 tools/**/

{lfeats-0.2.0 → lfeats-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: lfeats
-Version: 0.2.0
+Version: 0.2.2
 Summary: A unified interface to extract hidden representations from speech foundation models
 Project-URL: Documentation, https://takenori-y.github.io/lfeats/stable/
 Project-URL: Source, https://github.com/takenori-y/lfeats
@@ -28,11 +28,13 @@ Requires-Dist: onnxruntime>=1.19.0
 Requires-Dist: parfive>=2.1.0
 Requires-Dist: platformdirs>=2.0.0
 Requires-Dist: requests>=2.27.0
+Requires-Dist: scipy>=1.9.2
 Requires-Dist: soundfile>=0.10.2
 Requires-Dist: soxr>=0.4.0
 Requires-Dist: torch>=2.6.0
 Requires-Dist: torchaudio>=2.6.0
-Requires-Dist: transformers>=4.30.0
+Requires-Dist: tqdm>=4.63.0
+Requires-Dist: transformers>=5.3.0
 Provides-Extra: dev
 Requires-Dist: build; extra == 'dev'
 Requires-Dist: matplotlib; extra == 'dev'
@@ -40,7 +42,7 @@ Requires-Dist: mdformat; extra == 'dev'
 Requires-Dist: numpydoc; extra == 'dev'
 Requires-Dist: pkginfo; extra == 'dev'
 Requires-Dist: pydata-sphinx-theme; extra == 'dev'
-Requires-Dist: pyright; extra == 'dev'
+Requires-Dist: pyright[nodejs]; extra == 'dev'
 Requires-Dist: pytest; extra == 'dev'
 Requires-Dist: pytest-cov; extra == 'dev'
 Requires-Dist: ruff; extra == 'dev'
@@ -143,16 +145,33 @@ pip install git+https://github.com/takenori-y/lfeats.git@master
 | | `large-v2` | 32 | 1280 | | | [🤗](https://huggingface.co/openai/whisper-large-v2) |
 | | `large-v3` | 32 | 1280 | | | [🤗](https://huggingface.co/openai/whisper-large-v3) |
 
-### Utterance-Level Features
+### Token-Level Features
 
-| Model Name | Model Variant | Layers | Dimension | Paper | Source | Model Hub |
+| Model Name | Model Variant | Hop Size [ms] | Dimension | Paper | Source | Model Hub |
 | :--- | :--- | ---: | ---: | :---: | :---: | :---: |
-| `ecapa-tdnn` | `base` | 0 | 192 | [arXiv](https://arxiv.org/abs/2005.07143) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb) |
-| `next-tdnn` | `light` | 0 | 192 | [arXiv](https://arxiv.org/abs/2312.08603) | [GitHub](https://github.com/dmlguq456/NeXt_TDNN_ASV) | |
-| | `base` | 0 | 192 | | | |
-| | `base-v2` | 0 | 192 | | | |
-| `r-vector` | `base` | 0 | 256 | [arXiv](https://arxiv.org/pdf/1910.12592) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-resnet-voxceleb) |
-| `x-vector` | `base` | 0 | 512 | [IEEE](https://ieeexplore.ieee.org/document/8461375) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-xvect-voxceleb) |
+| `higgs-audio` | `v2` | 40 | 1024 | [Blog](https://github.com/Render-AI-Code/cog-higgs-audio/blob/main/tech_blogs/TOKENIZER_BLOG.md) | [GitHub](https://github.com/boson-ai/higgs-audio) | [🤗](https://huggingface.co/eustlb/higgs-audio-v2-tokenizer) |
+| `x-codec` | `hubert` | 20 | 1024 | [arXiv](https://arxiv.org/abs/2408.17175) | [GitHub](https://github.com/zhenye234/xcodec) | [🤗](https://huggingface.co/hf-audio/xcodec-hubert-librispeech) |
+| | `wavlm` | 20 | 1024 | | | [🤗](https://huggingface.co/hf-audio/xcodec-wavlm-more-data) |
+
+### Utterance-Level Features
+
+| Model Name | Model Variant | Dimension | Paper | Source | Model Hub |
+| :--- | :--- | ---: | :---: | :---: | :---: |
+| `ecapa-tdnn` | `base` | 192 | [arXiv](https://arxiv.org/abs/2005.07143) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb) |
+| `next-tdnn` | `light` | 192 | [arXiv](https://arxiv.org/abs/2312.08603) | [GitHub](https://github.com/dmlguq456/NeXt_TDNN_ASV) | |
+| | `base` | 192 | | | |
+| | `base-v2` | 192 | | | |
+| `r-vector` | `base` | 256 | [arXiv](https://arxiv.org/abs/1910.12592) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-resnet-voxceleb) |
+| `redimnet` | `b0` | 192 | [arXiv](https://arxiv.org/abs/2407.18223) | [GitHub](https://github.com/IDRnD/redimnet) | |
+| | `b1` | 192 | | | |
+| | `b2` | 192 | | | |
+| | `b3` | 192 | | | |
+| | `b4` | 192 | | | |
+| | `b5` | 192 | | | |
+| | `b6` | 192 | | | |
+| `wavlm-sv` | `base` | 512 | [arXiv](https://arxiv.org/abs/2110.13900) | [GitHub](https://github.com/microsoft/unilm/tree/master/wavlm) | [🤗](https://huggingface.co/microsoft/wavlm-base-sv) |
+| | `base+` | 512 | | | [🤗](https://huggingface.co/microsoft/wavlm-base-plus-sv) |
+| `x-vector` | `base` | 512 | [IEEE](https://ieeexplore.ieee.org/document/8461375) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-xvect-voxceleb) |
 
