Hugging Face's logo

qualcomm
/
Whisper-Small-Quantized

Automatic Speech Recognition
PyTorch
foundation
android
Model card Files
xet
 Community

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
QNN_CONTEXT_BINARY w8a16 qualcomm_qcm6690 QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_qcs8550_proxy QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_qcs9075 QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_sa7255p QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_sa8775p QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_snapdragon_7gen4 QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_snapdragon_8_elite_for_galaxy QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_snapdragon_8_elite_gen5 QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_snapdragon_8gen3 QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_snapdragon_x2_elite QAIRT 2.43 Download
QNN_CONTEXT_BINARY w8a16 qualcomm_snapdragon_x_elite 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
decoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 8 Elite Gen 5 Mobile 3.979 ms 38 - 48 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® X2 Elite 3.812 ms 185 - 185 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® X Elite 7.967 ms 185 - 185 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 8 Gen 3 Mobile 6.354 ms 40 - 48 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Qualcomm® QCS8550 (Proxy) 8.319 ms 29 - 32 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Qualcomm® QCS9075 9.06 ms 24 - 57 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Qualcomm® QCM6690 37.459 ms 29 - 39 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 8 Elite For Galaxy Mobile 4.757 ms 25 - 33 MB NPU
decoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 7 Gen 4 Mobile 10.823 ms 30 - 36 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 8 Elite Gen 5 Mobile 3.963 ms 30 - 40 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Snapdragon® X2 Elite 4.191 ms 30 - 30 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Snapdragon® X Elite 7.57 ms 30 - 30 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 8 Gen 3 Mobile 6.254 ms 30 - 39 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCS8275 (Proxy) 13.457 ms 13 - 20 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCS8550 (Proxy) 8.158 ms 30 - 32 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Qualcomm® SA8775P 9.257 ms 29 - 38 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCS9075 8.92 ms 25 - 60 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCM6690 36.842 ms 30 - 37 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Qualcomm® SA7255P 13.457 ms 13 - 20 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 8 Elite For Galaxy Mobile 4.751 ms 22 - 35 MB NPU
decoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 7 Gen 4 Mobile 10.848 ms 30 - 37 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 8 Elite Gen 5 Mobile 175.256 ms 63 - 73 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® X2 Elite 154.028 ms 127 - 127 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® X Elite 264.69 ms 127 - 127 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 8 Gen 3 Mobile 244.062 ms 63 - 75 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Qualcomm® QCS8550 (Proxy) 350.116 ms 55 - 58 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Qualcomm® QCS9075 255.908 ms 63 - 66 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Qualcomm® QCM6690 4170.586 ms 2 - 13 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 8 Elite For Galaxy Mobile 200.797 ms 63 - 75 MB NPU
encoder PRECOMPILED_QNN_ONNX w8a16 Snapdragon® 7 Gen 4 Mobile 457.805 ms 56 - 66 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 8 Elite Gen 5 Mobile 175.95 ms 1 - 11 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Snapdragon® X2 Elite 153.643 ms 0 - 0 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Snapdragon® X Elite 296.075 ms 0 - 0 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 8 Gen 3 Mobile 269.335 ms 1 - 8 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCS8275 (Proxy) 518.268 ms 1 - 9 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCS8550 (Proxy) 368.292 ms 1 - 2 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Qualcomm® SA8775P 312.155 ms 1 - 10 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCS9075 293.695 ms 0 - 29 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Qualcomm® QCM6690 4234.28 ms 0 - 7 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Qualcomm® SA7255P 518.268 ms 1 - 9 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 8 Elite For Galaxy Mobile 221.77 ms 1 - 14 MB NPU
encoder QNN_CONTEXT_BINARY w8a16 Snapdragon® 7 Gen 4 Mobile 468.035 ms 1 - 7 MB NPU

License

  • The license for the original implementation of Whisper-Small-Quantized can be found here.

References

Community

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
Automatic Speech Recognition
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support