--- library_name: pytorch license: mit pipeline_tag: automatic-speech-recognition tags: - foundation - android --- ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/whisper_small_en/web-assets/model_demo.png) # Whisper-Small-En: Optimized for Mobile Deployment ## Automatic speech recognition (ASR) model for English transcription as well as translation OpenAI’s Whisper ASR (Automatic Speech Recognition) model is a state-of-the-art system designed for transcribing spoken language into written text. 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 mean decoded length specified below. This model is an implementation of Whisper-Small-En found [here](https://github.com/openai/whisper/tree/main). This repository provides scripts to run Whisper-Small-En on Qualcomm® devices. More details on model performance across various devices, can be found [here](https://aihub.qualcomm.com/models/whisper_small_en). ### Model Details - **Model Type:** Speech recognition - **Model Stats:** - Model checkpoint: small.en - Input resolution: 80x3000 (30 seconds audio) - Mean decoded sequence length: 112 tokens - Number of parameters (WhisperEncoder): 102M - Model size (WhisperEncoder): 390 MB - Number of parameters (WhisperDecoder): 139M - Model size (WhisperDecoder): 531 MB | Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | WhisperEncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 722.7 ms | 69 - 449 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 820.248 ms | 0 - 209 MB | FP16 | NPU | [Whisper-Small-En.so](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.so) | | WhisperEncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 518.95 ms | 111 - 201 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 778.511 ms | 113 - 3977 MB | FP16 | NPU | [Whisper-Small-En.onnx](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.onnx) | | WhisperEncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 516.678 ms | 0 - 906 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 696.316 ms | 85 - 465 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 644.566 ms | 1 - 2 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | SA7255P ADP | SA7255P | TFLITE | 4426.504 ms | 108 - 142 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | SA7255P ADP | SA7255P | QNN | 3210.318 ms | 1 - 8 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 705.299 ms | 26 - 406 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 638.347 ms | 1 - 2 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | SA8295P ADP | SA8295P | QNN | 700.683 ms | 3 - 9 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 709.764 ms | 78 - 445 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 678.71 ms | 1 - 3 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | SA8775P ADP | SA8775P | TFLITE | 1293.65 ms | 108 - 140 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | SA8775P ADP | SA8775P | QNN | 603.983 ms | 1 - 6 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 969.375 ms | 110 - 205 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) | | WhisperEncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 504.049 ms | 0 - 0 MB | FP16 | NPU | Use Export Script | | WhisperEncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 1342.641 ms | 237 - 237 MB | FP16 | NPU | [Whisper-Small-En.onnx](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.onnx) | | WhisperDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 28.657 ms | 16 - 100 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 11.929 ms | 61 - 141 MB | FP16 | NPU | [Whisper-Small-En.so](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.so) | | WhisperDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 58.778 ms | 120 - 123 MB | FP16 | NPU | [Whisper-Small-En.onnx](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.onnx) | | WhisperDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 23.885 ms | 16 - 148 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 9.45 ms | 446 - 552 MB | FP16 | NPU | [Whisper-Small-En.so](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.so) | | WhisperDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 47.995 ms | 85 - 1135 MB | FP16 | NPU | [Whisper-Small-En.onnx](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.onnx) | | WhisperDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 16.628 ms | 16 - 263 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 8.06 ms | 53 - 188 MB | FP16 | NPU | Use Export Script | | WhisperDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 44.088 ms | 69 - 697 MB | FP16 | NPU | [Whisper-Small-En.onnx](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.onnx) | | WhisperDecoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 28.65 ms | 16 - 101 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 11.92 ms | 61 - 71 MB | FP16 | NPU | Use Export Script | | WhisperDecoder | SA7255P ADP | SA7255P | QNN | 74.962 ms | 56 - 64 MB | FP16 | NPU | Use Export Script | | WhisperDecoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 29.533 ms | 16 - 99 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 12.125 ms | 57 - 62 MB | FP16 | NPU | Use Export Script | | WhisperDecoder | SA8295P ADP | SA8295P | TFLITE | 30.807 ms | 16 - 162 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | SA8295P ADP | SA8295P | QNN | 14.596 ms | 57 - 62 MB | FP16 | NPU | Use Export Script | | WhisperDecoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 29.43 ms | 16 - 99 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 12.052 ms | 65 - 66 MB | FP16 | NPU | Use Export Script | | WhisperDecoder | SA8775P ADP | SA8775P | TFLITE | 33.02 ms | 16 - 174 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 34.145 ms | 16 - 139 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) | | WhisperDecoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 15.967 ms | 57 - 173 MB | FP16 | NPU | Use Export Script | | WhisperDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 52.917 ms | 232 - 232 MB | FP16 | NPU | [Whisper-Small-En.onnx](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.onnx) | ## Installation