 > [!IMPORTANT]
 > Users must comply with the respective licenses of the models.
@@ -163,12 +182,14 @@ pip install git+https://github.com/takenori-y/lfeats.git@master
 | Resampler Type | Quality Preset | Source | License |
 | :--- | :--- | :---: | :--- |
 | `lilfilter` | `base` | [GitHub](https://github.com/danpovey/filtering) | MIT |
+| `scipy` | `fft` | [Document](https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.resample.html) | BSD 3-Clause |
+| | `poly` | [Document](https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.resample_poly.html) | |
 | `soxr` | `quick` | [GitHub](https://github.com/dofuuz/python-soxr) | LGPL v2.1+ |
 | | `low` | | |
 | | `medium` | | |
 | | `high` | | |
 | | `very-high` | | |
-| `torchaudio` | `kaiser-fast` | [GitHub](https://github.com/pytorch/audio) | BSD 2-Clause |
+| `torchaudio` | `kaiser-fast` | [Document](https://docs.pytorch.org/audio/stable/tutorials/audio_resampling_tutorial.html) | BSD 2-Clause |
 | | `kaiser-best` | | |
 
 ## Examples

{lfeats-0.2.0 → lfeats-0.2.2}/README.md

@@ -93,16 +93,33 @@ pip install git+https://github.com/takenori-y/lfeats.git@master
 | | `large-v2` | 32 | 1280 | | | [🤗](https://huggingface.co/openai/whisper-large-v2) |
 | | `large-v3` | 32 | 1280 | | | [🤗](https://huggingface.co/openai/whisper-large-v3) |
 
