Voice Cloning PoC

2-day research spike comparing voice-cloning routes — Azure Personal Voice, XTTS-v2, F5-TTS, and Qwen3-TTS — on a fixed evaluation grid. Key finding: fidelity-to-speaker and naturalness-in-target-language are in structural tension; automated metrics (WER/UTMOS) are blind to L1 accent transfer.

Benchmark spike comparing four voice-cloning approaches on a fixed 8-prompt evaluation grid using a real speaker reference.

Tracks evaluated: XTTS-v2 (autoregressive), F5-TTS (flow-matching, replaces OpenVoice v2), Qwen3-TTS-Base (LLM-based), and Azure Personal Voice. Each track generated 64 WAVs scored on SECS (speaker similarity), WER (intelligibility), and UTMOS (naturalness).

Key finding: Qwen3-TTS produced the highest speaker similarity but transferred the speaker’s L1 accent into L2 output — a structural tradeoff invisible to automated metrics. The evaluation pipeline separates the two axes explicitly.