lfeats 0.1.4__tar.gz → 0.2.1__tar.gz

{lfeats-0.1.4 → lfeats-0.2.1}/.gitignore

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

{lfeats-0.1.4 → lfeats-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lfeats
-Version: 0.1.4
+Version: 0.2.1
 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
@@ -18,6 +18,7 @@ Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
 Classifier: Programming Language :: Python :: 3.14
 Requires-Python: >=3.10
+Requires-Dist: filelock>=3.10.0
 Requires-Dist: huggingface-hub>=0.23.0
 Requires-Dist: hydra-core>=1.3.0
 Requires-Dist: hyperpyyaml>=0.0.1
@@ -27,11 +28,12 @@ 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: transformers>=5.3.0
 Provides-Extra: dev
 Requires-Dist: build; extra == 'dev'
 Requires-Dist: matplotlib; extra == 'dev'
@@ -39,7 +41,7 @@ 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<=1.1.408; extra == 'dev'
 Requires-Dist: pytest; extra == 'dev'
 Requires-Dist: pytest-cov; extra == 'dev'
 Requires-Dist: ruff; extra == 'dev'
@@ -142,6 +144,12 @@ 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) |
 
+### Token-Level Features
+
+| Model Name | Model Variant | Hop Size [ms] | Dimension | Paper | Source | Model Hub |
+| :--- | :--- | ---: | ---: | :---: | :---: | :---: |
+| `higgs-audio` | `v2` | 40 | 1024 | | [GitHub](https://github.com/boson-ai/higgs-audio) | [🤗](https://huggingface.co/eustlb/higgs-audio-v2-tokenizer) |
+
 ### Utterance-Level Features
 
 | Model Name | Model Variant | Layers | Dimension | Paper | Source | Model Hub |
@@ -150,7 +158,14 @@ pip install git+https://github.com/takenori-y/lfeats.git@master
 | `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) |
+| `r-vector` | `base` | 0 | 256 | [arXiv](https://arxiv.org/abs/1910.12592) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-resnet-voxceleb) |
+| `redimnet` | `b0` | 0 | 192 | [arXiv](https://arxiv.org/abs/2407.18223) | [GitHub](https://github.com/IDRnD/redimnet) | |
+| | `b1` | 0 | 192 | | | |
+| | `b2` | 0 | 192 | | | |
+| | `b3` | 0 | 192 | | | |
+| | `b4` | 0 | 192 | | | |
+| | `b5` | 0 | 192 | | | |
+| | `b6` | 0 | 192 | | | |
 | `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) |
 
 > [!IMPORTANT]
@@ -162,12 +177,14 @@ 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-0.1.4 → lfeats-0.2.1}/README.md

@@ -93,6 +93,12 @@ 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) |
 
+### Token-Level Features
+
+| Model Name | Model Variant | Hop Size [ms] | Dimension | Paper | Source | Model Hub |
+| :--- | :--- | ---: | ---: | :---: | :---: | :---: |
+| `higgs-audio` | `v2` | 40 | 1024 | | [GitHub](https://github.com/boson-ai/higgs-audio) | [🤗](https://huggingface.co/eustlb/higgs-audio-v2-tokenizer) |
+
 ### Utterance-Level Features
 
 | Model Name | Model Variant | Layers | Dimension | Paper | Source | Model Hub |
@@ -101,7 +107,14 @@ pip install git+https://github.com/takenori-y/lfeats.git@master
 | `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) |
+| `r-vector` | `base` | 0 | 256 | [arXiv](https://arxiv.org/abs/1910.12592) | [GitHub](https://github.com/speechbrain/speechbrain) | [🤗](https://huggingface.co/speechbrain/spkrec-resnet-voxceleb) |
+| `redimnet` | `b0` | 0 | 192 | [arXiv](https://arxiv.org/abs/2407.18223) | [GitHub](https://github.com/IDRnD/redimnet) | |
+| | `b1` | 0 | 192 | | | |
+| | `b2` | 0 | 192 | | | |
+| | `b3` | 0 | 192 | | | |
+| | `b4` | 0 | 192 | | | |
+| | `b5` | 0 | 192 | | | |
+| | `b6` | 0 | 192 | | | |
 | `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) |
 
