Zipformer: Optimized for Qualcomm Devices
Zipformer streaming ASR (Automatic Speech Recognition) model is a state-of-the-art system designed for transcribing spoken language into written text streamingly. This model is based on the transformer architecture and has been optimized for edge inference by replacing linear layers with convolutional (conv) layers. It exhibits robust performance in realistic, noisy environments, making it highly reliable for real-world applications. Specifically, it excels in long-form transcription, capable of accurately transcribing audios. Time to the first token is the encoder's latency, while time to each additional token is joiner's latency, where we assume a max decoded length specified below.
This is based on the implementation of Zipformer 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 | float | Snapdragon® X Elite | QAIRT 2.37, ONNX Runtime 1.23.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.37, ONNX Runtime 1.23.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8550 (Proxy) | QAIRT 2.37, ONNX Runtime 1.23.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.37, ONNX Runtime 1.23.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.37, ONNX Runtime 1.23.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS9075 | QAIRT 2.37, ONNX Runtime 1.23.0 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® QCS8275 (Proxy) | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® QCS8550 (Proxy) | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA7255P | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® QCS9075 | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 (Proxy) | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Snapdragon® X Elite | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8275 (Proxy) | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8550 (Proxy) | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA8775P | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA7255P | QAIRT 2.42 | Download |
| QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS9075 | QAIRT 2.42 | Download |
For more device-specific assets and performance metrics, visit Zipformer 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 Zipformer on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.speech_recognition
Model Stats:
- Model checkpoint: pfluo/k2fsa-zipformer-chinese-english-mixed
- Input resolution: 80x71 (0.71 seconds audio)
- Max decoded sequence length: 200 tokens
- Number of parameters (ZipformerEncoder): 63.2M
- Model size (ZipformerEncoder) (float): 242 MB
- Number of parameters (ZipformerDecoder): 3.47M
- Model size (ZipformerDecoder) (float): 13.2 MB
- Number of parameters (ZipformerJoiner): 3.21M
- Model size (ZipformerJoiner) (float): 12.2 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| ZipformerDecoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | 0.154 ms | 7 - 7 MB | NPU |
| ZipformerDecoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.2 ms | 0 - 8 MB | NPU |
| ZipformerDecoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8550 (Proxy) | 0.211 ms | 0 - 8 MB | NPU |
| ZipformerDecoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS9075 | 0.331 ms | 0 - 3 MB | NPU |
| ZipformerDecoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.157 ms | 0 - 12 MB | NPU |
| ZipformerDecoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.155 ms | 0 - 11 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | 0.157 ms | 0 - 0 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | 0.054 ms | 0 - 7 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8275 (Proxy) | 0.317 ms | 0 - 8 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8550 (Proxy) | 0.072 ms | 0 - 2 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | 0.206 ms | 0 - 5 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS9075 | 0.119 ms | 2 - 4 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 (Proxy) | 0.099 ms | 0 - 9 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA7255P | 0.317 ms | 0 - 8 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | 0.328 ms | 0 - 5 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.045 ms | 0 - 13 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.043 ms | 0 - 9 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® X Elite | 0.151 ms | 0 - 0 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Gen 3 Mobile | 0.062 ms | 0 - 7 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8275 (Proxy) | 0.32 ms | 0 - 8 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8550 (Proxy) | 0.077 ms | 0 - 2 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA8775P | 0.228 ms | 0 - 5 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS9075 | 0.128 ms | 2 - 4 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA7255P | 0.32 ms | 0 - 8 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite For Galaxy Mobile | 0.046 ms | 0 - 13 MB | NPU |
| ZipformerDecoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite Gen 5 Mobile | 0.045 ms | 0 - 10 MB | NPU |
| ZipformerEncoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | 9.807 ms | 142 - 142 MB | NPU |
| ZipformerEncoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | 7.014 ms | 9 - 17 MB | NPU |
| ZipformerEncoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8550 (Proxy) | 9.665 ms | 10 - 11 MB | NPU |
| ZipformerEncoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS9075 | 11.071 ms | 8 - 12 MB | NPU |
| ZipformerEncoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 5.879 ms | 9 - 21 MB | NPU |
| ZipformerEncoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 5.367 ms | 10 - 20 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | 9.567 ms | 3 - 3 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | 6.437 ms | 2 - 12 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8275 (Proxy) | 23.669 ms | 3 - 10 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8550 (Proxy) | 9.198 ms | 3 - 5 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | 10.839 ms | 3 - 11 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS9075 | 10.155 ms | 3 - 7 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 (Proxy) | 11.987 ms | 2 - 12 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA7255P | 23.669 ms | 3 - 10 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | 13.259 ms | 1 - 6 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite For Galaxy Mobile | 5.425 ms | 17 - 26 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | 4.858 ms | 3 - 12 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® X Elite | 14.971 ms | 1 - 1 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Gen 3 Mobile | 12.47 ms | 1 - 9 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8275 (Proxy) | 28.245 ms | 1 - 9 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8550 (Proxy) | 16.135 ms | 1 - 3 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA8775P | 17.76 ms | 1 - 10 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS9075 | 17.93 ms | 1 - 3 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA7255P | 28.245 ms | 1 - 9 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite For Galaxy Mobile | 8.709 ms | 1 - 10 MB | NPU |
| ZipformerEncoder | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite Gen 5 Mobile | 8.026 ms | 1 - 11 MB | NPU |
| ZipformerJoiner | PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | 0.247 ms | 6 - 6 MB | NPU |
| ZipformerJoiner | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.328 ms | 0 - 7 MB | NPU |
| ZipformerJoiner | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8550 (Proxy) | 0.352 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS9075 | 0.455 ms | 0 - 3 MB | NPU |
| ZipformerJoiner | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.23 ms | 0 - 7 MB | NPU |
| ZipformerJoiner | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.226 ms | 0 - 11 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | 0.254 ms | 0 - 0 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | 0.158 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8275 (Proxy) | 0.456 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8550 (Proxy) | 0.186 ms | 0 - 1 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | 0.914 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Qualcomm® QCS9075 | 0.232 ms | 2 - 4 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 (Proxy) | 0.233 ms | 0 - 9 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Qualcomm® SA7255P | 0.456 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | 0.46 ms | 0 - 5 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.135 ms | 0 - 13 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.129 ms | 0 - 9 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® X Elite | 0.243 ms | 0 - 0 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Gen 3 Mobile | 0.156 ms | 0 - 7 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8275 (Proxy) | 0.426 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS8550 (Proxy) | 0.186 ms | 0 - 1 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA8775P | 0.894 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® QCS9075 | 0.231 ms | 0 - 2 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Qualcomm® SA7255P | 0.426 ms | 0 - 8 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite For Galaxy Mobile | 0.133 ms | 0 - 9 MB | NPU |
| ZipformerJoiner | QNN_CONTEXT_BINARY | w16a16 | Snapdragon® 8 Elite Gen 5 Mobile | 0.121 ms | 0 - 10 MB | NPU |
License
- The license for the original implementation of Zipformer 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.
