Vyvo/dacvae-tts-tr-demo

Turkish zero-shot voice-cloning TTS demo for VoiceHub/dacvae-tts-tr-w512 (66.5M flow-matching DiT in frozen Meta DACVAE latent space, trained with dacvae-tts).

• —Synthesis: 3–15 s reference (upload or microphone; transcript filled in by Whisper and editable) + any length of text. Long text is split into sentences and generated in one batched GPU call; numbers, dates, clock times, currencies, units, abbreviations, acronyms and symbols are rewritten into spoken Turkish.
• —Speaking rate: Otomatik follows the prompt but slows down prompts faster than 17 characters/s (rushed podcast prompts were the main source of errors on Freya-TR-Eval).
• —Best-of-N: N candidates per sentence in one batch; Whisper picks the one it transcribes best.
• —Research tools: any checkpoint from the Hub (log in for private repos — your own read permission is used, the Space holds no token), A/B comparison of 2–4 checkpoints, batch test on sentence lists (custom sentences or Freya-TR-Eval) with WER/CER, speaker similarity and DNSMOS, downloadable as a zip.
• —ZeroGPU: models are loaded once at start-up; GPU time per request is a few seconds.