Etikedoj kiujn la elektita modelo komprenas - klaku por meti unu en vian tekston kie ĝi okazas:
This model reads plain text, so inline tags are ignored. For tag-based emotion, switch to an expressive model like Orpheus or Bark.
Natural- language delivery hints. Honored by Qwen3- TTS today; other expression models coming soon.
Tiu modelo ne subtenas stilajn instrukciojn — ŝanĝu al modelo kiu subtenas ilin (ekz. Qwen3-TTS).
Difini proprajn elparolojn (vorto = elparolo):
-12
+12
Kontrolas esprimpovon (0 = neŭtrala, 1 = tre esprima)
Klasigilo- libera gvidpezo (pli alta = pli prompt- sekvanta)
Priskribi la voĉstilon en natura lingvo (Parler uzas tion anstataŭ antaŭdifinitaj voĉoj)
Dia dialoga formato: Uzu [S1] kaj [S2] por marki malsamajn parolantojn. Ekzemplo: [S1] Saluton! [S2] Saluton, kiel vi fartas?
FreyaTTS-small is a 183-million-parameter model built for one language and built well. It is a non-autoregressive conditional flow-matching diffusion transformer that reads Turkish directly at the character level — 92 symbols, no phonemizer and no grapheme-to-phoneme stage, which removes a whole class of mispronunciation that pronunciation dictionaries introduce. It generates in a frozen AudioVAE2 latent space and decodes to 48 kHz mono, more than double the sample rate of the piper Turkish voice, so the output carries treble detail that a 22 kHz model simply cannot represent. On the Freya-TR-Eval benchmark it reaches 8.0% word error rate, placing it ahead of both XTTS-v2 and F5-TTS among open sub-billion-parameter Turkish systems, and it runs fast enough for real-time use at roughly a tenth of real time.
Plej bona por: Turkish narration, voice agents, and any Turkish audio that needs high sample-rate output
It was trained from scratch on Turkish speech alone rather than adapted from a multilingual model. Specialising lets a 183M-parameter model compete with far larger multilingual systems on Turkish, but it means the model has no ability to read other languages — requests in another language are rejected rather than mispronounced.
Most open TTS models emit 22.05 kHz or 24 kHz, which caps reproducible audio at around 11-12 kHz and audibly dulls sibilants. FreyaTTS decodes to 48 kHz, the standard sample rate for video and broadcast, so its output drops into a production timeline without upsampling.
No. FreyaTTS has a single fixed speaker and does not support cloning. For a cloned Turkish voice use one of our zero-shot cloning models instead.
Many TTS systems first convert text into phonemes using a pronunciation dictionary, and anything missing from that dictionary — new words, names, loanwords — gets guessed. FreyaTTS reads the characters themselves, so Turkish spelling, which is highly regular, maps to sound without that lossy middle step.