-### Utterance-Level Features
+### Token-Level Features
 
-| Model Name | Model Variant | Layers | Dimension | Paper | Source | Model Hub |
+| Model Name | Model Variant | Hop Size [ms] | Dimension | Paper | Source | Model Hub |
 | :--- | :--- | ---: | ---: | :---: | :---: | :---: |
-| `ecapa-tdnn` | `base` | 0 | 192 | [arXiv](https://arxiv.org/abs/2005.07143) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb) |
-| `next-tdnn` | `light` | 0 | 192 | [arXiv](https://arxiv.org/abs/2312.08603) | [GitHub](https://github.com/dmlguq456/NeXt_TDNN_ASV) | |
-| | `base` | 0 | 192 | | | |
-| | `base-v2` | 0 | 192 | | | |
-| `r-vector` | `base` | 0 | 256 | [arXiv](https://arxiv.org/pdf/1910.12592) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-resnet-voxceleb) |
-| `x-vector` | `base` | 0 | 512 | [IEEE](https://ieeexplore.ieee.org/document/8461375) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-xvect-voxceleb) |
+| `higgs-audio` | `v2` | 40 | 1024 | [Blog](https://github.com/Render-AI-Code/cog-higgs-audio/blob/main/tech_blogs/TOKENIZER_BLOG.md) | [GitHub](https://github.com/boson-ai/higgs-audio) | [🤗](https://huggingface.co/eustlb/higgs-audio-v2-tokenizer) |
+| `x-codec` | `hubert` | 20 | 1024 | [arXiv](https://arxiv.org/abs/2408.17175) | [GitHub](https://github.com/zhenye234/xcodec) | [🤗](https://huggingface.co/hf-audio/xcodec-hubert-librispeech) |
+| | `wavlm` | 20 | 1024 | | | [🤗](https://huggingface.co/hf-audio/xcodec-wavlm-more-data) |
+
+### Utterance-Level Features
+
+| Model Name | Model Variant | Dimension | Paper | Source | Model Hub |
+| :--- | :--- | ---: | :---: | :---: | :---: |
+| `ecapa-tdnn` | `base` | 192 | [arXiv](https://arxiv.org/abs/2005.07143) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb) |
+| `next-tdnn` | `light` | 192 | [arXiv](https://arxiv.org/abs/2312.08603) | [GitHub](https://github.com/dmlguq456/NeXt_TDNN_ASV) | |
+| | `base` | 192 | | | |
+| | `base-v2` | 192 | | | |
+| `r-vector` | `base` | 256 | [arXiv](https://arxiv.org/abs/1910.12592) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-resnet-voxceleb) |
+| `redimnet` | `b0` | 192 | [arXiv](https://arxiv.org/abs/2407.18223) | [GitHub](https://github.com/IDRnD/redimnet) | |
+| | `b1` | 192 | | | |
+| | `b2` | 192 | | | |
+| | `b3` | 192 | | | |
+| | `b4` | 192 | | | |
+| | `b5` | 192 | | | |
+| | `b6` | 192 | | | |
+| `wavlm-sv` | `base` | 512 | [arXiv](https://arxiv.org/abs/2110.13900) | [GitHub](https://github.com/microsoft/unilm/tree/master/wavlm) | [🤗](https://huggingface.co/microsoft/wavlm-base-sv) |
+| | `base+` | 512 | | | [🤗](https://huggingface.co/microsoft/wavlm-base-plus-sv) |
+| `x-vector` | `base` | 512 | [IEEE](https://ieeexplore.ieee.org/document/8461375) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-xvect-voxceleb) |
 
 > [!IMPORTANT]
 > Users must comply with the respective licenses of the models.
@@ -113,12 +130,14 @@ pip install git+https://github.com/takenori-y/lfeats.git@master
 | Resampler Type | Quality Preset | Source | License |
 | :--- | :--- | :---: | :--- |
 | `lilfilter` | `base` | [GitHub](https://github.com/danpovey/filtering) | MIT |
+| `scipy` | `fft` | [Document](https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.resample.html) | BSD 3-Clause |
+| | `poly` | [Document](https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.resample_poly.html) | |
 | `soxr` | `quick` | [GitHub](https://github.com/dofuuz/python-soxr) | LGPL v2.1+ |
 | | `low` | | |
 | | `medium` | | |
 | | `high` | | |
 | | `very-high` | | |
-| `torchaudio` | `kaiser-fast` | [GitHub](https://github.com/pytorch/audio) | BSD 2-Clause |
+| `torchaudio` | `kaiser-fast` | [Document](https://docs.pytorch.org/audio/stable/tutorials/audio_resampling_tutorial.html) | BSD 2-Clause |
 | | `kaiser-best` | | |
 
 ## Examples

{lfeats-0.2.0 → lfeats-0.2.2}/lfeats/cli.py

@@ -12,6 +12,8 @@ import os
 import sys
 from pathlib import Path
 
+from tqdm import tqdm
+
 logger = logging.getLogger("lfeats")
 
 
@@ -125,6 +127,12 @@ def get_arguments() -> argparse.Namespace:
             "'mean', or 'auto'."
         ),
     )
+    parser.add_argument(
+        "--normalize",
+        action=argparse.BooleanOptionalAction,
+        default=False,
+        help="Enable or disable L2 normalization of the extracted features.",
+    )
     parser.add_argument(
         "-q",
         "--quiet",
@@ -147,9 +155,6 @@ def main() -> None:
         datefmt="%Y-%m-%d %H:%M:%S",
     )
 
