F5-TTS

This first-party Soniqo page documents F5-TTS from the local speech-swift / speech-core implementation. Hugging Face bundles are linked below after the integration notes.

Internal Page First

Landing cards and docs menus now point here first; source model and bundle links remain available from this page.

At a Glance

Model	F5-TTS
Role	Zero-shot voice cloning from a short reference plus transcript
Backend	MLX fp16
Output	24 kHz mono waveform
Languages	English and Mandarin, mixed EN/ZH supported
License	CC-BY-NC-4.0 weights; non-commercial bundle
Status	Ready through speech speak --engine f5 and the F5TTS Swift product
Source	SWivid F5-TTS
Swift product	`F5TTS`
CLI / runtime	`speech speak --engine f5`

Use

The snippet below mirrors the current speech-swift API or command exposed by the repo.

# Clone a voice from a ~10 s reference clip plus its transcript.
.build/release/speech speak "Hello from my cloned voice." \
  --engine f5 \
  --voice-sample reference.wav \
  --f5-reference-text "Text spoken in the reference clip." \
  -o cloned.wav

Model Links

Implementation Notes

16 flow steps are the default (RTF 0.57 on an M5 Pro, lower is faster); --f5-steps 32 for maximum fidelity, 12 runs at RTF 0.43 — ASR roundtrips are word-identical across all three.
--f5-reference-text is required: cloning conditions on the reference transcript, and quality tracks how cleanly the reference clip ends on a sentence boundary.
Mandarin runs through a bundled pinyin lexicon with tone sandhi baked in at export time (99.2% token-exact against the upstream rjieba/pypinyin frontend), so Swift needs no external tokenizer.
Deterministic per --f5-seed; --f5-speed scales pacing, CFG strength and sway sampling are exposed as flags; about 0.8 GB peak RSS makes it the smallest cloning engine in the package.