 > [!IMPORTANT]
@@ -113,12 +126,14 @@ 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-0.1.4 → lfeats-0.2.1}/lfeats/cli.py

@@ -147,9 +147,6 @@ 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
 
@@ -170,7 +167,8 @@ def main() -> None:
         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.
@@ -216,7 +214,7 @@ 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."
@@ -252,26 +250,7 @@ def main() -> None:
             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-0.1.4 → lfeats-0.2.1}/lfeats/interfaces/extractor.py

@@ -90,6 +90,18 @@ class Extractor:
         """
         self.model_manager.get_model().load(self.cache_dir, quiet)
 
+    def to(self, device: str) -> None:
+        """Move the model to the specified device.
+
+        Parameters
+        ----------
+        device : str
+            The device to move the model to (e.g., 'cpu' or 'cuda').
+
+        """
+        self.model_manager.to(device)
+        self.resampler_manager.to(device)
+
     def __call__(
         self,
         source: np.ndarray | torch.Tensor | Audio,

{lfeats-0.1.4 → lfeats-0.2.1}/lfeats/interfaces/resampler.py

@@ -45,6 +45,17 @@ class Resampler:
                 f"Supported resamplers are: {[k for k in RESAMPLER_MAP.keys()]}"
             ) from e
 
+    def to(self, device: str) -> None:
+        """Move the resampler to the specified device.
+
+        Parameters
+        ----------
+        device : str
+            The device to move the resampler to (e.g., 'cpu' or 'cuda').
+
+        """
+        self.resampler_manager.to(device)
+
     def __call__(
         self,
         source: np.ndarray | torch.Tensor | Audio,

{lfeats-0.1.4 → lfeats-0.2.1}/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.
 

{lfeats-0.1.4 → lfeats-0.2.1}/lfeats/models/__init__.py

@@ -9,11 +9,13 @@ 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
@@ -30,10 +32,12 @@ 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,

{lfeats-0.1.4 → lfeats-0.2.1}/lfeats/models/base.py

@@ -25,6 +25,7 @@ class BaseModel(ABC):
         """
         self.device = device
 
+        self.model = None
         self._model_id = None  # To be defined in subclasses
 
     @abstractmethod
@@ -42,6 +43,24 @@ class BaseModel(ABC):
         """
         raise NotImplementedError
 
+    def to(self, device: str) -> None:
+        """Move the model to the specified device.
+
+        Parameters
+        ----------
+        device : str
+            The device to move the model to (e.g., 'cpu' or 'cuda').
+
+        """
+        self.device = device
+        if self.model is not None and hasattr(self.model, "to"):
+            self.model.to(device)
+            if hasattr(self.model, "device"):
+                try:
+                    self.model.device = device
+                except AttributeError:
+                    pass
+
     def extract_features(self, audio: Audio, layers: list[int]) -> Features:
         """Extract features from the input audio data.
 
@@ -237,6 +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(40.0 * 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-0.1.4 → lfeats-0.2.1}/lfeats/models/contentvec.py

@@ -68,8 +68,6 @@ class ContentVecModel(FrameLevelFeatureModel):
         )
         self._model_id = f"contentvec-{self.variant.value}"
 
-        self.model = None
-
     def load(self, model_dir: str, quiet: bool = False) -> None:
         """Load the model from the specified directory.
 

{lfeats-0.1.4 → lfeats-0.2.1}/lfeats/models/data2vec.py

@@ -58,9 +58,6 @@ class Data2VecModel(FrameLevelFeatureModel):
         self.variant = validate_enum(variant, Data2VecVariant, Data2VecVariant.BASE)
         self._model_id = f"data2vec-{self.variant.value}"
 
-        self.processor = None
-        self.model = None
-
     def load(self, model_dir: str, quiet: bool = False) -> None:
         """Load the model from the specified directory.
 

