KitsuMate/neutts-2e-onnx

NeuTTS-2E ONNX

ONNX conversion of Neuphonic NeuTTS-2E, prepared for KitsuMate ONNX TTS.

The model supports four fixed speakers (emily, paul, sophie, and steven) and seven emotions (angry, disgusted, fearful, happy, neutral, sad, and surprised). It generates English, 24 kHz mono speech.

Files

• onnx/backbone_fp16.onnx — backbone with FP16 weight storage, FP32 computation, and WebGPU-compatible fused attention. It supports batch size 1 and a 2048-token context.
• onnx/codec_decoder.onnx — decoder prepared for WebGPU and CPU execution. It accepts int32 speech codes shaped [1, 1, N] and returns float32 audio.
• tokenizer.json and neutts.json — tokenizer and fixed-speaker reference data required at runtime.

The graphs were validated with ONNX Runtime 1.30.0 on CPU and WebGPU. Native WebGPU testing used an RTX 3070 Laptop GPU; browser execution has not been tested.

Speech generation uses about 50 tokens per second of audio. Literal digits and times are unreliable in the upstream checkpoint, so spell them out in words. The decoder can produce raw waveform peaks outside [-1, 1].

This conversion does not apply the optional Perth watermark. Custom voice cloning and the NeuTTS Air/Nano variants are not included.

License

The backbone is distributed under the NeuTTS Open License v1.0, including its commercial-use restriction and USD 5 million annual-revenue threshold. See LICENSE for the exact terms. The codec and reference assets are Apache-2.0; see CODEC_LICENSE and ATTRIBUTION.md.