Jade 1.1 by ETM-Network: on-device text-to-speech in 23 languages and 9 voices; at the default settings about 4 times real time on a phone CPU and 250 times on an RTX 4090
GitHub: jade Python: pip install etm-tts Licence: OpenRAIL-M

Jade 1.1

A small text-to-speech model that runs on your own device: nine voices, 23 languages, about four times faster than real time on a mid-range phone CPU at the default settings.

An early release. Jade 1.1 is our first public model. Training didn't go entirely to plan, and some of that is still audible (see Limitations). Jade 1.2 follows soon with more voices, steadier long-form reading and higher quality, again with open weights and code.

Willow, English
Silas, English
Juliette, German
Vivienne, Spanish
Elias, Japanese
Willow, English, q4 at 6 steps
Silas, English, q4 at 6 steps

The samples use the default, q8 at 12 steps, except the last two, which use the smaller and faster q4 at 6 steps. Every voice in English and every language are in samples/.

Quick start

pip install etm-tts
from etm_tts import TTS

tts = TTS()  # downloads onnx/quants/q8 once, about 150 MB
tts.speak("Hello! It's nice to finally meet you.").save("hello.wav")
tts.speak("Guten Morgen, wie geht es dir?", voice="Henry", lang="de").save("morgen.wav")

On Apple silicon, pip install "etm-tts[mlx]" runs the MLX files on the GPU. On an NVIDIA GPU, install onnxruntime-gpu in place of onnxruntime and use TTS(device="cuda"). The Android library and the source are at github.com/ETM-Network/jade.

Which variant and how many steps

The default, q8 at 12 steps, is what we recommend, especially for long-form reading such as audiobooks; 12 to 16 steps are steadier still. Use q4 or q3, and fewer steps, only when size or speed requires it. At low precision and few steps, a word can now and then jump in pitch. That is rare in short lines but noticeable over long passages, most of all with q3 at the lowest step counts.

tts = TTS(variant="q4")                     # 81 MB instead of 150 MB
tts.speak("Short reply.", steps=6).save("reply.wav")

Voices and languages

Female: Juliette (default), Vivienne, Sienna, Molly, Willow. Male: Henry, Elias, Silas, Gabriel. Every voice speaks every language.

English, German, French, Spanish, Italian, Portuguese, Dutch, Danish, Polish, Czech, Slovak, Slovenian, Croatian, Hungarian, Romanian, Bulgarian, Russian, Greek, Latvian, Lithuanian, Estonian, Hindi and Japanese. Pass the language; the model doesn't detect it. English and German read numbers, dates and abbreviations; in the other languages, write numbers as words.

Files

folder size
onnx/quants/q8 150 MB default: 8 bits, decoder in fp16
onnx/quants/q4 81 MB step model about 4 bits, text side 8 bits, decoder about 4.3 bits
onnx/quants/q3 65 MB step model about 3.5 bits, text side about 4.2 bits
onnx/raw 437 MB float32
mlx/... same tiers the same models for MLX, 0.5 to 1 MB smaller
pytorch 437 MB float32 safetensors, for research; no inference code

The quantized files use standard formats (ONNX MatMulNBits, MLX quantized layers) with values tuned per layer (learned rounding and per-layer bit widths). The ONNX files need ONNX Runtime 1.28 or newer.

default (q8, 12 steps) phone CPU desktop CPU RTX 4090 Apple GPU
real-time factor about 4× about 7× about 250× about 240×

Phone: Snapdragon 778G+, 4 threads. Desktop: Intel i9-13900KF, 4 threads (8.7× with 8). RTX 4090: TTS(device="cuda") with onnxruntime-gpu, about 340× on long texts. Apple: M5 Ultra, MLX. Fewer steps are faster: at 6 steps the phone runs about 6× real time.

