torchaudio 2.7.1__cp312-cp312-win_amd64.whl → 2.9.0__cp312-cp312-win_amd64.whl

torchaudio/_torchcodec.py

@@ -0,0 +1,340 @@
+"""TorchCodec integration for TorchAudio."""
+
+import os
+from typing import BinaryIO, Optional, Tuple, Union
+
+import torch
+
+
+def load_with_torchcodec(
+    uri: Union[BinaryIO, str, os.PathLike],
+    frame_offset: int = 0,
+    num_frames: int = -1,
+    normalize: bool = True,
+    channels_first: bool = True,
+    format: Optional[str] = None,
+    buffer_size: int = 4096,
+    backend: Optional[str] = None,
+) -> Tuple[torch.Tensor, int]:
+    """Load audio data from source using TorchCodec's AudioDecoder.
+
+    .. note::
+
+        This function supports the same API as :func:`~torchaudio.load`, and
+        relies on TorchCodec's decoding capabilities under the hood. It is
+        provided for convenience, but we do recommend that you port your code to
+        natively use ``torchcodec``'s ``AudioDecoder`` class for better
+        performance:
+        https://docs.pytorch.org/torchcodec/stable/generated/torchcodec.decoders.AudioDecoder.
+        As of TorchAudio 2.9, :func:`~torchaudio.load` relies on
+        :func:`~torchaudio.load_with_torchcodec`. Note that some parameters of
+        :func:`~torchaudio.load`, like ``normalize``, ``buffer_size``, and
+        ``backend``, are ignored by :func:`~torchaudio.load_with_torchcodec`.
+        To install torchcodec, follow the instructions at https://github.com/pytorch/torchcodec#installing-torchcodec.
+
+
+    Args:
+        uri (path-like object or file-like object):
+            Source of audio data. The following types are accepted:
+
+            * ``path-like``: File path or URL.
+            * ``file-like``: Object with ``read(size: int) -> bytes`` method.
+
+        frame_offset (int, optional):
+            Number of samples to skip before start reading data.
+        num_frames (int, optional):
+            Maximum number of samples to read. ``-1`` reads all the remaining samples,
+            starting from ``frame_offset``.
+        normalize (bool, optional):
+            TorchCodec always returns normalized float32 samples. This parameter
+            is ignored and a warning is issued if set to False.
+            Default: ``True``.
+        channels_first (bool, optional):
+            When True, the returned Tensor has dimension `[channel, time]`.
+            Otherwise, the returned Tensor's dimension is `[time, channel]`.
+        format (str or None, optional):
+            Format hint for the decoder. May not be supported by all TorchCodec
+            decoders. (Default: ``None``)
+        buffer_size (int, optional):
+            Not used by TorchCodec AudioDecoder. Provided for API compatibility.
+        backend (str or None, optional):
+            Not used by TorchCodec AudioDecoder. Provided for API compatibility.
+
+    Returns:
+        (torch.Tensor, int): Resulting Tensor and sample rate.
+        Always returns float32 tensors. If ``channels_first=True``, shape is
+        `[channel, time]`, otherwise `[time, channel]`.
+
+    Raises:
+        ImportError: If torchcodec is not available.
+        ValueError: If unsupported parameters are used.
+        RuntimeError: If TorchCodec fails to decode the audio.
+
+    Note:
+        - TorchCodec always returns normalized float32 samples, so the ``normalize``
+          parameter has no effect.
+        - The ``buffer_size`` and ``backend`` parameters are ignored.
+        - Not all audio formats supported by torchaudio backends may be supported
+          by TorchCodec.
+    """
+    # Import torchcodec here to provide clear error if not available
+    try:
+        from torchcodec.decoders import AudioDecoder
+    except ImportError as e:
+        raise ImportError(
+            "TorchCodec is required for load_with_torchcodec. " "Please install torchcodec to use this function."
+        ) from e
+
+    # Parameter validation and warnings
+    if not normalize:
+        import warnings
+
+        warnings.warn(
+            "TorchCodec AudioDecoder always returns normalized float32 samples. "
+            "The 'normalize=False' parameter is ignored.",
+            UserWarning,
+            stacklevel=2,
+        )
+
+    if buffer_size != 4096:
+        import warnings
+
+        warnings.warn("The 'buffer_size' parameter is not used by TorchCodec AudioDecoder.", UserWarning, stacklevel=2)
+
+    if backend is not None:
+        import warnings
+
+        warnings.warn("The 'backend' parameter is not used by TorchCodec AudioDecoder.", UserWarning, stacklevel=2)
+
