speecht5-voxpopuli-nl

Fine-tuned microsoft/speecht5_tts on the VoxPopuli Dutch subset (7433 train / 826 test). eval_loss 0.457. For inference also load microsoft/speecht5_hifigan as vocoder and supply a speaker embedding via SpeechT5ForTextToSpeech.generate_speech.

Metrics

• training loss converged

Usage

# text-to-speech