Distil-Whisper: Optimized for Qualcomm Devices

Distil-Whisper Small English is a distilled version of Whisper Small, optimized for fast and efficient automatic speech recognition.

This is based on the implementation of Distil-Whisper 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
ONNX	float	Universal	QAIRT 2.42, ONNX Runtime 1.24.3	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.19.1	Download

For more device-specific assets and performance metrics, visit Distil-Whisper 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 Distil-Whisper on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.speech_recognition

Model Stats:

Model checkpoint: distil-whisper/distil-small.en
Input resolution: 80x3000 (30 seconds audio)
Max decoded sequence length: 200 tokens
Number of parameters (encoder): 166M
Model size (encoder) (float): 332 MB
Number of parameters (decoder): 211M
Model size (decoder) (float): 450MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
decoder	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	5.544 ms	51 - 375 MB	NPU
decoder	ONNX	float	Snapdragon® X2 Elite	5.102 ms	178 - 178 MB	NPU
decoder	ONNX	float	Snapdragon® X Elite	11.168 ms	178 - 178 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Gen 3 Mobile	8.615 ms	0 - 364 MB	NPU
decoder	ONNX	float	Qualcomm® QCS8550 (Proxy)	11.602 ms	0 - 195 MB	NPU
decoder	ONNX	float	Qualcomm® QCS9075	13.163 ms	40 - 82 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	7.186 ms	16 - 476 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	5.544 ms	17 - 326 MB	NPU
decoder	QNN_DLC	float	Snapdragon® X2 Elite	5.615 ms	40 - 40 MB	NPU
decoder	QNN_DLC	float	Snapdragon® X Elite	10.866 ms	40 - 40 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	8.59 ms	17 - 330 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	18.893 ms	30 - 246 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	11.254 ms	40 - 42 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA8775P	12.908 ms	19 - 236 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS9075	13.082 ms	40 - 86 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	18.076 ms	36 - 345 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA7255P	18.893 ms	30 - 246 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA8295P	14.217 ms	20 - 271 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	7.215 ms	5 - 447 MB	NPU
decoder	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	5.627 ms	4 - 480 MB	NPU
decoder	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	8.471 ms	4 - 556 MB	NPU
decoder	TFLITE	float	Qualcomm® QCS8275 (Proxy)	18.822 ms	5 - 450 MB	NPU
decoder	TFLITE	float	Qualcomm® QCS8550 (Proxy)	11.489 ms	5 - 7 MB	NPU
decoder	TFLITE	float	Qualcomm® SA8775P	12.895 ms	5 - 450 MB	NPU
decoder	TFLITE	float	Qualcomm® QCS9075	13.074 ms	2 - 266 MB	NPU
decoder	TFLITE	float	Qualcomm® QCS8450 (Proxy)	18.592 ms	5 - 524 MB	NPU
decoder	TFLITE	float	Qualcomm® SA7255P	18.822 ms	5 - 450 MB	NPU
decoder	TFLITE	float	Qualcomm® SA8295P	14.156 ms	5 - 443 MB	NPU
decoder	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	6.994 ms	3 - 504 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	50.657 ms	77 - 836 MB	NPU
encoder	ONNX	float	Snapdragon® X2 Elite	50.82 ms	183 - 183 MB	NPU
encoder	ONNX	float	Snapdragon® X Elite	124.045 ms	182 - 182 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Gen 3 Mobile	82.002 ms	86 - 1245 MB	NPU
encoder	ONNX	float	Qualcomm® QCS8550 (Proxy)	118.492 ms	0 - 197 MB	NPU
encoder	ONNX	float	Qualcomm® QCS9075	150.326 ms	79 - 83 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	60.397 ms	81 - 774 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	49.034 ms	0 - 742 MB	NPU
encoder	QNN_DLC	float	Snapdragon® X2 Elite	48.393 ms	1 - 1 MB	NPU
encoder	QNN_DLC	float	Snapdragon® X Elite	121.663 ms	1 - 1 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	82.41 ms	0 - 1046 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	397.445 ms	1 - 731 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	118.837 ms	1 - 3 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA8775P	139.846 ms	1 - 733 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS9075	152.427 ms	1 - 39 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	276.295 ms	1 - 917 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA7255P	397.445 ms	1 - 731 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA8295P	207.259 ms	1 - 668 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	60.835 ms	1 - 665 MB	NPU
encoder	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	419.256 ms	41 - 81 MB	GPU
encoder	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	479.748 ms	0 - 144 MB	GPU
encoder	TFLITE	float	Qualcomm® QCS8275 (Proxy)	3116.266 ms	38 - 82 MB	GPU
encoder	TFLITE	float	Qualcomm® QCS8550 (Proxy)	655.878 ms	0 - 313 MB	GPU
encoder	TFLITE	float	Qualcomm® SA8775P	1326.077 ms	27 - 71 MB	GPU
encoder	TFLITE	float	Qualcomm® QCS9075	1270.615 ms	0 - 40 MB	GPU
encoder	TFLITE	float	Qualcomm® QCS8450 (Proxy)	843.956 ms	39 - 193 MB	GPU
encoder	TFLITE	float	Qualcomm® SA7255P	3116.266 ms	38 - 82 MB	GPU
encoder	TFLITE	float	Qualcomm® SA8295P	667.309 ms	40 - 83 MB	GPU
encoder	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	409.596 ms	41 - 80 MB	GPU

License

The license for the original implementation of Distil-Whisper 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.

Downloads last month: -; Downloads are not tracked for this model. How to track

Paper for qualcomm/Distil-Whisper

Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling

Paper • 2311.00430 • Published Nov 1, 2023 • 56