DSP Modules

Beamforming

class asteroid.dsp.beamforming.Beamformer(*args: Any, **kwargs: Any)

Bases: torch.nn.Module

Base class for beamforming modules.

static apply_beamforming_vector(bf_vector: torch.Tensor, mix: torch.Tensor)

Apply the beamforming vector to the mixture. Output (batch, freqs, frames).

Parameters

• bf_vector – shape (batch, mics, freqs)

• mix – shape (batch, mics, freqs, frames).

static get_reference_mic_vects(ref_mic, bf_mat: torch.Tensor, target_scm: torch.Tensor = None, noise_scm: torch.Tensor = None)

Return the reference channel indices over the batch.

Parameters

• ref_mic (Optional[Union[int, torch.Tensor]]) – The reference channel. If torch.Tensor (ndim>1), return it, it is the reference mic vector, If torch.LongTensor of size batch, select independent reference mic of the batch. If int, select the corresponding reference mic, If None, the optimal reference mics are computed with get_optimal_reference_mic(), If None, and either SCM is None, ref_mic is set to 0,

• bf_mat – beamforming matrix of shape (batch, freq, mics, mics).

• target_scm (torch.ComplexTensor) – (batch, freqs, mics, mics).

• noise_scm (torch.ComplexTensor) – (batch, freqs, mics, mics).

Returns

torch.LongTensor of size batch to select with the reference channel indices.

class asteroid.dsp.beamforming.SDWMWFBeamformer(mu=1.0)

Bases: asteroid.dsp.beamforming.Beamformer

forward(mix: torch.Tensor, target_scm: torch.Tensor, noise_scm: torch.Tensor, ref_mic: Union[torch.Tensor, torch.LongTensor, int] = None)

Compute and apply SDW-MWF beamformer.

\(\mathbf{w} = \displaystyle (\Sigma_{ss} + \mu \Sigma_{nn})^{-1} \Sigma_{ss}\).

Parameters

• mix (torch.ComplexTensor) – shape (batch, mics, freqs, frames)

• target_scm (torch.ComplexTensor) – (batch, mics, mics, freqs)

• noise_scm (torch.ComplexTensor) – (batch, mics, mics, freqs)

• ref_mic (int) – reference microphone.

Returns

Filtered mixture. torch.ComplexTensor (batch, freqs, frames)

class asteroid.dsp.beamforming.GEVBeamformer(*args: Any, **kwargs: Any)

Bases: asteroid.dsp.beamforming.Beamformer

forward(mix: torch.Tensor, target_scm: torch.Tensor, noise_scm: torch.Tensor)

Compute and apply the GEV beamformer.

\(\mathbf{w} = \displaystyle MaxEig\{ \Sigma_{nn}^{-1}\Sigma_{ss} \}\), where MaxEig extracts the eigenvector corresponding to the maximum eigenvalue (using the GEV decomposition).

Parameters

• mix – shape (batch, mics, freqs, frames)

• target_scm – (batch, mics, mics, freqs)

• noise_scm – (batch, mics, mics, freqs)

Returns

Filtered mixture. torch.ComplexTensor (batch, freqs, frames)

class asteroid.dsp.beamforming.RTFMVDRBeamformer(*args: Any, **kwargs: Any)

Bases: asteroid.dsp.beamforming.Beamformer

forward(mix: torch.Tensor, target_scm: torch.Tensor, noise_scm: torch.Tensor)

Compute and apply MVDR beamformer from the speech and noise SCM matrices.

\(\mathbf{w} = \displaystyle \frac{\Sigma_{nn}^{-1} \mathbf{a}}{ \mathbf{a}^H \Sigma_{nn}^{-1} \mathbf{a}}\) where \(\mathbf{a}\) is the ATF estimated from the target SCM.

Parameters

• mix (torch.ComplexTensor) – shape (batch, mics, freqs, frames)

• target_scm (torch.ComplexTensor) – (batch, mics, mics, freqs)

• noise_scm (torch.ComplexTensor) – (batch, mics, mics, freqs)

Returns

Filtered mixture. torch.ComplexTensor (batch, freqs, frames)

from_rtf_vect(mix: torch.Tensor, rtf_vec: torch.Tensor, noise_scm: torch.Tensor)

Compute and apply MVDR beamformer from the ATF vector and noise SCM matrix.