+    if format is not None:
+        import warnings
+
+        warnings.warn("The 'format' parameter is not supported by TorchCodec AudioDecoder.", UserWarning, stacklevel=2)
+
+    # Create AudioDecoder
+    try:
+        decoder = AudioDecoder(uri)
+    except Exception as e:
+        raise RuntimeError(f"Failed to create AudioDecoder for {uri}: {e}") from e
+
+    # Get sample rate from metadata
+    sample_rate = decoder.metadata.sample_rate
+    if sample_rate is None:
+        raise RuntimeError("Unable to determine sample rate from audio metadata")
+
+    # Decode the entire file first, then subsample manually
+    # This is the simplest approach since torchcodec uses time-based indexing
+    try:
+        audio_samples = decoder.get_all_samples()
+    except Exception as e:
+        raise RuntimeError(f"Failed to decode audio samples: {e}") from e
+
+    data = audio_samples.data
+
+    # Apply frame_offset and num_frames (which are actually sample offsets)
+    if frame_offset > 0:
+        if frame_offset >= data.shape[1]:
+            # Return empty tensor if offset is beyond available data
+            empty_shape = (data.shape[0], 0) if channels_first else (0, data.shape[0])
+            return torch.zeros(empty_shape, dtype=torch.float32), sample_rate
+        data = data[:, frame_offset:]
+
+    if num_frames == 0:
+        # Return empty tensor if num_frames is 0
+        empty_shape = (data.shape[0], 0) if channels_first else (0, data.shape[0])
+        return torch.zeros(empty_shape, dtype=torch.float32), sample_rate
+    elif num_frames > 0:
+        data = data[:, :num_frames]
+
+    # TorchCodec returns data in [channel, time] format by default
+    # Handle channels_first parameter
+    if not channels_first:
+        data = data.transpose(0, 1)  # [channel, time] -> [time, channel]
+
+    return data, sample_rate
+
+
+def save_with_torchcodec(
+    uri: Union[str, os.PathLike],
+    src: torch.Tensor,
+    sample_rate: int,
+    channels_first: bool = True,
+    format: Optional[str] = None,
+    encoding: Optional[str] = None,
+    bits_per_sample: Optional[int] = None,
+    buffer_size: int = 4096,
+    backend: Optional[str] = None,
+    compression: Optional[Union[float, int]] = None,
+) -> None:
+    """Save audio data to file using TorchCodec's AudioEncoder.
+
+    .. note::
+
+        This function supports the same API as :func:`~torchaudio.save`, and
+        relies on TorchCodec's encoding capabilities under the hood. It is
+        provided for convenience, but we do recommend that you port your code to
+        natively use ``torchcodec``'s ``AudioEncoder`` class for better
+        performance:
+        https://docs.pytorch.org/torchcodec/stable/generated/torchcodec.encoders.AudioEncoder.
+        As of TorchAudio 2.9, :func:`~torchaudio.save` relies on
+        :func:`~torchaudio.save_with_torchcodec`. Note that some parameters of
+        :func:`~torchaudio.save`, like ``format``, ``encoding``,
+        ``bits_per_sample``, ``buffer_size``, and ``backend``, are ignored by
+        are ignored by :func:`~torchaudio.save_with_torchcodec`.
+        To install torchcodec, follow the instructions at https://github.com/pytorch/torchcodec#installing-torchcodec.
+
+    This function provides a TorchCodec-based alternative to torchaudio.save
+    with the same API. TorchCodec's AudioEncoder provides efficient encoding
+    with FFmpeg under the hood.
+
+    Args:
+        uri (path-like object):
+            Path to save the audio file. The file extension determines the format.
+
+        src (torch.Tensor):
+            Audio data to save. Must be a 1D or 2D tensor with float32 values
+            in the range [-1, 1]. If 2D, shape should be [channel, time] when
+            channels_first=True, or [time, channel] when channels_first=False.
+
+        sample_rate (int):
+            Sample rate of the audio data.
+
+        channels_first (bool, optional):
+            Indicates whether the input tensor has channels as the first dimension.
+            If True, expects [channel, time]. If False, expects [time, channel].
+            Default: True.
+
+        format (str or None, optional):
+            Audio format hint. Not used by TorchCodec (format is determined by
+            file extension). A warning is issued if provided.
+            Default: None.
+
+        encoding (str or None, optional):
+            Audio encoding. Not fully supported by TorchCodec AudioEncoder.
+            A warning is issued if provided. Default: None.
+
+        bits_per_sample (int or None, optional):
