Whisper-Small: Optimized for Qualcomm Devices
HuggingFace Whisper-Small ASR (Automatic Speech Recognition) model is a state-of-the-art system designed for transcribing spoken language into written text. This model is based on the transformer architecture and has been optimized for edge inference by replacing Multi-Head Attention (MHA) with Single-Head Attention (SHA) and 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 audio clips up to 30 seconds long. Time to the first token is the encoder's latency, while time to each additional token is decoder's latency, where we assume a max decoded length specified below.
This is based on the implementation of Whisper-Small 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® 8 Elite Gen 5 Mobile | QAIRT 2.42, ONNX Runtime 1.25.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® X2 Elite | QAIRT 2.42, ONNX Runtime 1.25.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | QAIRT 2.42, ONNX Runtime 1.25.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.42, ONNX Runtime 1.25.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8550 (Proxy) | QAIRT 2.42, ONNX Runtime 1.25.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.42, ONNX Runtime 1.25.0 | Download |
| PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS9075 | QAIRT 2.42, ONNX Runtime 1.25.0 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® X2 Elite | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® QCS8550 (Proxy) | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA7255P | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® QCS9075 | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 (Proxy) | QAIRT 2.45 | Download |
For more device-specific assets and performance metrics, visit Whisper-Small 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 Whisper-Small on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.speech_recognition
Model Stats:
- Model checkpoint: openai/whisper-small
- Input resolution: 80x3000 (30 seconds audio)
- Max decoded sequence length: 200 tokens
- Number of parameters (encoder): 102M
- Model size (encoder) (float): 391 MB
- Number of parameters (decoder): 139M
- Model size (decoder) (float): 533 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| decoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 7.216 ms | 75 - 84 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® X2 Elite | 6.455 ms | 60 - 60 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | 10.125 ms | 286 - 286 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | 9.758 ms | 52 - 59 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8550 (Proxy) | 12.404 ms | 0 - 319 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 8.365 ms | 54 - 65 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS9075 | 13.338 ms | 60 - 122 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8750 | 8.365 ms | 54 - 65 MB | NPU |
| decoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS7181 | 10.125 ms | 286 - 286 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | 7.379 ms | 43 - 53 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Snapdragon® X2 Elite | 7.039 ms | 60 - 60 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | 10.535 ms | 60 - 60 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | 9.912 ms | 57 - 65 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8550 (Proxy) | 12.185 ms | 44 - 46 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | 13.757 ms | 43 - 50 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8650P | 13.757 ms | 43 - 50 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8255P | 13.757 ms | 43 - 50 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | 14.322 ms | 57 - 63 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 (Proxy) | 17.769 ms | 60 - 69 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite For Galaxy Mobile | 8.347 ms | 0 - 9 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS9075 | 13.409 ms | 60 - 128 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8750 | 8.347 ms | 0 - 9 MB | NPU |
| decoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS7181 | 10.535 ms | 60 - 60 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 45.515 ms | 127 - 137 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® X2 Elite | 47.779 ms | 133 - 133 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | 114.36 ms | 202 - 202 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | 82.234 ms | 129 - 139 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8550 (Proxy) | 113.014 ms | 126 - 128 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 59.531 ms | 112 - 124 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS9075 | 136.903 ms | 130 - 133 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8750 | 59.531 ms | 112 - 124 MB | NPU |
| encoder | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS7181 | 114.36 ms | 202 - 202 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | 58.619 ms | 1 - 11 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Snapdragon® X2 Elite | 60.02 ms | 0 - 0 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | 132.182 ms | 0 - 0 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | 97.213 ms | 1 - 8 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8550 (Proxy) | 129.745 ms | 0 - 3 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | 154.359 ms | 1 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8650P | 154.359 ms | 1 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8255P | 154.359 ms | 1 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | 200.0 ms | 1 - 6 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 (Proxy) | 272.084 ms | 1 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite For Galaxy Mobile | 70.337 ms | 1 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS9075 | 154.389 ms | 1 - 55 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8750 | 70.337 ms | 1 - 10 MB | NPU |
| encoder | QNN_CONTEXT_BINARY | float | Qualcomm® QCS7181 | 132.182 ms | 0 - 0 MB | NPU |
License
- The license for the original implementation of Whisper-Small 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.