-    import numpy as np
-    import torch
-
     import lfeats
     from lfeats.utils.io import load_audio
 
@@ -166,11 +171,13 @@ def main() -> None:
         for root, _, files in os.walk(args.source):
             for file in files:
                 input_files.append(os.path.join(root, file))
+        input_files = sorted(input_files)
     else:
         raise ValueError(f"Invalid source: {args.source}")
 
     if len(input_files) == 0:
-        raise ValueError(f"No audio files found in the source: {args.source}")
+        logging.info(f"No audio files found in the source: {args.source}")
+        sys.exit(0)
     logger.info(f"Found {len(input_files)} audio files to process.")
 
     # Parse the layers argument.
@@ -199,12 +206,15 @@ def main() -> None:
         device=args.device,
         cache_dir=args.cache_dir,
     )
+    logger.info(f"Loading {extractor.get_model_id()} model...")
     extractor.load(quiet=args.quiet)
 
     num_errors = 0
 
     # Process each input file and extract features.
-    for input_file in input_files:
+    for input_file in tqdm(
+        input_files, desc="Processing files", unit="file", disable=args.quiet
+    ):
         if not os.path.isfile(input_file):
             logger.error(f"Could not find file: {input_file}. Skipping.")
             num_errors += 1
@@ -216,7 +226,7 @@ def main() -> None:
             path = Path(input_file).parent
             # Remove the root part of the path.
             dirs = path.relative_to(path.anchor).parts
-            if args.subdir_offset >= len(dirs):
+            if args.subdir_offset > len(dirs):
                 logger.error(
                     f"Subdir offset {args.subdir_offset} is too large for file: "
                     f"{input_file}. Skipping."
@@ -246,32 +256,14 @@ def main() -> None:
                 overlap_length_sec=args.overlap_length_sec,
                 upsample_factor=args.upsample_factor,
                 reduction=args.reduction,
+                normalize=args.normalize,
             )
         except Exception as e:
             logger.error(f"Error processing file {input_file}: {e}. Skipping.")
             num_errors += 1
             continue
 
-        if args.output_format == "npz":
-            result = {
-                "features": features.array,
-                "source": features.source,
-                "layers": features.layers,
-            }
-            np.savez_compressed(output_file, **result)
-        elif args.output_format == "pt":
-            result = {
-                "features": features.tensor.cpu(),
-                "source": features.source,
-                "layers": features.layers,
-            }
-            torch.save(result, output_file)
-        elif args.output_format == "float":
-            features.array.tofile(output_file)
-        elif args.output_format == "double":
-            features.array.astype(np.float64).tofile(output_file)
-        else:
-            raise ValueError(f"Unsupported output format: {args.output_format}")
+        features.tofile(output_file, double=args.output_format == "double")
 
     if num_errors > 0:
         logger.error(f"{num_errors} files were skipped due to errors.")

{lfeats-0.2.0 → lfeats-0.2.2}/lfeats/interfaces/extractor.py

@@ -102,6 +102,17 @@ class Extractor:
         self.model_manager.to(device)
         self.resampler_manager.to(device)
 
+    def get_model_id(self) -> str:
+        """Return the identifier of the model being used.
+
+        Returns
+        -------
+        out : str
+            The identifier of the model being used.
+
+        """
+        return self.model_manager.get_model().model_id
+
     def __call__(
         self,
         source: np.ndarray | torch.Tensor | Audio,
@@ -113,6 +124,7 @@ class Extractor:
         overlap_length_sec: int = 5,
         upsample_factor: int = 1,
         reduction: Literal["none", "mean", "auto"] = "auto",
+        normalize: bool = False,
     ) -> Features:
         """Extract features from the input waveform.
 
@@ -148,6 +160,9 @@ class Extractor:
             determined based on the feature granularity ('none' for frame-level features
             and 'mean' for utterance-level features).
 
+        normalize : bool, optional
+            If True, the extracted features will be L2-normalized.
+
         Returns
         -------
         out : Features
@@ -188,6 +203,7 @@ class Extractor:
                 chunk_length_sec=chunk_length_sec,
                 overlap_length_sec=overlap_length_sec,
                 reduction=reduction,
+                normalize=normalize,
             )
 
         model = self.model_manager.get_model()
@@ -243,6 +259,7 @@ class Extractor:
         chunk_length_sec: int = 30,
         overlap_length_sec: int = 5,
         reduction: Literal["none", "mean", "auto"] = "auto",
+        normalize: bool = False,
     ) -> Features:
         """Extract features from the input waveform.
 
