Resample

korvax.resample

resample(
    x,
    /,
    orig_sr,
    target_sr,
    lowpass_filter_width=6,
    rolloff=0.99,
    resampling_method="sinc_interp_hann",
    beta=None,
    scale=False,
)

Resample a waveform using sinc interpolation.

This function is a JAX port of torchaudio.resample. When jitted, it is just as fast as soxr (HQ) on CPU, but the JAX implementation is also fully differentiable and works on GPU/TPU.

Note: Unlike the rest of korvax, this function requires sampling rates to be specified as integers.

Parameters:

Name	Type	Description	Default
x	Float[Array, '*channels old_n_samples']	The input signal of dimension (..., time).	required
orig_sr	int	The original frequency of the signal.	required
target_sr	int	The desired frequency.	required
lowpass_filter_width	int	Controls the sharpness of the filter. A larger value gives a sharper filter but is less efficient. Defaults to 6.	6
rolloff	float	The roll-off frequency of the filter as a fraction of the Nyquist frequency. Lower values reduce aliasing but also attenuate high frequencies. Defaults to 0.99.	0.99
resampling_method	str	The windowing function to use. Options: ["sinc_interp_hann", "sinc_interp_kaiser"]. Defaults to "sinc_interp_hann".	'sinc_interp_hann'
beta	float | None	The shape parameter for the Kaiser window. Only used if resampling_method is "sinc_interp_kaiser". Defaults to 14.7696....	None

Returns:

Type	Description
Float[Array, '*channels new_n_samples']	The waveform at the new frequency. The new shape is (..., int(ceil(target_sr * old_n_samples / orig_sr))).

Source code in src/korvax/_resample.py

def resample(
    x: Float[Array, "*channels old_n_samples"],
    /,
    orig_sr: int,
    target_sr: int,
    lowpass_filter_width: int = 6,
    rolloff: float = 0.99,
    resampling_method: str = "sinc_interp_hann",
    beta: float | None = None,
    scale: bool = False,
) -> Float[Array, "*channels new_n_samples"]:
    """Resample a waveform using sinc interpolation.

    This function is a JAX port of [`torchaudio.resample`](https://docs.pytorch.org/audio/main/generated/torchaudio.functional.resample.html).
    When jitted, it is just as fast as [`soxr`](https://github.com/dofuuz/python-soxr) (HQ) on CPU,
    but the JAX implementation is also fully differentiable and works on GPU/TPU.

    Note: Unlike the rest of korvax, this function requires sampling rates to be specified as integers.

    Args:
        x: The input signal of dimension `(..., time)`.
        orig_sr: The original frequency of the signal.
        target_sr: The desired frequency.
        lowpass_filter_width: Controls the sharpness of the filter.
            A larger value gives a sharper filter but is less efficient. Defaults to ``6``.
        rolloff: The roll-off frequency of the filter as a
            fraction of the Nyquist frequency. Lower values reduce aliasing but
            also attenuate high frequencies. Defaults to ``0.99``.
        resampling_method: The windowing function to use.
            Options: [``"sinc_interp_hann"``, ``"sinc_interp_kaiser"``].
            Defaults to ``"sinc_interp_hann"``.
        beta: The shape parameter for the Kaiser window.
            Only used if `resampling_method` is ``"sinc_interp_kaiser"``.
            Defaults to ``14.7696...``.

    Returns:
        The waveform at the new frequency. The new shape is `(..., int(ceil(target_sr * old_n_samples / orig_sr)))`.
    """
    if orig_sr <= 0 or target_sr <= 0:
        raise ValueError("Original and new frequencies must be positive.")

    if orig_sr == target_sr:
        return x

    with jax.ensure_compile_time_eval():
        gcd = math.gcd(orig_sr, target_sr)
        kernel, width = _get_sinc_resample_kernel(
            orig_sr,
            target_sr,
            gcd,
            lowpass_filter_width,
            rolloff,
            resampling_method,
            beta,
            dtype=x.dtype,
        )

    x = _apply_sinc_resample_kernel(x, orig_sr, target_sr, gcd, kernel, width)

    if scale:
        x = x / jnp.sqrt((target_sr / orig_sr))

    return x