Whisper-Small-Quantized: Optimized for Qualcomm Devices
We have applied w8a16 quantization to significantly enhance performance and efficiency. 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-Quantized 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 | w8a16 | Snapdragon® X Elite | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCS8550 (Proxy) | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® X2 Elite | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCM6690 | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCS9075 | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.42, ONNX Runtime 1.24.1 | Download |
| QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS8275 (Proxy) | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS8550 (Proxy) | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | w8a16 | Qualcomm® SA8775P | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | w8a16 | Snapdragon® X2 Elite | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | w8a16 | Qualcomm® SA7255P | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCM6690 | QAIRT 2.43 | Download |
| QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS9075 | QAIRT 2.43 | Download |
For more device-specific assets and performance metrics, visit Whisper-Small-Quantized 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-Quantized 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
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® X Elite | 7.991 ms | 185 - 185 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 6.451 ms | 38 - 50 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCS8550 (Proxy) | 8.335 ms | 28 - 29 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCS9075 | 9.124 ms | 24 - 57 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCM6690 | 35.618 ms | 28 - 38 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 4.767 ms | 25 - 37 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 10.888 ms | 28 - 35 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 3.998 ms | 30 - 40 MB | NPU |
| WhisperSmallDecoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® X2 Elite | 3.718 ms | 186 - 186 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® X Elite | 7.662 ms | 30 - 30 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 8 Gen 3 Mobile | 6.366 ms | 30 - 39 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS8275 (Proxy) | 13.482 ms | 29 - 37 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS8550 (Proxy) | 8.23 ms | 30 - 33 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® SA8775P | 9.233 ms | 19 - 27 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS9075 | 8.915 ms | 25 - 60 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCM6690 | 31.639 ms | 30 - 37 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® SA7255P | 13.482 ms | 29 - 37 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 4.74 ms | 21 - 35 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 7 Gen 4 Mobile | 10.815 ms | 30 - 37 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 3.972 ms | 30 - 40 MB | NPU |
| WhisperSmallDecoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® X2 Elite | 4.089 ms | 30 - 30 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® X Elite | 264.922 ms | 127 - 127 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 240.754 ms | 64 - 76 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCS8550 (Proxy) | 334.502 ms | 0 - 129 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCS9075 | 254.756 ms | 63 - 66 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Qualcomm® QCM6690 | 4091.143 ms | 2 - 12 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 198.385 ms | 63 - 73 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 457.801 ms | 56 - 63 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 175.574 ms | 63 - 73 MB | NPU |
| WhisperSmallEncoderQuantizable | PRECOMPILED_QNN_ONNX | w8a16 | Snapdragon® X2 Elite | 154.538 ms | 127 - 127 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® X Elite | 294.934 ms | 0 - 0 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 8 Gen 3 Mobile | 268.023 ms | 1 - 8 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS8275 (Proxy) | 517.1 ms | 1 - 9 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS8550 (Proxy) | 358.638 ms | 1 - 2 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® SA8775P | 310.466 ms | 0 - 8 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCS9075 | 291.599 ms | 0 - 29 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® QCM6690 | 4113.051 ms | 0 - 6 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Qualcomm® SA7255P | 517.1 ms | 1 - 9 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 224.015 ms | 1 - 10 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 7 Gen 4 Mobile | 467.841 ms | 1 - 7 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 187.818 ms | 1 - 10 MB | NPU |
| WhisperSmallEncoderQuantizable | QNN_CONTEXT_BINARY | w8a16 | Snapdragon® X2 Elite | 153.916 ms | 0 - 0 MB | NPU |
License
- The license for the original implementation of Whisper-Small-Quantized 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.