librosax.istft

istft(stft_matrix: Array, hop_length: int = None, win_length: int = None, n_fft: int = None, window: str = 'hann', center: bool = True, length: int = None)[source]

Compute the Inverse Short-Time Fourier Transform (ISTFT).

This function reconstructs a waveform from an STFT matrix using JAX’s scipy.signal.istft implementation.

Parameters:

  • stft_matrix –

    Complex STFT matrix. The second-to-last axis must be frequency (n_fft // 2 + 1 bins) and the last axis must be time frames. Supported shapes:

    • (F, N) - single spectrogram

    • (B, F, N) - batch of B spectrograms

    • (B, C, F, N) - batch of B spectrograms with C channels

    where F = n_fft // 2 + 1 and N = number of frames.

  • hop_length – Number of samples between successive frames. Default is win_length // 4.

  • win_length – Window size. Default is n_fft.

  • n_fft – FFT size. Default is (stft_matrix.shape[-2] - 1) * 2, inferred from the frequency axis of the input.

  • window – Window function type. Default is "hann". Also supports "sqrt_hann".

  • center – If True, assumes the waveform was padded so that frames were centered. Default is True.

  • length – Target length for the reconstructed signal. If None, the entire signal is returned.

Returns:

Reconstructed time-domain signal. The last axis is time (samples). Output shapes correspond to inputs:

  • (F, N) → (T,)

  • (B, F, N) → (B, T)

  • (B, C, F, N) → (B, C, T)

where T is the reconstructed signal length.

Raises:

AssertionError – If center is False because the function is only tested for center=True.

Examples

>>> import jax.numpy as jnp
>>> import librosax
>>> # Single STFT: (n_freq, n_frames) -> (n_samples,)
>>> S = jnp.zeros((1025, 44), dtype=jnp.complex64)
>>> y = librosax.istft(S, hop_length=512)
>>> y.shape
(22016,)
>>> # Batched STFT: (batch, n_freq, n_frames) -> (batch, n_samples)
>>> S_batch = jnp.zeros((4, 1025, 44), dtype=jnp.complex64)
>>> y_batch = librosax.istft(S_batch, hop_length=512)
>>> y_batch.shape
(4, 22016)
>>> # With target length
>>> y_trimmed = librosax.istft(S, hop_length=512, length=22050)
>>> y_trimmed.shape
(22050,)