{lfeats-0.1.4 → lfeats-0.2.1}/lfeats/models/data2vec2.py

@@ -58,9 +58,6 @@ class Data2Vec2Model(FrameLevelFeatureModel):
         self.variant = validate_enum(variant, Data2Vec2Variant, Data2Vec2Variant.BASE)
         self._model_id = f"data2vec2-{self.variant.value}"
 
-        self.processor = None
-        self.model = None
-
     def load(self, model_dir: str, quiet: bool = False) -> None:
         """Load the model from the specified directory.
 

{lfeats-0.1.4 → lfeats-0.2.1}/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
 
 
@@ -40,8 +40,6 @@ class EcapaTDNNModel(UtteranceLevelFeatureModel):
         self.variant = validate_enum(variant, EcapaTDNNVariant, EcapaTDNNVariant.BASE)
         self._model_id = f"ecapa-tdnn-{self.variant.value}"
 
-        self.model = None
-
     def load(self, model_dir: str, quiet: bool = False) -> None:
         """Load the model from the specified directory.
 
@@ -78,11 +76,11 @@ class EcapaTDNNModel(UtteranceLevelFeatureModel):
             self.model = EncoderClassifier.from_hparams(
                 source="speechbrain/spkrec-ecapa-voxceleb",
                 fetch_config=fetch_config,
+                run_opts={"device": self.device},
             )
             if self.model is None:
                 raise RuntimeError("Failed to load the model.")
             self.model.eval()
-            self.model.to(self.device)
 
     def extract_features_impl(self, audio: Audio, layers: list[int]) -> Features:
         """Extract features from the input audio using the model.
@@ -110,6 +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)

{lfeats-0.1.4 → lfeats-0.2.1}/lfeats/models/emotion2vec.py

@@ -55,8 +55,6 @@ class Emotion2VecModel(FrameLevelFeatureModel):
         )
         self._model_id = f"emotion2vec-{self.variant.value}"
 
-        self.model = None
-
     def load(self, model_dir: str, quiet: bool = False) -> None:
         """Load the model from the specified directory.
 
@@ -78,7 +76,7 @@ class Emotion2VecModel(FrameLevelFeatureModel):
                 repo_id=repo_id,
                 filename=filename,
                 repo_type="model",
-                local_dir=model_dir,
+                cache_dir=model_dir,
             )
 
         setup_third_party_path()

{lfeats-0.1.4 → lfeats-0.2.1}/lfeats/models/emotion2vec_plus.py

@@ -58,8 +58,6 @@ class Emotion2VecPlusModel(FrameLevelFeatureModel):
         )
         self._model_id = f"emotion2vec+-{self.variant.value}"
 
-        self.model = None
-
     def load(self, model_dir: str, quiet: bool = False) -> None:
         """Load the model from the specified directory.
 
@@ -81,13 +79,13 @@ class Emotion2VecPlusModel(FrameLevelFeatureModel):
                 repo_id=repo_id,
                 filename="model.pt",
                 repo_type="model",
-                local_dir=os.path.join(model_dir, sanitize(self.model_id)),
+                cache_dir=os.path.join(model_dir, sanitize(self.model_id)),
             )
             config = hf_hub_download(
                 repo_id=repo_id,
                 filename="config.yaml",
                 repo_type="model",
-                local_dir=os.path.join(model_dir, sanitize(self.model_id)),
+                cache_dir=os.path.join(model_dir, sanitize(self.model_id)),
             )
 