Parameters

• mix (torch.ComplexTensor) – shape (batch, mics, freqs, frames)

• rtf_vec (torch.ComplexTensor) – (batch, mics, freqs)

• noise_scm (torch.ComplexTensor) – (batch, mics, mics, freqs)

Returns

Filtered mixture. torch.ComplexTensor (batch, freqs, frames)

class asteroid.dsp.beamforming.SoudenMVDRBeamformer(*args: Any, **kwargs: Any)

Bases: asteroid.dsp.beamforming.Beamformer

forward(mix: torch.Tensor, target_scm: torch.Tensor, noise_scm: torch.Tensor, ref_mic: Union[torch.Tensor, torch.LongTensor, int] = 0, eps=1e-08)

Compute and apply MVDR beamformer from the speech and noise SCM matrices. This class uses Souden’s formulation [1].

\(\mathbf{w} = \displaystyle \frac{\Sigma_{nn}^{-1} \Sigma_{ss}}{ Tr\left( \Sigma_{nn}^{-1} \Sigma_{ss} \right) }\mathbf{u}\) where \(\mathbf{a}\) is the steering vector.

Parameters

• mix (torch.ComplexTensor) – shape (batch, mics, freqs, frames)

• target_scm (torch.ComplexTensor) – (batch, mics, mics, freqs)

• noise_scm (torch.ComplexTensor) – (batch, mics, mics, freqs)

• ref_mic (int) – reference microphone.

• eps – numerical stabilizer.

Returns

Filtered mixture. torch.ComplexTensor (batch, freqs, frames)

References

[1] Souden, M., Benesty, J., & Affes, S. (2009). On optimal frequency-domain multichannel linear filtering for noise reduction. IEEE Transactions on audio, speech, and language processing, 18(2), 260-276.

class asteroid.dsp.beamforming.SCM(*args: Any, **kwargs: Any)

Bases: torch.nn.Module

forward(x: torch.Tensor, mask: torch.Tensor = None, normalize: bool = True)

See compute_scm().

LambdaOverlapAdd

class asteroid.dsp.LambdaOverlapAdd(nnet, n_src, window_size, hop_size=None, window='hann', reorder_chunks=True, enable_grad=False)

Bases: torch.nn.Module

Overlap-add with lambda transform on segments (not scriptable).

Segment input signal, apply lambda function (a neural network for example) and combine with OLA.

LambdaOverlapAdd can be used with asteroid.separate and the asteroid-infer CLI.

Parameters

• nnet (callable) – Function to apply to each segment.

• n_src (Optional[int]) – Number of sources in the output of nnet. If None, the number of sources is determined by the network’s output, but some correctness checks cannot be performed.

• window_size (int) – Size of segmenting window.

• hop_size (int) – Segmentation hop size.

• window (str) – Name of the window (see scipy.signal.get_window) used for the synthesis.

• reorder_chunks – Whether to reorder each consecutive segment. This might be useful when nnet is permutation invariant, as source assignements might change output channel from one segment to the next (in classic speech separation for example). Reordering is performed based on the correlation between the overlapped part of consecutive segment.

ola_forward(x)

Heart of the class: segment signal, apply func, combine with OLA.

forward(x)

Forward module: segment signal, apply func, combine with OLA.

Parameters

x (torch.Tensor) – waveform signal of shape (batch, 1, time).

Returns

torch.Tensor – The output of the lambda OLA.

DualPath Processing

class asteroid.dsp.DualPathProcessing(chunk_size, hop_size)

Bases: torch.nn.Module

Perform Dual-Path processing via overlap-add as in DPRNN [1].

Parameters

• chunk_size (int) – Size of segmenting window.

• hop_size (int) – segmentation hop size.

References

[1] Yi Luo, Zhuo Chen and Takuya Yoshioka. “Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation” https://arxiv.org/abs/1910.06379

unfold(x)

Unfold the feature tensor from $(batch, channels, time)$ to $(batch, channels, chunksize, nchunks)$.

Parameters

x (torch.Tensor) – feature tensor of shape $(batch, channels, time)$.

Returns

torch.Tensor – spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$.

fold(x, output_size=None)

Folds back the spliced feature tensor. Input shape $(batch, channels, chunksize, nchunks)$ to original shape $(batch, channels, time)$ using overlap-add.

Parameters

• x (torch.Tensor) – spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$.

• output_size (int, optional) – sequence length of original feature tensor. If None, the original length cached by the previous call of unfold() will be used.

Returns

torch.Tensor – feature tensor of shape $(batch, channels, time)$.

