Speaker Independent Voice Conv Eval

speaker-independent-voice-conv-eval

Converting Anyone's Emotion: Towards Speaker-Independent Emotional Voice Conversion — Zhou et al. (2020) (arXiv:2005.07025, 2020)

What this evaluates

Evaluates a model's ability to convert speech emotion (neutral to angry) while preserving speaker identity across both seen and unseen speakers. It measures spectral and prosody conversion quality objectively, and assesses perceived speech quality, emotion similarity, and speaker similarity subjectively.

Datasets

• English emotional speech corpus — total ?; splits: train (-1), test_seen (-1), test_unseen (-1)
• EmoV-DB — total ?; splits: train (-1)
• JL-Corpus — total ?; splits: test_seen (-1), test_unseen (-1)

Metrics

• MCD (primary) — range: other
  • Mel Cepstral Distortion measures the spectral difference between converted and target speech in dB. Lower is better.
• LSD (primary) — range: other
  • Log Spectral Distance evaluates spectral distortion in dB. Lower values indicate better spectral fidelity.
• PCC (primary) — range: [0, 1]
  • Pearson Correlation Coefficient measures the linear correlation between converted and target prosody features. Higher is better.
• MOS — range: other
  • Mean Opinion Score from subjective listening tests, typically on a 1-5 scale, assessing overall speech quality.
• XAB emotion similarity — range: [0, 1]
  • Preference rate from a forced-choice test where listeners pick the sample closest to the reference in emotional expression.
• XAB speaker similarity — range: [0, 1]
  • Preference rate from a forced-choice test where listeners pick the sample closest to the reference in speaker identity.

Input / output format

Input: Raw audio utterances in a neutral emotional state from source speakers.

Output: Synthesized audio utterances converted to an angry emotional state.

Scoring recipe

def evaluate(converted, target, prosody_conv, prosody_target, listeners):
    mcd = compute_mcd(converted, target)
    lsd = compute_lsd(converted, target)
    pcc = pearson(prosody_conv, prosody_target)
    mos = mean([l.rate_quality(converted) for l in listeners])
    xab_em = fraction([l.choose_closest(converted, target, ref, 'emotion') for l in listeners])
    xab_spk = fraction([l.choose_closest(converted, target, ref, 'speaker') for l in listeners])
    return mcd, lsd, pcc, mos, xab_em, xab_spk

Common pitfalls

• Splits are defined by speakers, not utterances; 'seen' vs 'unseen' refers to whether the speaker's data was in the training set.
• Objective metrics (MCD, LSD) are computed on converted vs target, while subjective tests use a reference-based XAB paradigm.
• Emotion conversion is strictly neutral-to-angry; other emotions are not evaluated.

Evidence (verbatim from paper)

We use MCD and LSD for spectrum conversion evaluation, while PCC is used for prosody conversion evaluation. In this section, the proposed VAW-GAN-based EVC framework given in Figure 1 is denoted as CWT-C-VAWGAN.

Citation

@misc{zhou2020converting,
  title={Converting Anyone's Emotion: Towards Speaker-Independent Emotional Voice Conversion},
  author={Zhou et al. (2020)},
  year={2020},
  note={arXiv:2005.07025}
}

• arXiv: 2005.07025