         from hyperpyyaml import load_hyperpyyaml

lfeats-0.2.1/lfeats/models/higgs_audio.py

@@ -0,0 +1,136 @@
+# Copyright (c) 2026 Takenori Yoshimura
+# Released under the MIT License.
+
+"""A module for the Higgs Audio tokenizer."""
+
+from enum import Enum
+from typing import Any
+
+import torch
+
+from ..interfaces.types import Audio, Features
+from ..utils.io import silence_transformers
+from ..utils.validation import validate_enum
+from .base import TokenLevelFeatureModel
+
+
+class HiggsAudioTokenizerVariant(str, Enum):
+    """Enumeration of supported Higgs Audio tokenizer variants."""
+
+    V2 = "v2"
+
+    @property
+    def model_name(self) -> str:
+        """Return the model name corresponding to the variant.
+
+        Returns
+        -------
+        out : str
+            The model name corresponding to the variant.
+
+        """
+        return f"eustlb/higgs-audio-{self.value}-tokenizer"
+
+
+class HiggsAudioTokenizerModel(TokenLevelFeatureModel):
+    """A class for the Higgs Audio tokenizer model."""
+
+    def __init__(self, variant: str | None = None, device: str = "cpu") -> None:
+        """Initialize the Higgs Audio tokenizer model.
+
+        Parameters
+        ----------
+        variant : str | None, optional
+            The variant of the model to use.
+
+        device : str, optional
+            The device to run the model on (e.g., 'cpu' or 'cuda').
+
+        """
+        super().__init__(variant, device)
+
+        self.variant = validate_enum(
+            variant, HiggsAudioTokenizerVariant, HiggsAudioTokenizerVariant.V2
+        )
+        self._model_id = f"higgs-audio-{self.variant.value}"
+
+        self.feature_extractor = None
+
+    def load(self, model_dir: str, quiet: bool = False) -> None:
+        """Load the model from the specified directory.
+
+        Parameters
+        ----------
+        model_dir : str
+            The directory where the model checkpoint will be stored.
+
+        quiet : bool, optional
+            Whether to suppress output during the loading process.
+
+        """
+        if self.model is not None:
+            return
+
+        from transformers import AutoFeatureExtractor, HiggsAudioV2TokenizerModel
+
+        with silence_transformers(quiet):
+            self.feature_extractor = AutoFeatureExtractor.from_pretrained(
+                self.variant.model_name, cache_dir=model_dir
+            )
+            self.model = HiggsAudioV2TokenizerModel.from_pretrained(
+                self.variant.model_name, cache_dir=model_dir
+            )
+            self.model.eval()
+            self.model.to(self.device)  # type: ignore
+
+    def extract_features_impl(self, audio: Audio, layers: list[int]) -> Features:
+        """Extract features from the input audio using the model.
+
+        Parameters
+        ----------
+        audio : Audio
+            The input audio data with shape (B, T).
+
+        layers : list[int]
+            The layer(s) from which to extract features.
+
+        Returns
+        -------
+        out : Features
+            The extracted features.
+
+        Raises
+        ------
+        RuntimeError
+            If the model is not loaded.
+
+        """
+        if self.feature_extractor is None or self.model is None:
+            raise RuntimeError("Model not loaded. Call 'load' method first.")
+
+        with torch.inference_mode():
+            inputs = self.feature_extractor(
+                raw_audio=[x for x in audio.array],
+                sampling_rate=self.feature_extractor.sampling_rate,
+                return_tensors="pt",
+            ).to(self.device)
+
+            encoder_outputs: Any = self.model.encode(inputs["input_values"])
+            indices = encoder_outputs.audio_codes  # (B, Q, N)
+            indices = indices.transpose(0, 1)
+            vectors = self.model.quantizer.decode(indices)  # (B, D, N)
+            vectors = vectors.transpose(1, 2)
+
+        return Features(data=vectors, source=self.model_id)
+
+    @property
+    def sample_rate(self) -> int:
+        """Get the sample rate required by the model.
+
+        Returns
+        -------
+        out : int
+            The sample rate in Hz.
+
+        """
+        return 24000