Skip to main content
Supertone offers five TTS models with different trade-offs between quality, latency, language coverage, and configurability. Use this page to choose the model that fits your product.

​
How to choose

The model is selected per-request via the model field. If omitted, the default is sona_speech_1.

​
Model summary

​
Models in detail

​
sona_speech_2

The most natural, highest-quality voice on the platform with broad multilingual coverage. Recommended for narration, audiobooks, character dialogue, and production-quality marketing audio — anywhere quality matters more than latency.
  • Languages (23): en, ko, ja, bg, cs, da, el, es, et, fi, hu, it, nl, pl, pt, ro, ar, de, fr, hi, id, ru, vi
  • Voice settings: all parameters except subharmonic_amplitude_control
  • Extras: include_phonemes (timestamps for lip-sync), normalized_text (pronunciation control)
  • Streaming: not supported

​
sona_speech_2_flash

A lightweight variant of sona_speech_2 optimized for lower latency while keeping the same multilingual coverage. Use it when you care about response time and want acceptable quality — for example, interactive agents or batch generation at scale.
  • Languages (23): same as sona_speech_2
  • Voice settings: pitch_shift, pitch_variance, speed, duration
  • Extras: include_phonemes, normalized_text
  • Streaming: not supported

​
supertonic_api_3

The next-generation successor to supertonic_api_1 with significantly improved speech stability. Trained differently from the open-weights Supertonic 3 release, this API variant inherits the ultra-low latency profile of supertonic_api_1 while delivering far more reliable pronunciation and reduced reading errors. The best default for voice agents, chatbots, and any real-time experience where time-to-first-audio is the top priority.
  • Languages (31): en, ko, ja, ar, bg, cs, da, de, el, es, et, fi, fr, hi, hr, hu, id, it, lt, lv, nl, pl, pt, ro, ru, sk, sl, sv, tr, uk, vi
  • Voice settings: speed only — all other settings are silently ignored
  • Extras: —
  • Streaming: not supported (but per-call latency is so low that streaming is usually unnecessary)

​
supertonic_api_1

The legacy supertonic model. Superseded by supertonic_api_3, which offers broader language coverage and dramatically better speech stability at the same latency profile. Pick supertonic_api_1 only if you have an existing integration pinned to it; new projects should use supertonic_api_3.
  • Languages (5): en, ko, ja, es, pt
  • Voice settings: speed only — all other settings are silently ignored
  • Extras: —
  • Streaming: not supported

​
sona_speech_1

The legacy flagship. It supports the full voice-settings surface and is the only model that currently supports chunked streaming (stream_speech). For most use cases the newer models are a better starting point; pick sona_speech_1 if you specifically need stream_speech output or the full set of fine-tuning parameters (similarity, text_guidance, subharmonic_amplitude_control).
  • Languages (3): en, ko, ja
  • Voice settings: all parameters
  • Extras: include_phonemes
  • Streaming: supported

​
Supported languages

language is required on every TTS request and must be a value supported by both the model and the chosen voice (check the voice’s language array). Pass the language as a lowercase ISO code string:
For multilingual content, fire one request per language rather than mixing languages inside a single text. For Japanese inputs with kanji, numbers, units, or symbols, see Normalized text.

​
Feature support matrix

Voice settings

Reference for every voice-setting parameter and its supported models.

Voices

Find a voice ID that matches your language and style requirements.

​
On-device TTS

Looking to run TTS locally on CPU, with no API call and no network round-trip? Supertone also publishes an open-weights model in the same Supertonic 3 family — Supertonic 3 (99M parameters, ONNX Runtime, OpenRAIL-M license).
Supertonic 3 (open-weights) is a different model from supertonic_api_3. They share the same family name and lineage, but were trained differently and produce different audio. The API model (supertonic_api_3) is what’s exposed by this API; the open-weights model is a separate on-device release. Don’t assume parity in voice quality, supported voices, or behavior.

Supertonic 3 — On-device TTS ↗

99M-parameter open-weights TTS that runs locally on CPU via ONNX Runtime — 31 languages, no GPU, no cloud, no API. A separate model from supertonic_api_3; visit the project site for weights, samples, and SDKs (Python, Node.js, Web, iOS, Android, C++).