+            Bits per sample. Not directly supported by TorchCodec AudioEncoder.
+            A warning is issued if provided. Default: None.
+
+        buffer_size (int, optional):
+            Not used by TorchCodec AudioEncoder. Provided for API compatibility.
+            A warning is issued if not default value. Default: 4096.
+
+        backend (str or None, optional):
+            Not used by TorchCodec AudioEncoder. Provided for API compatibility.
+            A warning is issued if provided. Default: None.
+
+        compression (float, int or None, optional):
+            Compression level or bit rate. Maps to bit_rate parameter in
+            TorchCodec AudioEncoder. Default: None.
+
+    Raises:
+        ImportError: If torchcodec is not available.
+        ValueError: If input parameters are invalid.
+        RuntimeError: If TorchCodec fails to encode the audio.
+
+    Note:
+        - TorchCodec AudioEncoder expects float32 samples in [-1, 1] range.
+        - Some parameters (format, encoding, bits_per_sample, buffer_size, backend)
+          are not used by TorchCodec but are provided for API compatibility.
+        - The output format is determined by the file extension in the uri.
+        - TorchCodec uses FFmpeg under the hood for encoding.
+    """
+    # Import torchcodec here to provide clear error if not available
+    try:
+        from torchcodec.encoders import AudioEncoder
+    except ImportError as e:
+        raise ImportError(
+            "TorchCodec is required for save_with_torchcodec. " "Please install torchcodec to use this function."
+        ) from e
+
+    # Parameter validation and warnings
+    if format is not None:
+        import warnings
+
+        warnings.warn(
+            "The 'format' parameter is not used by TorchCodec AudioEncoder. "
+            "Format is determined by the file extension.",
+            UserWarning,
+            stacklevel=2,
+        )
+
+    if encoding is not None:
+        import warnings
+
+        warnings.warn(
+            "The 'encoding' parameter is not fully supported by TorchCodec AudioEncoder.", UserWarning, stacklevel=2
+        )
+
+    if bits_per_sample is not None:
+        import warnings
+
+        warnings.warn(
+            "The 'bits_per_sample' parameter is not directly supported by TorchCodec AudioEncoder.",
+            UserWarning,
+            stacklevel=2,
+        )
+
+    if buffer_size != 4096:
+        import warnings
+
+        warnings.warn("The 'buffer_size' parameter is not used by TorchCodec AudioEncoder.", UserWarning, stacklevel=2)
+
+    if backend is not None:
+        import warnings
+
+        warnings.warn("The 'backend' parameter is not used by TorchCodec AudioEncoder.", UserWarning, stacklevel=2)
+
+    # Input validation
+    if not isinstance(src, torch.Tensor):
+        raise ValueError(f"Expected src to be a torch.Tensor, got {type(src)}")
+
+    if src.dtype != torch.float32:
+        src = src.float()
+
+    if sample_rate <= 0:
+        raise ValueError(f"sample_rate must be positive, got {sample_rate}")
+
+    # Handle tensor shape and channels_first
+    if src.ndim == 1:
+        # Convert to 2D: [1, time] for channels_first=True
+        if channels_first:
+            data = src.unsqueeze(0)  # [1, time]
+        else:
+            # For channels_first=False, input is [time] -> reshape to [time, 1] -> transpose to [1, time]
+            data = src.unsqueeze(1).transpose(0, 1)  # [time, 1] -> [1, time]
+    elif src.ndim == 2:
+        if channels_first:
+            data = src  # Already [channel, time]
+        else:
+            data = src.transpose(0, 1)  # [time, channel] -> [channel, time]
+    else:
+        raise ValueError(f"Expected 1D or 2D tensor, got {src.ndim}D tensor")
+
+    # Create AudioEncoder
+    try:
+        encoder = AudioEncoder(data, sample_rate=sample_rate)
+    except Exception as e:
+        raise RuntimeError(f"Failed to create AudioEncoder: {e}") from e
+
+    # Determine bit_rate from compression parameter
+    bit_rate = None
+    if compression is not None:
+        if isinstance(compression, (int, float)):
+            bit_rate = int(compression)
+        else:
+            import warnings
+
+            warnings.warn(
+                f"Unsupported compression type {type(compression)}. "
+                "TorchCodec AudioEncoder expects int or float for bit_rate.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+    # Save to file
+    try:
+        encoder.to_file(uri, bit_rate=bit_rate)
+    except Exception as e:
+        raise RuntimeError(f"Failed to save audio to {uri}: {e}") from e

