NTNU · Faculty of Information Technology and Electrical Engineering · Department of Computer Science · 2026
Benchmarking Deployed Real-Time Voice Conversion Tools
A unified, reproducible quality comparison of open-source and commercial voice changers.
This page supplements the thesis. It reports the benchmark’s measured results and lets you listen to the converted audio they were computed from. Six deployed voice changers, in seven configurations, were each run over the same 138 utterances on one machine and scored on intelligibility, naturalness, and speaker identity. The audio below is a curated subset of that benchmark.
Results
Mean over all 138 converted utterances per configuration. Arrows give the better direction; the best value in each quality column is marked in green. “—” means the metric is not available for that tool.
WER / CER word / character error rate (intelligibility). UTMOS, DNSMOS predicted naturalness, 1–5. SECS speaker-embedding cosine similarity: src = to the source voice (lower = identity moved away), cons = output consistency, tgt = to a clean target recording (the canonical identity metric, computable only for Seed-VC). The two reference-free SECS proxies are best read together with the quality metrics, not alone.
Resource usage
Steady-state usage during conversion, read from the operating system. Indicative, not precise.
Audio samples
The same source clip converted by every tool. Each tool renders one female and one male target voice; the two tables below correspond to those two target genders. Seed-VC is conditioned on a reference clip — its target voices are: