Hugging Face's logo

qualcomm
/
Whisper-Small-V2

Automatic Speech Recognition
PyTorch
LiteRT
ONNX
foundation
android
Model card Files Community

Whisper-Small-V2: Optimized for Mobile Deployment

Transformer-based automatic speech recognition (ASR) model for multilingual transcription and translation available on HuggingFace

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 model is an implementation of Whisper-Small-V2 found here.

This repository provides scripts to run Whisper-Small-V2 on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

  • Model Type: Speech recognition
  • Model Stats:
    • Model checkpoint: openai/whisper-small
    • Input resolution: 80x3000 (30 seconds audio)
    • Max decoded sequence length: 200 tokens
    • Number of parameters (HfWhisperEncoder): 102M
    • Model size (HfWhisperEncoder): 391 MB
    • Number of parameters (HfWhisperDecoder): 139M
    • Model size (HfWhisperDecoder): 533 MB
Model Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Precision Primary Compute Unit Target Model
HfWhisperEncoder Samsung Galaxy S23 Snapdragon® 8 Gen 2 TFLITE 752.637 ms 18 - 162 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder Samsung Galaxy S23 Snapdragon® 8 Gen 2 QNN 136.234 ms 1 - 3 MB FP16 NPU Whisper-Small-V2.so
HfWhisperEncoder Samsung Galaxy S23 Snapdragon® 8 Gen 2 ONNX 363.241 ms 227 - 400 MB FP16 NPU Whisper-Small-V2.onnx
HfWhisperEncoder Samsung Galaxy S24 Snapdragon® 8 Gen 3 TFLITE 1012.44 ms 32 - 221 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder Samsung Galaxy S24 Snapdragon® 8 Gen 3 QNN 106.405 ms 0 - 19 MB FP16 NPU Whisper-Small-V2.so
HfWhisperEncoder Samsung Galaxy S24 Snapdragon® 8 Gen 3 ONNX 284.349 ms 228 - 2122 MB FP16 NPU Whisper-Small-V2.onnx
HfWhisperEncoder Snapdragon 8 Elite QRD Snapdragon® 8 Elite TFLITE 509.76 ms 112 - 157 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder Snapdragon 8 Elite QRD Snapdragon® 8 Elite QNN 90.562 ms 0 - 15 MB FP16 NPU Use Export Script
HfWhisperEncoder Snapdragon 8 Elite QRD Snapdragon® 8 Elite ONNX 234.79 ms 228 - 1679 MB FP16 NPU Whisper-Small-V2.onnx
HfWhisperEncoder SA7255P ADP SA7255P TFLITE 4593.502 ms 108 - 158 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder SA7255P ADP SA7255P QNN 2646.985 ms 0 - 10 MB FP16 NPU Use Export Script
HfWhisperEncoder SA8255 (Proxy) SA8255P Proxy TFLITE 762.434 ms 28 - 138 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder SA8255 (Proxy) SA8255P Proxy QNN 137.83 ms 1 - 3 MB FP16 NPU Use Export Script
HfWhisperEncoder SA8295P ADP SA8295P TFLITE 802.005 ms 108 - 159 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder SA8295P ADP SA8295P QNN 246.174 ms 0 - 17 MB FP16 NPU Use Export Script
HfWhisperEncoder SA8650 (Proxy) SA8650P Proxy TFLITE 759.628 ms 18 - 205 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder SA8650 (Proxy) SA8650P Proxy QNN 135.873 ms 1 - 3 MB FP16 NPU Use Export Script
HfWhisperEncoder SA8775P ADP SA8775P TFLITE 1435.317 ms 96 - 146 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder SA8775P ADP SA8775P QNN 188.977 ms 0 - 9 MB FP16 NPU Use Export Script
HfWhisperEncoder QCS8275 (Proxy) QCS8275 Proxy TFLITE 4593.502 ms 108 - 158 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder QCS8275 (Proxy) QCS8275 Proxy QNN 2646.985 ms 0 - 10 MB FP16 NPU Use Export Script
HfWhisperEncoder QCS8550 (Proxy) QCS8550 Proxy TFLITE 746.162 ms 18 - 228 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder QCS8550 (Proxy) QCS8550 Proxy QNN 136.442 ms 0 - 3 MB FP16 NPU Use Export Script
HfWhisperEncoder QCS9075 (Proxy) QCS9075 Proxy TFLITE 1435.317 ms 96 - 146 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder QCS9075 (Proxy) QCS9075 Proxy QNN 188.977 ms 0 - 9 MB FP16 NPU Use Export Script
HfWhisperEncoder QCS8450 (Proxy) QCS8450 Proxy TFLITE 1059.574 ms 67 - 263 MB FP16 GPU Whisper-Small-V2.tflite
HfWhisperEncoder QCS8450 (Proxy) QCS8450 Proxy QNN 319.419 ms 0 - 16 MB FP16 NPU Use Export Script
HfWhisperEncoder Snapdragon X Elite CRD Snapdragon® X Elite QNN 132.379 ms 0 - 0 MB FP16 NPU Use Export Script
HfWhisperEncoder Snapdragon X Elite CRD Snapdragon® X Elite ONNX 356.931 ms 265 - 265 MB FP16 NPU Whisper-Small-V2.onnx
HfWhisperDecoder Samsung Galaxy S23 Snapdragon® 8 Gen 2 TFLITE 24.091 ms 14 - 36 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder Samsung Galaxy S23 Snapdragon® 8 Gen 2 QNN 12.1 ms 60 - 63 MB FP16 NPU Whisper-Small-V2.so
HfWhisperDecoder Samsung Galaxy S23 Snapdragon® 8 Gen 2 ONNX 24.895 ms 74 - 206 MB FP16 NPU Whisper-Small-V2.onnx
HfWhisperDecoder Samsung Galaxy S24 Snapdragon® 8 Gen 3 TFLITE 18.75 ms 14 - 664 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder Samsung Galaxy S24 Snapdragon® 8 Gen 3 QNN 9.89 ms 60 - 79 MB FP16 NPU Whisper-Small-V2.so
HfWhisperDecoder Samsung Galaxy S24 Snapdragon® 8 Gen 3 ONNX 20.73 ms 78 - 573 MB FP16 NPU Whisper-Small-V2.onnx
HfWhisperDecoder Snapdragon 8 Elite QRD Snapdragon® 8 Elite TFLITE 14.245 ms 0 - 372 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder Snapdragon 8 Elite QRD Snapdragon® 8 Elite QNN 8.085 ms 58 - 73 MB FP16 NPU Use Export Script
HfWhisperDecoder Snapdragon 8 Elite QRD Snapdragon® 8 Elite ONNX 16.576 ms 145 - 280 MB FP16 NPU Whisper-Small-V2.onnx
HfWhisperDecoder SA7255P ADP SA7255P TFLITE 94.243 ms 14 - 388 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder SA7255P ADP SA7255P QNN 73.516 ms 53 - 63 MB FP16 NPU Use Export Script
HfWhisperDecoder SA8255 (Proxy) SA8255P Proxy TFLITE 23.345 ms 14 - 35 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder SA8255 (Proxy) SA8255P Proxy QNN 11.969 ms 60 - 62 MB FP16 NPU Use Export Script
HfWhisperDecoder SA8295P ADP SA8295P TFLITE 24.64 ms 14 - 390 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder SA8295P ADP SA8295P QNN 14.534 ms 55 - 69 MB FP16 NPU Use Export Script
HfWhisperDecoder SA8650 (Proxy) SA8650P Proxy TFLITE 24.211 ms 14 - 35 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder SA8650 (Proxy) SA8650P Proxy QNN 11.914 ms 58 - 61 MB FP16 NPU Use Export Script
HfWhisperDecoder SA8775P ADP SA8775P TFLITE 27.981 ms 14 - 387 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder SA8775P ADP SA8775P QNN 13.956 ms 53 - 62 MB FP16 NPU Use Export Script
HfWhisperDecoder QCS8275 (Proxy) QCS8275 Proxy TFLITE 94.243 ms 14 - 388 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder QCS8275 (Proxy) QCS8275 Proxy QNN 73.516 ms 53 - 63 MB FP16 NPU Use Export Script
HfWhisperDecoder QCS8550 (Proxy) QCS8550 Proxy TFLITE 23.945 ms 14 - 32 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder QCS8550 (Proxy) QCS8550 Proxy QNN 11.636 ms 60 - 63 MB FP16 NPU Use Export Script
HfWhisperDecoder QCS9075 (Proxy) QCS9075 Proxy TFLITE 27.981 ms 14 - 387 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder QCS9075 (Proxy) QCS9075 Proxy QNN 13.956 ms 53 - 62 MB FP16 NPU Use Export Script
HfWhisperDecoder QCS8450 (Proxy) QCS8450 Proxy TFLITE 25.235 ms 14 - 670 MB FP16 NPU Whisper-Small-V2.tflite
HfWhisperDecoder QCS8450 (Proxy) QCS8450 Proxy QNN 19.162 ms 60 - 82 MB FP16 NPU Use Export Script
HfWhisperDecoder Snapdragon X Elite CRD Snapdragon® X Elite QNN 9.929 ms 60 - 60 MB FP16 NPU Use Export Script
HfWhisperDecoder Snapdragon X Elite CRD Snapdragon® X Elite ONNX 22.802 ms 227 - 227 MB FP16 NPU Whisper-Small-V2.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[whisper-small-v2]"