Note

fold caches the original length of the input.

static intra_process(x, module)

Performs intra-chunk processing.

Parameters

• x (torch.Tensor) – spliced feature tensor of shape (batch, channels, chunk_size, n_chunks).

• module (torch.nn.Module) – module one wish to apply to each chunk of the spliced feature tensor.

Returns

torch.Tensor – processed spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$.

Note

the module should have the channel first convention and accept a 3D tensor of shape $(batch, channels, time)$.

static inter_process(x, module)

Performs inter-chunk processing.

Parameters

• x (torch.Tensor) – spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$.

• module (torch.nn.Module) – module one wish to apply between each chunk of the spliced feature tensor.

Returns

x (torch.Tensor) – processed spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$.

Note

the module should have the channel first convention and accept a 3D tensor of shape $(batch, channels, time)$.

Mixture Consistency

asteroid.dsp.mixture_consistency(mixture: torch.Tensor, est_sources: torch.Tensor, src_weights: Optional[torch.Tensor] = None, dim: int = 1) → torch.Tensor

Applies mixture consistency to a tensor of estimated sources.

Parameters

• mixture (torch.Tensor) – Mixture waveform or TF representation.

• est_sources (torch.Tensor) – Estimated sources waveforms or TF representations.

• src_weights (torch.Tensor) – Consistency weight for each source. Shape needs to be broadcastable to est_source. We make sure that the weights sum up to 1 along dim dim. If src_weights is None, compute them based on relative power.

• dim (int) – Axis which contains the sources in est_sources.

Returns

torch.Tensor with same shape as est_sources, after applying mixture consistency.

Examples

>>> # Works on waveforms
>>> mix = torch.randn(10, 16000)
>>> est_sources = torch.randn(10, 2, 16000)
>>> new_est_sources = mixture_consistency(mix, est_sources, dim=1)
>>> # Also works on spectrograms
>>> mix = torch.randn(10, 514, 400)
>>> est_sources = torch.randn(10, 2, 514, 400)
>>> new_est_sources = mixture_consistency(mix, est_sources, dim=1)

Note

This method can be used only in ‘complete’ separation tasks, otherwise the residual error will contain unwanted sources. For example, this won’t work with the task “sep_noisy” from WHAM.

References

Scott Wisdom et al. “Differentiable consistency constraints for improved deep speech enhancement”, ICASSP 2019.

VAD

asteroid.dsp.vad.ebased_vad(mag_spec, th_db: int = 40)

Compute energy-based VAD from a magnitude spectrogram (or equivalent).

Parameters

• mag_spec (torch.Tensor) – the spectrogram to perform VAD on. Expected shape (batch, *, freq, time). The VAD mask will be computed independently for all the leading dimensions until the last two. Independent of the ordering of the last two dimensions.

• th_db (int) – The threshold in dB from which a TF-bin is considered silent.

Returns

torch.BoolTensor, the VAD mask.

Examples

>>> import torch
>>> mag_spec = torch.abs(torch.randn(10, 2, 65, 16))
>>> batch_src_mask = ebased_vad(mag_spec)

Delta Features

asteroid.dsp.deltas.compute_delta(feats: torch.Tensor, dim: int = -1) → torch.Tensor

Compute delta coefficients of a tensor.

Parameters

• feats – Input features to compute deltas with.

• dim – feature dimension in the feats tensor.

Returns

Tensor – Tensor of deltas.

Examples

>>> import torch
>>> phase = torch.randn(2, 257, 100)
>>> # Compute instantaneous frequency
>>> inst_freq = compute_delta(phase, dim=-1)
>>> # Or group delay
>>> group_delay = compute_delta(phase, dim=-2)

asteroid.dsp.deltas.concat_deltas(feats: torch.Tensor, order: int = 1, dim: int = -1) → torch.Tensor

Concatenate delta coefficients of a tensor to itself.

Parameters

• feats – Input features to compute deltas with.

• order – Order of the delta e.g with order==2, compute delta of delta as well.

• dim – feature dimension in the feats tensor.

Returns

Tensor – Concatenation of the features, the deltas and subsequent deltas.

Examples

>>> import torch
>>> phase = torch.randn(2, 257, 100)
>>> # Compute second order instantaneous frequency
>>> phase_and_inst_freq = concat_deltas(phase, order=2, dim=-1)
>>> # Or group delay
>>> phase_and_group_delay = concat_deltas(phase, order=2, dim=-2)