Maximilian228/Geode-Supertonic-3

Supertonic 3 (int8) for Geode

The voice Geode reads its answers with when "Geode's voice" is chosen in its settings: a copy of Supertonic 3 by Supertone Inc., 31 languages, quantized to int8 ONNX by the sherpa-onnx project, which also runs it in the app. The app downloads these files only when its user picks this voice; they stay on the phone's storage and are opened in memory (about 300 MB) only while an answer is read.

File	What it is
text_encoder.int8.onnx, duration_predictor.int8.onnx, vector_estimator.int8.onnx, vocoder.int8.onnx	the model
tts.json, unicode_indexer.bin	its configuration and text index
voice.bin	its ten voice styles: speaker ids 0-4 the female voices F1-F5, 5-9 the male M1-M5
LICENSE	the model's licence: BigScience Open RAIL-M (Supertone/supertonic-3)
LICENSE-supertonic-code	the MIT licence of Supertone's code that came with the files

Sources

• Model: Supertone/supertonic-3 at revision 3cadd1ee6394adea1bd021217a0e650ede09a323 (its LICENSE copied here unchanged).
• int8 conversion: sherpa-onnx's sherpa-onnx-supertonic-3-tts-int8-2026-05-11 (GitHub release tts-models, mirrored at csukuangfj2/sherpa-onnx-supertonic-3-tts-int8-2026-05-11 revision cca5a0e6c96e1d2c720986bf7e75fcc81dee3ae4); the files here are byte for byte those (sha256 checked).

Licence and use restrictions

The model is licensed under the BigScience Open RAIL-M License (LICENSE). Its use-based restrictions (Attachment A) apply to everyone who uses these files, through Geode or otherwise: among them, not to use the model in any way that violates applicable law, to harm minors, to generate or spread verifiably false information meant to harm others, to present what it generates without disclosing that it is machine-generated, to defame, disparage or harass, or to impersonate a person without their consent (deepfakes). Read the full text before use. Nothing here changes the licence or adds terms of its own; Geode only redistributes the files unchanged.

Speed (measured)

On a Redmi Note 9 Pro (Snapdragon 720G, two Cortex-A76 cores), sherpa-onnx 2.0.1 with 2 threads: a 7.6 s Ukrainian sentence in 3.6 s (real-time factor 0.48; about the same for Russian and English), the model opened in ~1.7 s, 300 MB resident while open. Four threads were no faster (the other cores are the little A55s).