@@ -275,6 +292,9 @@ class Extractor:
             determined based on the feature granularity ('none' for frame-level features
             and 'mean' for utterance-level features).
 
+        normalize : bool, optional
+            If True, the extracted features will be L2-normalized.
+
         Returns
         -------
         out : Features
@@ -377,6 +397,10 @@ class Extractor:
         ):
             features = features.reduce("mean")
 
+        # Apply normalization if specified.
+        if normalize:
+            features = features.normalize()
+
         return features
 
     @staticmethod

{lfeats-0.2.0 → lfeats-0.2.2}/lfeats/interfaces/types.py

@@ -9,6 +9,7 @@ from dataclasses import dataclass
 from enum import Enum
 
 import numpy as np
+import soundfile as sf
 import torch
 import torch.nn.functional as F
 
@@ -164,6 +165,28 @@ class Audio(Container):
         """
         return self.data.shape[1]
 
+    def tofile(self, path: str) -> None:
+        """Save the audio to a file.
+
+        Parameters
+        ----------
+        path : str
+            The path to save the audio to.
+
+        """
+        ext = path.split(".")[-1].lower()
+        if ext in ("wav", "flac"):
+            sf.write(path, self.array.T, self.sample_rate)
+        elif ext == "npz":
+            np.savez_compressed(path, samples=self.array, sample_rate=self.sample_rate)
+        elif ext == "pt":
+            torch.save(
+                {"samples": self.tensor.cpu(), "sample_rate": self.sample_rate},
+                path,
+            )
+        else:
+            self.array.tofile(path)
+
     def normalize(self, eps: float = 1e-5) -> Audio:
         """Normalize the audio samples to have zero mean and unit variance.
 
@@ -237,6 +260,40 @@ class Features(Container):
         """
         return self.data.shape[1]
 
+    def tofile(self, path: str, double: bool = False) -> None:
+        """Save the features to a file.
+
+        Parameters
+        ----------
+        path : str
+            The path to save the features to.
+
+        double : bool, optional
+            Whether to save the features in double precision instead of single one.
+
+        """
+        ext = path.split(".")[-1].lower()
+        if ext == "npz":
+            np.savez_compressed(
+                path,
+                features=self.array.astype(np.float64 if double else np.float32),
+                source=self.source,
+                layers=self.layers or [],
+            )
+        elif ext == "pt":
+            torch.save(
+                {
+                    "features": self.tensor.cpu().to(
+                        torch.float64 if double else torch.float32
+                    ),
+                    "source": self.source,
+                    "layers": self.layers,
+                },
+                path,
+            )
+        else:
+            self.array.astype(np.float64 if double else np.float32).tofile(path)
+
     def trim(self, start: int, end: int) -> Features:
         """Trim the features along the time dimension.
 
@@ -299,6 +356,26 @@ class Features(Container):
 
         return Features(data=reduced_data, source=self.source, layers=self.layers)
 
+    def normalize(self) -> Features:
+        """Normalize the features in the L2 sense along the feature dimension.
+
+        Returns
+        -------
+        out : Features
+            A new Features instance with normalized data.
+
+        """
+        eps = 1e-12
+        if isinstance(self.data, np.ndarray):
+            norm = np.linalg.norm(self.array, ord=2, axis=-1, keepdims=True)
+            normalized_data = self.array / np.maximum(norm, eps)
+        else:
+            normalized_data = torch.nn.functional.normalize(
+                self.tensor, p=2, dim=-1, eps=eps
+            )
+
+        return Features(data=normalized_data, source=self.source, layers=self.layers)
+
     def concat(self, other: Features) -> Features:
         """Concatenate this Features instance with another one along the time dimension.
 