torchaudio/datasets/cmuarctic.py

@@ -129,7 +129,7 @@ class CMUARCTIC(Dataset):
         self._text = os.path.join(self._path, self._folder_text, self._file_text)
 
         with open(self._text, "r") as text:
-            walker = csv.reader(text, delimiter="\n")
+            walker = csv.reader(text)
             self._walker = list(walker)
 
     def __getitem__(self, n: int) -> Tuple[Tensor, int, str, str]:

torchaudio/datasets/utils.py

@@ -2,7 +2,7 @@ import logging
 import os
 import tarfile
 import zipfile
-from typing import Any, List, Optional
+from typing import Any, List, Optional  # noqa: F401
 
 import torchaudio
 

torchaudio/functional/__init__.py

@@ -1,4 +1,6 @@
-from ._alignment import forced_align, merge_tokens, TokenSpan
+from torchaudio._internal.module_utils import dropping_support
+
+from ._alignment import forced_align as _forced_align, merge_tokens, TokenSpan
 from .filtering import (
     allpass_biquad,
     band_biquad,
@@ -23,11 +25,13 @@ from .filtering import (
     treble_biquad,
     vad,
 )
+
+forced_align = dropping_support(_forced_align)
+
 from .functional import (
     add_noise,
     amplitude_to_DB,
     apply_beamforming,
-    apply_codec,
     compute_deltas,
     convolve,
     create_dct,
@@ -106,7 +110,6 @@ __all__ = [
     "riaa_biquad",
     "treble_biquad",
     "vad",
-    "apply_codec",
     "resample",
     "edit_distance",
     "pitch_shift",

torchaudio/functional/_alignment.py

@@ -70,7 +70,7 @@ def forced_align(
     assert target_lengths is not None
 
     paths, scores = torch.ops.torchaudio.forced_align(log_probs, targets, input_lengths, target_lengths, blank)
-    return paths, scores
+    return paths, scores[:, torch.arange(scores.shape[1]), paths[0]]
 
 
 @dataclass

torchaudio/functional/filtering.py

@@ -3,6 +3,7 @@ import warnings
 from typing import Optional
 
 import torch
+import torch.nn.functional as F
 from torch import Tensor
 
 from torchaudio._extension import _IS_TORCHAUDIO_EXT_AVAILABLE
@@ -932,69 +933,83 @@ def _lfilter_core_generic_loop(input_signal_windows: Tensor, a_coeffs_flipped: T
 
 
 if _IS_TORCHAUDIO_EXT_AVAILABLE:
-    _lfilter_core_cpu_loop = torch.ops.torchaudio._lfilter_core_loop
+    _lfilter_core_loop = torch.ops.torchaudio._lfilter_core_loop
 else:
-    _lfilter_core_cpu_loop = _lfilter_core_generic_loop
+    _lfilter_core_loop = _lfilter_core_generic_loop
+
+
+class DifferentiableFIR(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, waveform, b_coeffs):
+        n_order = b_coeffs.size(1)
+        n_channel = b_coeffs.size(0)
+        b_coeff_flipped = b_coeffs.flip(1).contiguous()
+        padded_waveform = F.pad(waveform, (n_order - 1, 0))
+        output = F.conv1d(padded_waveform, b_coeff_flipped.unsqueeze(1), groups=n_channel)
+        ctx.save_for_backward(waveform, b_coeffs, output)
+        return output
+
+    @staticmethod
+    def backward(ctx, dy):
+        x, b_coeffs, y = ctx.saved_tensors
+        n_batch = x.size(0)
+        n_channel = x.size(1)
+        n_order = b_coeffs.size(1)
+        db = (
+            F.conv1d(
+                F.pad(x, (n_order - 1, 0)).view(1, n_batch * n_channel, -1),
+                dy.view(n_batch * n_channel, 1, -1),
+                groups=n_batch * n_channel,
+            )
+            .view(n_batch, n_channel, -1)
+            .sum(0)
+            .flip(1)
+            if b_coeffs.requires_grad
+            else None
+        )
+        dx = F.conv1d(F.pad(dy, (0, n_order - 1)), b_coeffs.unsqueeze(1), groups=n_channel) if x.requires_grad else None
+        return (dx, db)
 
