Higgs TTS 3

This first-party Soniqo page documents Higgs TTS 3 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	Higgs TTS 3
Role	Conversational TTS and zero-shot cloning with inline emotion, style, and SFX tags
Backend	MLX bf16 backbone with float32 codec
Output	24 kHz mono waveform, 25 frames per second
Languages	100+ languages
License	Boson Higgs TTS 3 Research and Non-Commercial License
Status	Ready through speech speak --engine higgs and the HiggsTTS Swift product
Source	Boson AI higgs-tts-3-4b
Swift product	`HiggsTTS`
CLI / runtime	`speech speak --engine higgs`

Use

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

# Expressive cloned speech with an inline emotion tag.
.build/release/speech speak "<|emotion:elation|> We did it — the demo works!" \
  --engine higgs \
  --voice-sample reference.wav \
  --higgs-ref-text "Text spoken in the reference clip." \
  -o higgs.wav

Model Links

Implementation Notes

RTF 0.78 on an M5 Pro for cloned and plain synthesis alike: sampling stays on the GPU with the delay ramp masked on-device, and asyncEval overlaps decode with the end-of-codes state machine.
Cloning gates: English lexical overlap 1.00 and Mandarin CER 0.000 on ASR roundtrips; the language model and codec are parity-validated against the upstream implementations in both directions.
Control tags are ordinary tokens inside the target text: <|emotion:...|>, <|style:whispering|>, <|sfx:laughter|>, <|prosody:pause|> and friends.
Encoding a 12 s reference costs about 0.6 s once and the codes are reusable across generations; omit --voice-sample for reference-free synthesis in a stock voice.
The CLI defaults to temperature 0.8 because the upstream 1.0 wobbles on single sentences; expect about 8.6 GB peak RSS for the 4B backbone.