MeloTTS-ZH: Optimized for Qualcomm Devices
MeloTTS is a high-quality multi-lingual text-to-speech library for English, Chinese and Spanish language.
This is based on the implementation of MeloTTS-ZH found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.
Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.
Getting Started
There are two ways to deploy this model on your device:
Option 1: Download Pre-Exported Models
Below are pre-exported model assets ready for deployment.
| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X2 Elite | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X Elite | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS8550 (Proxy) | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS9075 | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X2 Elite | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X Elite | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8550 (Proxy) | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8775P | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA7255P | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8295P | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS9075 | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8450 (Proxy) | QAIRT 2.43 | Download |
For more device-specific assets and performance metrics, visit MeloTTS-ZH on Qualcomm® AI Hub.
Option 2: Export with Custom Configurations
Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:
- Custom weights (e.g., fine-tuned checkpoints)
- Custom input shapes
- Target device and runtime configurations
This option is ideal if you need to customize the model beyond the default configuration provided here.
See our repository for MeloTTS-ZH on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.audio_generation
Model Stats:
- Model checkpoint: myshell-ai/MeloTTS-Chinese
- Max decoded sequence length: 512 tokens
- Number of parameters (encoder): 8.34M
- Model size (encoder) (float): 31.9 MB
- Number of parameters (flow): 20.1M
- Model size (flow) (float): 76.9 MB
- Number of parameters (decoder): 14.5M
- Model size (decoder) (float): 55.5 MB
- Number of parameters (bert_wrapper): 152M
- Model size (bert_wrapper) (float): 581 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| bert_wrapper | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 2.71 ms | 0 - 10 MB | NPU |
| bert_wrapper | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X2 Elite | 3.073 ms | 293 - 293 MB | NPU |
| bert_wrapper | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X Elite | 7.338 ms | 293 - 293 MB | NPU |
| bert_wrapper | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 4.737 ms | 0 - 10 MB | NPU |
| bert_wrapper | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 6.784 ms | 1 - 4 MB | NPU |
| bert_wrapper | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS9075 | 8.894 ms | 1 - 4 MB | NPU |
| bert_wrapper | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 3.468 ms | 0 - 11 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 2.728 ms | 0 - 9 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X2 Elite | 3.376 ms | 0 - 0 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X Elite | 7.344 ms | 0 - 0 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 4.78 ms | 0 - 7 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8275 (Proxy) | 30.061 ms | 0 - 8 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 6.78 ms | 0 - 1 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8775P | 9.121 ms | 0 - 9 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS9075 | 8.857 ms | 2 - 4 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8450 (Proxy) | 10.259 ms | 0 - 9 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA7255P | 30.061 ms | 0 - 8 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8295P | 11.297 ms | 0 - 5 MB | NPU |
| bert_wrapper | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 3.467 ms | 0 - 8 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 42.209 ms | 0 - 10 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X2 Elite | 38.578 ms | 20 - 20 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X Elite | 76.531 ms | 18 - 18 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 57.739 ms | 1 - 7 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 77.89 ms | 6 - 27 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS9075 | 76.416 ms | 0 - 3 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 46.943 ms | 1 - 7 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 41.169 ms | 0 - 10 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X2 Elite | 38.809 ms | 0 - 0 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X Elite | 77.005 ms | 0 - 0 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 56.773 ms | 0 - 7 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8275 (Proxy) | 127.513 ms | 0 - 9 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 78.388 ms | 1 - 3 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8775P | 330.323 ms | 0 - 8 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS9075 | 77.193 ms | 0 - 2 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8450 (Proxy) | 106.049 ms | 1 - 10 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA7255P | 127.513 ms | 0 - 9 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8295P | 93.005 ms | 0 - 5 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 47.347 ms | 0 - 13 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 17.476 ms | 4 - 14 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X2 Elite | 19.545 ms | 14 - 14 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X Elite | 32.856 ms | 13 - 13 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 24.767 ms | 4 - 12 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 33.332 ms | 0 - 17 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS9075 | 39.63 ms | 4 - 6 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 20.552 ms | 3 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 17.559 ms | 4 - 13 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X2 Elite | 19.88 ms | 4 - 4 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X Elite | 32.812 ms | 4 - 4 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 25.022 ms | 4 - 11 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8275 (Proxy) | 77.602 ms | 2 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 33.406 ms | 4 - 5 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8775P | 39.263 ms | 2 - 11 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS9075 | 39.21 ms | 4 - 9 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8450 (Proxy) | 47.505 ms | 4 - 12 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA7255P | 77.602 ms | 2 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8295P | 44.065 ms | 2 - 7 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 20.819 ms | 2 - 10 MB | NPU |
| flow | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 90.668 ms | 6 - 16 MB | NPU |
| flow | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X2 Elite | 84.275 ms | 31 - 31 MB | NPU |
| flow | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® X Elite | 147.768 ms | 30 - 30 MB | NPU |
| flow | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 120.552 ms | 6 - 13 MB | NPU |
| flow | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 159.257 ms | 0 - 37 MB | NPU |
| flow | PRECOMPILED_QNN_ONNX | mixed_with_float | Qualcomm® QCS9075 | 149.956 ms | 4 - 10 MB | NPU |
| flow | PRECOMPILED_QNN_ONNX | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 103.125 ms | 3 - 10 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite Gen 5 Mobile | 91.112 ms | 4 - 13 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X2 Elite | 85.084 ms | 4 - 4 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® X Elite | 147.836 ms | 4 - 4 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Gen 3 Mobile | 119.018 ms | 4 - 11 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8275 (Proxy) | 279.016 ms | 1 - 10 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8550 (Proxy) | 155.047 ms | 5 - 6 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8775P | 150.731 ms | 1 - 10 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS9075 | 149.511 ms | 4 - 10 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® QCS8450 (Proxy) | 175.631 ms | 4 - 13 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA7255P | 279.016 ms | 1 - 10 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Qualcomm® SA8295P | 188.834 ms | 0 - 5 MB | NPU |
| flow | QNN_CONTEXT_BINARY | mixed_with_float | Snapdragon® 8 Elite For Galaxy Mobile | 102.343 ms | 3 - 16 MB | NPU |
License
- The license for the original implementation of MeloTTS-ZH can be found here.
References
Community
- Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
- For questions or feedback please reach out to us.
