Fast Vgan Vc Eval

fast-vgan-vc-eval

Fast-VGAN: Lightweight Voice Conversion with Explicit Control of F0 and Duration Parameters — Abrassart et al. (2025) (arXiv:2507.04817, 2025)

What this evaluates

Evaluates a GAN-based voice conversion model's ability to transfer speaker timbre while explicitly controlling prosodic features like F0 and duration. It tests static prosodic manipulation, dynamic expressive transfer without expressive training data, and real-time inference efficiency.

Datasets

• VCTK Corpus — total ?; splits: test (16)
• Expresso Dataset — total ?; splits: train (-1), test (-1)

Metrics

• timbre transfer quality (primary) — range: other
  • Not explicitly defined in the provided excerpt; standard subjective and objective measures are implied to benchmark speaker similarity and conversion fidelity.
• naturalness — range: other
  • Not explicitly defined in the provided excerpt; implied subjective or objective assessment of speech quality after prosodic manipulation.
• intelligibility — range: other
  • Not explicitly defined in the provided excerpt; implied assessment of speech clarity and word recognition after prosodic manipulation.
• speaker similarity — range: other
  • Not explicitly defined in the provided excerpt; implied objective or subjective measure of how closely the converted voice matches the target speaker.

Input / output format

Input: Source speech waveform, target speaker ID, phoneme sequence, F0 contour, intensity curve, and optional prosodic control parameters.

Output: Converted speech waveform at 16 kHz.

Scoring recipe

# 1. Voice Identity Conversion Benchmark
for src, tgt_id in test_set:
    gen = model(src, tgt_id)
    # Evaluate timbre transfer quality, naturalness, intelligibility

# 2. Prosodic Control Evaluation
for factor in pitch_shifts, duration_stretches:
    gen = model(src, tgt_id, f0_factor=factor, dur_factor=factor)
    # Evaluate naturalness and intelligibility

# 3. Expressive Transfer
for expr_src in expresso_test:
    gen = model(neutral_src, expr_src.prosody)
    # Assess expressive synthesis capability

Common pitfalls

• Test set is extremely small (only 2 utterances per speaker across 8 speakers)
• Model is trained only on neutral speech but evaluated on expressive prosody transfer
• All outputs are down-sampled to 16 kHz for fair baseline comparison, which may affect high-frequency quality metrics
• Inference speed varies significantly between CPU decoder and vocoder stages

Evidence (verbatim from paper)

We first compare our method to two representative voice conversion models from the literature, ControlVC and HiFi-VC, to benchmark the timbre transfer quality in a standard many-to-many conversion setting. We evaluate to what extent those transformations remain natural and intelligible with respect to the transformation factor of the speech parameters.

Citation

@misc{abrassart2025fastvgan,
  title={Fast-VGAN: Lightweight Voice Conversion with Explicit Control of F0 and Duration Parameters},
  author={Abrassart et al. (2025)},
  year={2025},
  note={arXiv:2507.04817}
}

• arXiv: 2507.04817