This model can be installed as a Python package via pip. ```bash pip install "qai-hub-models[whisper_small_en]" ``` ## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device Sign-in to [Qualcomm® AI Hub](https://app.aihub.qualcomm.com/) with your Qualcomm® ID. Once signed in navigate to `Account -> Settings -> API Token`. With this API token, you can configure your client to run models on the cloud hosted devices. ```bash qai-hub configure --api_token API_TOKEN ``` Navigate to [docs](https://app.aihub.qualcomm.com/docs/) for more information. ## Demo off target The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input. ```bash python -m qai_hub_models.models.whisper_small_en.demo ``` The above demo runs a reference implementation of pre-processing, model inference, and post processing. **NOTE**: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above). ``` %run -m qai_hub_models.models.whisper_small_en.demo ``` ### Run model on a cloud-hosted device In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following: * Performance check on-device on a cloud-hosted device * Downloads compiled assets that can be deployed on-device for Android. * Accuracy check between PyTorch and on-device outputs. ```bash python -m qai_hub_models.models.whisper_small_en.export ``` ``` Profiling Results ------------------------------------------------------------ WhisperEncoder Device : Samsung Galaxy S23 (13) Runtime : TFLITE Estimated inference time (ms) : 722.7 Estimated peak memory usage (MB): [69, 449] Total # Ops : 911 Compute Unit(s) : GPU (900 ops) CPU (11 ops) ------------------------------------------------------------ WhisperDecoder Device : Samsung Galaxy S23 (13) Runtime : TFLITE Estimated inference time (ms) : 28.7 Estimated peak memory usage (MB): [16, 100] Total # Ops : 2573 Compute Unit(s) : NPU (2573 ops) ``` ## How does this work? This [export script](https://aihub.qualcomm.com/models/whisper_small_en/qai_hub_models/models/Whisper-Small-En/export.py) leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) to optimize, validate, and deploy this model on-device. Lets go through each step below in detail: Step 1: **Compile model for on-device deployment** To compile a PyTorch model for on-device deployment, we first trace the model in memory using the `jit.trace` and then call the `submit_compile_job` API. ```python import torch import qai_hub as hub from qai_hub_models.models.whisper_small_en import Model # Load the model model = Model.from_pretrained() encoder_model = model.encoder decoder_model = model.decoder # Device device = hub.Device("Samsung Galaxy S23") # Trace model encoder_input_shape = encoder_model.get_input_spec() encoder_sample_inputs = encoder_model.sample_inputs() traced_encoder_model = torch.jit.trace(encoder_model, [torch.tensor(data[0]) for _, data in encoder_sample_inputs.items()]) # Compile model on a specific device encoder_compile_job = hub.submit_compile_job( model=traced_encoder_model , device=device, input_specs=encoder_model.get_input_spec(), ) # Get target model to run on-device encoder_target_model = encoder_compile_job.get_target_model() # Trace model decoder_input_shape = decoder_model.get_input_spec() decoder_sample_inputs = decoder_model.sample_inputs() traced_decoder_model = torch.jit.trace(decoder_model, [torch.tensor(data[0]) for _, data in decoder_sample_inputs.items()]) # Compile model on a specific device decoder_compile_job = hub.submit_compile_job( model=traced_decoder_model , device=device, input_specs=decoder_model.get_input_spec(), ) # Get target model to run on-device decoder_target_model = decoder_compile_job.get_target_model() ``` Step 2: **Performance profiling on cloud-hosted device** After compiling models from step 1. Models can be profiled model on-device using the `target_model`. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics. ```python encoder_profile_job = hub.submit_profile_job( model=encoder_target_model, device=device, ) decoder_profile_job = hub.submit_profile_job( model=decoder_target_model, device=device, ) ``` Step 3: **Verify on-device accuracy** To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device. ```python encoder_input_data = encoder_model.sample_inputs() encoder_inference_job = hub.submit_inference_job( model=encoder_target_model, device=device, inputs=encoder_input_data, ) encoder_inference_job.download_output_data() decoder_input_data = decoder_model.sample_inputs() decoder_inference_job = hub.submit_inference_job( model=decoder_target_model, device=device, inputs=decoder_input_data, ) decoder_inference_job.download_output_data() ``` With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output. **Note**: This on-device profiling and inference requires access to Qualcomm® AI Hub. [Sign up for access](https://myaccount.qualcomm.com/signup). ## Deploying compiled model to Android The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): [This tutorial](https://www.tensorflow.org/lite/android/quickstart) provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This [sample app](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/sample_app.html) provides instructions on how to use the `.so` shared library in an Android application. ## View on Qualcomm® AI Hub Get more details on Whisper-Small-En's performance across various devices [here](https://aihub.qualcomm.com/models/whisper_small_en). Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/) ## License * The license for the original implementation of Whisper-Small-En can be found [here](https://github.com/openai/whisper/blob/main/LICENSE). * The license for the compiled assets for on-device deployment can be found [here](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/Qualcomm+AI+Hub+Proprietary+License.pdf) ## References * [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) * [Source Model Implementation](https://github.com/openai/whisper/tree/main) ## Community * Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI. * For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).