Metrics and Evaluation API

Hyperion separates metric primitives from evaluators. Use NumPy functions for in-memory score arrays; use evaluators or hyperion-eval-verification-metrics when scores and keys are stored in tables/files.

Verification metrics

The primary verification contract is a TrialKey plus an aligned TrialScores object. The evaluator extracts target/non-target scores, then computes EER and min/actual DCF for the requested target priors.

hyperion.np.metrics.compute_eer(tar: ndarray, non: ndarray) → float[source]

Computes equal error rate.

Parameters:

  • tar – Scores of target trials.

  • non – Scores of non-target trials.

Returns:

EER

hyperion.np.metrics.compute_min_dcf(tar: ndarray, non: ndarray, prior: float | ndarray | Sequence[float], normalize: bool = True) → Tuple[float | ndarray, float | ndarray, float | ndarray][source]

Computes minimum DCF.

min_DCF = min_t prior * p_miss(t) + (1 - prior) * p_fa(t), where t is the decision threshold.

Parameters:

  • tar – Target scores.

  • non – Non-target scores.

  • prior – Target prior or vector of target priors.

  • normalize – if true, return normalized DCF, else unnormalized.

Returns:

Vector Minimum DCF for each prior. Vector of P_miss corresponding to each min DCF. Vector of P_fa corresponding to each min DCF.

Use the command-line workflow in Extract, Score, and Evaluate X-Vectors for CSV/HDF5 result files and DET/DCF plots. EER is threshold-independent; DCF depends on the declared prior and costs, so only compare DCF values computed under the same policy.

Other evaluator families

The anonymization/voice-conversion workflows that feed these evaluators are experimental. TPM-backed quality and VoxProfile evaluators may require their corresponding optional external packages.

Torch metrics

hyperion.torch.metrics contains training-loop metrics such as categorical accuracy. These are distinct from speaker-verification EER/DCF evaluation and should not be used as a substitute for trial-based evaluation.

See also