aoiandroid/supertonic-3-litert

Supertonic-3 — LiteRT (.tflite, INT4) + ONNX vector_estimator

LiteRT / TensorFlow Lite conversion of Supertone/supertonic-3, a 99M-parameter multilingual TTS model. 3 of the 4 components convert cleanly to true INT4 weight-only quantization via Google's ai-edge-quantizer and run on the `ai_edge_litert` runtime. vector_estimator (the diffusion denoiser) is kept as ONNX — its rotary multi-head attention defeats onnx2tf's NCW↔NHWC shape inference, and litert_torch.convert deadlocks in MLIR lowering when fed the model with loaded weights. The ONNX VE is shipped in both fp32 (vector_estimator.onnx) and INT8 dynamic quantization (vector_estimator_int8.onnx, 65 MB) — INT8 is the recommended config.

Configurations

Quickstart

pip install ai-edge-litert onnxruntime soundfile numpy supertonic
git clone https://huggingface.co/Reza2kn/supertonic-3-litert
cd supertonic-3-litert

# Recommended INT4 + INT8 VE config (default)
python inference.py --text "Hello, world." --voice F1 --out hello.wav

# Long prompt — use --auto-pad for full content rendering
python inference.py \
  --text "<longer prompt>" \
  --voice F5 --auto-pad --out long.wav

# Explicit FP32 baseline (uses fp32 vector_estimator.onnx)
python inference.py --text "Hello" --dp-quant fp32 --te-quant fp32 --voc-quant fp32 --ve-fp32

10 voice styles ship in voice_styles/: F1–F5 (female), M1–M5 (male). 31 languages supported via unicode_indexer.json.

⚠️ Known limitation: rushed pacing on long prompts (vs CoreML build)

The supertonic-3 model has a soft content cap per utterance (~13.7 s of speech for the included longenF5 prompt). The LiteRT pipeline runs vector_estimator at native input shapes via ONNX Runtime, which respects the model's hard limit and truncates long prompts.

The CoreML build of this same model benefits from an accidental "bucket-leak" in the CoreML conversion (padded latent positions leak through ConvNeXt's dilated convolutions), which extends content by ~3 s and gives more natural pacing. This extension does not exist in LiteRT — we tested padding the ONNX VE inputs to the same bucket: 13.00s → 13.05s (essentially no extension).

In practice:

• —Short prompts (under ~10 s of speech): fine.
• —Long prompts (over ~13 s): LiteRT will sound rushed and may truncate the last words. Use the CoreML build for those if you're on Apple.

--auto-pad is still useful — it appends a filler sentence that the model partially renders, then trims at the silence gap. It recovers some content but cannot match CoreML's bucket-leak extension.

Conversion pipeline

Supertone/supertonic-3 (ONNX)
  -> onnxsim.simplify (T=L=320)
  -> fuse_gelu (Div/Erf/Add/Mul/Mul -> ONNX Gelu opset 20)
  -> onnx2tf -kt -coion (TF SavedModel)
  -> tf.lite.TFLiteConverter (fp32 .tflite)
  -> ai-edge-quantizer weight_only_wi4_afp32() (true INT4)
  -> ai_edge_litert.Interpreter at runtime

vector_estimator:
  -> onnxruntime.quantization.quantize_dynamic(QInt8, per_channel=True)
     (4× compression, kept ONNX because onnx2tf/litert_torch both
      fail on the rotary multi-head attention)

The GELU fuse is the key unlock for INT4 LiteRT. Without it, onnx2tf emits FlexErf ops which disqualify the model from ai_edge_litert (the runtime that supports INT4). Replacing the Erf-based GELU expansion with a single ONNX Gelu op (opset 20) keeps the model in pure-TFLite ops and unblocks INT4 inference.

vector_estimator is kept as ONNX because onnx2tf's transpose optimization breaks rotary attention masking, and litert_torch.convert deadlocks on its loaded weights. INT8 dynamic quantization via onnxruntime.quantization.quantize_dynamic works cleanly on Conv + MatMul ops and gives 4× compression with audio-identical output to fp32.

License

OpenRAIL — same as the original Supertone/supertonic-3.

Credits

• —Original model: Supertone/supertonic-3
• —LiteRT conversion + auto-pad workflow: this repo
• —Companion CoreML build: Reza2kn/supertonic-3-coreml
• —Quantization: `ai-edge-quantizer`, onnxruntime.quantization
• —Runtime: `ai_edge_litert`, onnxruntime