{lfeats-0.2.0 → lfeats-0.2.2}/lfeats/models/__init__.py

@@ -9,18 +9,22 @@ from .data2vec2 import Data2Vec2Model
 from .ecapa_tdnn import EcapaTDNNModel
 from .emotion2vec import Emotion2VecModel
 from .emotion2vec_plus import Emotion2VecPlusModel
+from .higgs_audio import HiggsAudioTokenizerModel
 from .hubert import HuBERTModel
 from .manager import ModelManager
 from .next_tdnn import NeXtTDNNModel
 from .r_spin import RSpinModel
 from .r_vector import RVectorModel
+from .redimnet import ReDimNetModel
 from .spidr import SpidRModel
 from .spin import SpinModel
 from .sslzip import SSLZipModel
 from .unispeech_sat import UniSpeechSATModel
 from .wav2vec2 import Wav2Vec2Model
 from .wavlm import WavLMModel
+from .wavlm_sv import WavLMSVModel
 from .whisper import WhisperModel
+from .x_codec import XCodecModel
 from .x_vector import XVectorModel
 
 MODEL_MAP = {
@@ -30,21 +34,25 @@ MODEL_MAP = {
     "ecapa-tdnn": EcapaTDNNModel,
     "emotion2vec": Emotion2VecModel,
     "emotion2vec+": Emotion2VecPlusModel,
+    "higgs-audio": HiggsAudioTokenizerModel,
     "hubert": HuBERTModel,
     "next-tdnn": NeXtTDNNModel,
     "r-spin": RSpinModel,
     "r-vector": RVectorModel,
+    "redimnet": ReDimNetModel,
     "spidr": SpidRModel,
     "spin": SpinModel,
     "sslzip": SSLZipModel,
     "unispeech-sat": UniSpeechSATModel,
     "wav2vec2": Wav2Vec2Model,
     "wavlm": WavLMModel,
+    "wavlm-sv": WavLMSVModel,
     "whisper": WhisperModel,
+    "x-codec": XCodecModel,
     "x-vector": XVectorModel,
 }
 
 __all__ = [
-    "ModelManager",
     "MODEL_MAP",
+    "ModelManager",
 ]

{lfeats-0.2.0 → lfeats-0.2.2}/lfeats/models/base.py

@@ -256,6 +256,58 @@ class FrameLevelFeatureModel(BaseModel):
         return Granularity.FRAME
 
 
+class TokenLevelFeatureModel(BaseModel):
+    """An abstract base class for frame-level feature extraction models."""
+
+    @property
+    def num_layers(self) -> int:
+        """Get the number of layers in the model.
+
+        Returns
+        -------
+        out : int
+            The number of layers.
+
+        """
+        return 0
+
+    @property
+    def frame_shift(self) -> int:
+        """Get the frame shift of the model.
+
+        Returns
+        -------
+        out : int
+            The frame shift in samples.
+
+        """
+        return int(20.0 * self.sample_rate / 1000)
+
+    @property
+    def center_offset(self) -> int:
+        """Get the center offset of the model.
+
+        Returns
+        -------
+        out : int
+            The center offset in samples.
+
+        """
+        return 0
+
+    @property
+    def granularity(self) -> Granularity:
+        """Get the granularity of the features extracted by the model.
+
+        Returns
+        -------
+        out : str
+            The granularity of the features.
+
+        """
+        return Granularity.FRAME
+
+
 class UtteranceLevelFeatureModel(BaseModel):
     """An abstract base class for utterance-level feature extraction models."""
 

{lfeats-0.2.0 → lfeats-0.2.2}/lfeats/models/ecapa_tdnn.py

@@ -10,7 +10,7 @@ import torch
 from ..interfaces.types import Audio, Features
 from ..utils.io import silence_hf_hub
 from ..utils.paths import setup_third_party_path
-from ..utils.validation import validate_enum
+from ..utils.validation import validate_enum, validate_length
 from .base import UtteranceLevelFeatureModel
 
 
@@ -108,6 +108,8 @@ class EcapaTDNNModel(UtteranceLevelFeatureModel):
             raise RuntimeError("Model is not loaded. Call 'load' method first.")
 
         with torch.inference_mode():
-            vectors = self.model.encode_batch(audio.tensor.to(self.device))  # (B, N, D)
+            inputs = audio.tensor.to(self.device)
+            inputs = validate_length(inputs, 640)
+            vectors = self.model.encode_batch(inputs)
 
         return Features(data=vectors, source=self.model_id)