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub 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.

qai-hub configure --api_token API_TOKEN

Navigate to 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.

python -m qai_hub_models.models.whisper_small_v2.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_v2.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.
python -m qai_hub_models.models.whisper_small_v2.export

Profiling Results
------------------------------------------------------------
HfWhisperEncoder
Device                          : Samsung Galaxy S23 (13)   
Runtime                         : TFLITE                    
Estimated inference time (ms)   : 752.6                     
Estimated peak memory usage (MB): [18, 162]                 
Total # Ops                     : 1806                      
Compute Unit(s)                 : GPU (1798 ops) CPU (8 ops)

------------------------------------------------------------
HfWhisperDecoder
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 24.1                   
Estimated peak memory usage (MB): [14, 36]               
Total # Ops                     : 3136                   
Compute Unit(s)                 : NPU (3136 ops)

How does this work?

This export script leverages Qualcomm® AI Hub 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.

import torch

import qai_hub as hub
from qai_hub_models.models.whisper_small_v2 import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S24")

# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()

pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])

# Compile model on a specific device
compile_job = hub.submit_compile_job(
    model=pt_model,
    device=device,
    input_specs=torch_model.get_input_spec(),
)

# Get target model to run on-device
target_model = 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.

profile_job = hub.submit_profile_job(
    model=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.

input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
    model=target_model,
    device=device,
    inputs=input_data,
)
    on_device_output = 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.

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

  • TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.

  • QNN (.so export ): This sample app 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-V2's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of Whisper-Small-V2 can be found here.
  • The license for the compiled assets for on-device deployment 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