How it works

How it works: text, encoder, a 12-block diffusion transformer run in 12 steps by default, the Supertonic 3 decoder, and 44.1 kHz audio

A 12-block diffusion transformer generates the latents of the Supertonic 3 speech autoencoder from plain characters, with no phonemizer. Supertonic 3's decoder turns them into 44.1 kHz audio. The released model is distilled for few-step sampling with guidance built in; it runs at 2 to 32 steps, 12 by default. It has 109 M parameters, of which 84 M are ours and 25 M are the Supertonic 3 decoder.

Training, distillation and the full evaluation are in DETAILS.md.

This model is not made, endorsed or reviewed by Supertone Inc.

Evaluation

181 held-out lines, 8 per language, never seen in training, rendered from these files by etm-tts: the default, and every variant at 6 steps.

q8, 12 steps (default) raw, 6 q8, 6 q4, 6 q3, 6
Character error rate (Whisper read-back) 1.1 % 1.1 % 1.1 % 1.0 % 1.0 %
Lines read back without an error 147 144 144 147 146
Similarity to the intended voice (WavLM) 0.818 0.811 0.811 0.803 0.800
Pitch match with raw at 6 steps 0.978 0.962 0.924

At 6 steps one English line is left out of the error rate: Whisper stopped transcribing at a 1.5 s pause, though the line is spoken in full. With it, the 6-step rates are 1.6, 1.6, 1.5 and 1.0 %. At 12 steps Whisper reads that line through, and the default's 1.1 % covers all 181 lines.

On these short lines the quantized files read as accurately as full precision; the sets are small, so differences of a few tenths of a percent are noise. In our blind A/B listening on short clips, q4 and raw were indistinguishable at 6 steps. Long listening is harsher: over minutes of audio, q4 and q3 at few steps show more pitch errors than q8 at 12 steps, which is why q8 at 12 steps is the default. Per-language results are in DETAILS.md. We haven't run a formal listening test (MOS) or compared against other models.

Limitations

  • Short phrases. On two or three words, several voices can sound alike.
  • Pitch errors at low precision. At q4 and q3 with few steps, a word can jump in pitch now and then. It's rare in short lines but noticeable over long passages; see Which variant and how many steps.
  • Reading errors. No digits outside English and German. Japanese is read from kanji as written, so rare kanji and place names can be misread. Estonian and Hindi have the highest read-back error.
  • Formal style. Much of the training speech for Greek, Bulgarian, Croatian, Slovenian, Estonian, Romanian, Slovak, Czech, Hungarian and Polish is parliament speech, often with machine transcripts. Expect a formal pace and an occasional mispronounced word.
  • One reading style. Neutral, like an audiobook narrator, with no emotion or emphasis control.

Intended use

Reading text aloud on the device: narration, accessibility, apps, prototypes and research on small TTS models. Out of scope: anything the licence restricts, imitating a real person, presenting the output as a human recording, and safety-critical announcements.

Licence

The model is released under the BigScience OpenRAIL-M licence (LICENSE), which allows commercial use within its use-based restrictions. It includes the Supertonic 3 decoder by Supertone Inc. under the same licence; NOTICE lists our changes to it. The runtime code in etm-tts is Apache 2.0.

Credits

  • Voices: modelled on nine LibriVox volunteer readers, credited in ATTRIBUTION.md. The voice names are invented and refer to no real person.
  • Decoder and latent space: Supertonic 3 by Supertone Inc., OpenRAIL-M. "Copyright (c) 2026 Supertone Inc."
  • Training data: see ATTRIBUTION.md, among others HiFi-TTS and CML-TTS (CC BY 4.0).
  • Architecture: follows EMA Lightning by Canberk Aslan (Apache 2.0), scaled up and trained on Supertonic 3's latents.

Citation

@misc{etm2026jade,
  title  = {Jade 1.1: on-device multilingual text-to-speech},
  author = {{ETM-Network}},
  year   = {2026},
  url    = {https://huggingface.co/ETM-Network/jade-1.1}
}

ETM-Network is a brand of Mensch, Nina! UG (haftungsbeschränkt), Geretsried, Germany.

Downloads last month

-

Downloads are not tracked for this model. How to track
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support