 
-def _lfilter_core(
-    waveform: Tensor,
-    a_coeffs: Tensor,
-    b_coeffs: Tensor,
-) -> Tensor:
+class DifferentiableIIR(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, waveform, a_coeffs_normalized):
+        n_batch, n_channel, n_sample = waveform.shape
+        n_order = a_coeffs_normalized.size(1)
+        n_sample_padded = n_sample + n_order - 1
 
-    if a_coeffs.size() != b_coeffs.size():
-        raise ValueError(
-            "Expected coeffs to be the same size."
-            f"Found a_coeffs size: {a_coeffs.size()}, b_coeffs size: {b_coeffs.size()}"
+        a_coeff_flipped = a_coeffs_normalized.flip(1).contiguous()
+        padded_output_waveform = torch.zeros(
+            n_batch, n_channel, n_sample_padded, device=waveform.device, dtype=waveform.dtype
         )
-    if waveform.ndim != 3:
-        raise ValueError(f"Expected waveform to be 3 dimensional. Found: {waveform.ndim}")
-    if not (waveform.device == a_coeffs.device == b_coeffs.device):
-        raise ValueError(
-            "Expected waveform and coeffs to be on the same device."
-            f"Found: waveform device:{waveform.device}, a_coeffs device: {a_coeffs.device}, "
-            f"b_coeffs device: {b_coeffs.device}"
+        _lfilter_core_loop(waveform, a_coeff_flipped, padded_output_waveform)
+        output = padded_output_waveform[:, :, n_order - 1 :]
+        ctx.save_for_backward(waveform, a_coeffs_normalized, output)
+        return output
+
+    @staticmethod
+    def backward(ctx, dy):
+        x, a_coeffs_normalized, y = ctx.saved_tensors
+        n_channel = x.size(1)
+        n_order = a_coeffs_normalized.size(1)
+        tmp = DifferentiableIIR.apply(dy.flip(2).contiguous(), a_coeffs_normalized).flip(2)
+        dx = tmp if x.requires_grad else None
+        da = (
+            -(
+                tmp.transpose(0, 1).reshape(n_channel, 1, -1)
+                @ F.pad(y, (n_order - 1, 0)).unfold(2, n_order, 1).transpose(0, 1).reshape(n_channel, -1, n_order)
+            )
+            .squeeze(1)
+            .flip(1)
+            if a_coeffs_normalized.requires_grad
+            else None
         )
+        return (dx, da)
 
-    n_batch, n_channel, n_sample = waveform.size()
-    n_order = a_coeffs.size(1)
-    if n_order <= 0:
-        raise ValueError(f"Expected n_order to be positive. Found: {n_order}")
-
-    # Pad the input and create output
 
-    padded_waveform = torch.nn.functional.pad(waveform, [n_order - 1, 0])
-    padded_output_waveform = torch.zeros_like(padded_waveform)
-
-    # Set up the coefficients matrix
-    # Flip coefficients' order
-    a_coeffs_flipped = a_coeffs.flip(1)
-    b_coeffs_flipped = b_coeffs.flip(1)
-
-    # calculate windowed_input_signal in parallel using convolution
-    input_signal_windows = torch.nn.functional.conv1d(padded_waveform, b_coeffs_flipped.unsqueeze(1), groups=n_channel)
-
-    input_signal_windows.div_(a_coeffs[:, :1])
-    a_coeffs_flipped.div_(a_coeffs[:, :1])
-
-    if (
-        input_signal_windows.device == torch.device("cpu")
-        and a_coeffs_flipped.device == torch.device("cpu")
-        and padded_output_waveform.device == torch.device("cpu")
-    ):
-        _lfilter_core_cpu_loop(input_signal_windows, a_coeffs_flipped, padded_output_waveform)
-    else:
-        _lfilter_core_generic_loop(input_signal_windows, a_coeffs_flipped, padded_output_waveform)
-
-    output = padded_output_waveform[:, :, n_order - 1 :]
-    return output
-
-
-if _IS_TORCHAUDIO_EXT_AVAILABLE:
-    _lfilter = torch.ops.torchaudio._lfilter
-else:
-    _lfilter = _lfilter_core
+def _lfilter(waveform, a_coeffs, b_coeffs):
+    filtered_waveform = DifferentiableFIR.apply(waveform, b_coeffs / a_coeffs[:, 0:1])
+    return DifferentiableIIR.apply(filtered_waveform, a_coeffs / a_coeffs[:, 0:1])
 
 
 def lfilter(waveform: Tensor, a_coeffs: Tensor, b_coeffs: Tensor, clamp: bool = True, batching: bool = True) -> Tensor: