calibrax.metrics.functional.audio

Audio quality metrics for evaluating signal reconstruction and synthesis. Includes spectral convergence, mel cepstral distortion, and signal-to-noise ratio -- all operating on JAX arrays.

Audio quality metrics -- FFT-based, pure math.

All metrics are pure mathematical operations on audio signal arrays. No pretrained models, no external audio processing libraries.

Includes: spectral_convergence, mel_cepstral_distortion, signal_to_noise_ratio. Registered with domain="audio".

spectral_convergence(predictions: Any, targets: Any) -> Any

Spectral convergence between predicted and target signals.

Computes ||STFT(targets) - STFT(predictions)||_F / ||STFT(targets)||_F.

Parameters:

Name	Type	Description	Default
predictions	Any	Predicted signal array, shape (n,).	required
targets	Any	Target signal array, shape (n,).	required

Returns:

Type	Description
Any	Spectral convergence ratio. Lower is better. 0.0 = identical spectra.

Examples:

>>> import jax.numpy as jnp
>>> sig = jnp.sin(jnp.linspace(0, 10, 256))
>>> spectral_convergence(sig, sig)
0.0

mel_cepstral_distortion(predictions: Any, targets: Any, *, num_mels: int = 80) -> Any

Mel Cepstral Distortion between two signals.

Computes (10/ln(10)) * sqrt(2 * sum((c_pred - c_tgt)^2)) on cepstral coefficients derived from the signal spectra.

Parameters:

Name	Type	Description	Default
predictions	Any	Predicted signal array, shape (n,).	required
targets	Any	Target signal array, shape (n,).	required
num_mels	int	Number of mel-frequency coefficients to use.	80

Returns:

Type	Description
Any	MCD value in dB. Lower is better. 0.0 = identical signals.

Examples:

>>> import jax.numpy as jnp
>>> sig = jnp.sin(jnp.linspace(0, 10, 256))
>>> mel_cepstral_distortion(sig, sig)
0.0

signal_to_noise_ratio(signal: Any, noise: Any) -> Any

Signal-to-Noise Ratio in dB.

SNR = 10 * log10(|signal|^2 / |noise|^2).

Parameters:

Name	Type	Description	Default
signal	Any	Clean signal array.	required
noise	Any	Noise signal array (same shape as signal).	required

Returns:

Type	Description
Any	SNR value in dB. Higher is better.

Examples:

>>> import jax.numpy as jnp
>>> signal = jnp.ones(100)
>>> noise = jnp.ones(100) * 0.01
>>> signal_to_noise_ratio(signal, noise)  # ~40 dB
...