[FEATURE] Add split-model export for MeloTTS and TTS support for RKNN3/RK1828 NPU

Describe the feature

I would like to propose experimental support for running MeloTTS/VITS on a discrete Rockchip RK1828 PCIe NPU using RKNN3.

The initial target model is:

• vits-melo-tts-zh_en
• Chinese + English
• 1 speaker
• 44.1 kHz

Our prototype splits the original dynamic VITS graph into:

1. Step1, fixed T=96: text encoder and deterministic duration predictor.
2. CPU dynamic path: duration rounding, monotonic alignment, prior expansion, latent sampling, padding and cropping.
3. Step2, fixed W=384: inverse flow and waveform decoder.

Both Step1 and Step2 are converted to RKNN3 FP16 models with do_quantization=False.

The proposed upstream contribution would be divided into:

1. Split ONNX export and RKNN3 conversion tools.
2. ONNX/RKNN numerical parity tools.
3. A generic RKNN3 TTS runtime integrated with sherpa-onnx OfflineTts.
4. Reproducible model packaging, benchmarks and documentation.

The contribution would not contain product-specific server protocols, ALSA configuration, device paths, systemd files, or Rockchip SDK binaries.

Motivation / Use Case

sherpa-onnx already provides the vits-melo-tts-zh_en model and an RKNN backend for several ASR, VAD and KWS models.

However, the existing RKNN implementation is based mainly on RKNN Toolkit2/RKNPU2 and currently does not provide a VITS/MeloTTS runtime for a discrete RK1828 device using RKNN3.

Supporting this configuration would make high-quality Chinese and English TTS available on embedded systems where CPU-only VITS inference does not provide sufficient latency or power efficiency.

It may also provide a reusable implementation for other VITS-family models and future RKNN3 devices.

Additional context

Environment used by the current prototype:

• Host: RK3588
• Accelerator: discrete RK1828 PCIe NPU
• RKNN3 Toolkit/runtime: 1.0.4
• RKNN conversion target: rk1820
• Model: vits-melo-tts-zh_en
• Precision: FP16, without W4A16 quantization

Current experimental board results:

• PyTorch vs split ONNX validation: passed
• Two rounds of 20 synthesis requests: 40/40 completed
• Aggregate synthesis RTF: approximately 0.416–0.417
• First-audio latency: approximately 0.93–1.03 seconds
• Single-character RTF: 1.577–1.760 because Step2 always executes the fixed W384 graph

Known numerical gaps:

• Step1 logw RMSE: 0.005756
• Predicted duration changed from 289 to 291 frames
• Step2 waveform cosine similarity: 0.991028
• Step2 waveform RMSE: 0.009526

This is currently an experimental prototype. We are not claiming production readiness or full numerical parity yet.

Before preparing the implementation PRs, I would appreciate guidance from the maintainers on the following questions:

1. Should RKNN3 use a separate build option and backend directory from the existing RKNN Toolkit2 implementation?
2. Would you prefer the split-export and numerical-parity tools as the first standalone PR?
3. Is a converter that requires a user-supplied RKNN3 SDK acceptable?

References:

• sherpa-onnx MeloTTS scripts: https://github.com/k2-fsa/sherpa-onnx/tree/master/scripts/melo-tts
• vits-melo-tts-zh_en model: https://k2-fsa.github.io/sherpa/onnx/tts/pretrained_models/vits.html#vits-melo-tts-zh-en
• Existing sherpa-onnx RKNN backend: https://github.com/k2-fsa/sherpa-onnx/tree/master/sherpa-onnx/csrc/rknn