Whisper-Large-V3-Turbo: Optimized for Mobile Deployment

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

Whisper large-v3-turbo is a finetuned version of a pruned Whisper large-v3. In other words, it's the exact same model, except that the number of decoding layers have reduced from 32 to 4. As a result, the model is way faster, at the expense of a minor quality degradation. 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-Large-V3-Turbo found here.

This repository provides scripts to run Whisper-Large-V3-Turbo on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

• Model Type: Model_use_case.speech_recognition
• Model Stats:
  • Model checkpoint: openai/whisper-large-v3-turbo
  • Input resolution: 128x3000 (30 seconds audio)
  • Max decoded sequence length: 200 tokens

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
HfWhisperEncoder	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_CONTEXT_BINARY	2169.924 ms	2 - 10 MB	NPU	Use Export Script
HfWhisperEncoder	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_CONTEXT_BINARY	592.709 ms	2 - 4 MB	NPU	Use Export Script
HfWhisperEncoder	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	PRECOMPILED_QNN_ONNX	805.415 ms	0 - 1537 MB	NPU	Use Export Script
HfWhisperEncoder	float	SA7255P ADP	Qualcomm® SA7255P	QNN_CONTEXT_BINARY	2169.924 ms	2 - 10 MB	NPU	Use Export Script
HfWhisperEncoder	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_CONTEXT_BINARY	596.904 ms	0 - 3 MB	NPU	Use Export Script
HfWhisperEncoder	float	SA8295P ADP	Qualcomm® SA8295P	QNN_CONTEXT_BINARY	1047.515 ms	2 - 15 MB	NPU	Use Export Script
HfWhisperEncoder	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_CONTEXT_BINARY	597.193 ms	2 - 8 MB	NPU	Use Export Script
HfWhisperEncoder	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_CONTEXT_BINARY	436.222 ms	2 - 18 MB	NPU	Use Export Script
HfWhisperEncoder	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	PRECOMPILED_QNN_ONNX	611.332 ms	65 - 80 MB	NPU	Use Export Script
HfWhisperEncoder	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_CONTEXT_BINARY	321.191 ms	2 - 15 MB	NPU	Use Export Script
HfWhisperEncoder	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	PRECOMPILED_QNN_ONNX	472.462 ms	125 - 141 MB	NPU	Use Export Script
HfWhisperEncoder	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	QNN_CONTEXT_BINARY	260.094 ms	4 - 11 MB	NPU	Use Export Script
HfWhisperEncoder	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	PRECOMPILED_QNN_ONNX	401.758 ms	125 - 136 MB	NPU	Use Export Script
HfWhisperEncoder	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_CONTEXT_BINARY	588.236 ms	1 - 1 MB	NPU	Use Export Script
HfWhisperEncoder	float	Snapdragon X Elite CRD	Snapdragon® X Elite	PRECOMPILED_QNN_ONNX	800.958 ms	1395 - 1395 MB	NPU	Use Export Script
HfWhisperDecoder	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_CONTEXT_BINARY	19.253 ms	56 - 65 MB	NPU	Use Export Script
HfWhisperDecoder	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_CONTEXT_BINARY	13.161 ms	67 - 74 MB	NPU	Use Export Script
HfWhisperDecoder	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	PRECOMPILED_QNN_ONNX	13.068 ms	0 - 379 MB	NPU	Use Export Script
HfWhisperDecoder	float	SA7255P ADP	Qualcomm® SA7255P	QNN_CONTEXT_BINARY	19.253 ms	56 - 65 MB	NPU	Use Export Script
HfWhisperDecoder	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_CONTEXT_BINARY	13.049 ms	67 - 70 MB	NPU	Use Export Script
HfWhisperDecoder	float	SA8295P ADP	Qualcomm® SA8295P	QNN_CONTEXT_BINARY	15.054 ms	66 - 81 MB	NPU	Use Export Script
HfWhisperDecoder	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_CONTEXT_BINARY	13.052 ms	66 - 70 MB	NPU	Use Export Script
HfWhisperDecoder	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_CONTEXT_BINARY	9.846 ms	66 - 85 MB	NPU	Use Export Script
HfWhisperDecoder	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	PRECOMPILED_QNN_ONNX	10.319 ms	84 - 102 MB	NPU	Use Export Script
HfWhisperDecoder	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_CONTEXT_BINARY	8.994 ms	33 - 51 MB	NPU	Use Export Script
HfWhisperDecoder	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	PRECOMPILED_QNN_ONNX	9.449 ms	46 - 56 MB	NPU	Use Export Script
HfWhisperDecoder	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	QNN_CONTEXT_BINARY	7.879 ms	66 - 78 MB	NPU	Use Export Script
HfWhisperDecoder	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	PRECOMPILED_QNN_ONNX	8.117 ms	83 - 94 MB	NPU	Use Export Script
HfWhisperDecoder	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_CONTEXT_BINARY	10.801 ms	67 - 67 MB	NPU	Use Export Script
HfWhisperDecoder	float	Snapdragon X Elite CRD	Snapdragon® X Elite	PRECOMPILED_QNN_ONNX	10.577 ms	367 - 367 MB	NPU	Use Export Script

Installation

Install the package via pip:

pip install "qai-hub-models[whisper-large-v3-turbo]"

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_large_v3_turbo.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_large_v3_turbo.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_large_v3_turbo.export

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_large_v3_turbo import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S25")

# 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-Large-V3-Turbo's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

• The license for the original implementation of Whisper-Large-V3-Turbo can be found here.
• The license for the compiled assets for on-device deployment can be found here

References

• Robust Speech Recognition via Large-Scale Weak Supervision
• Source Model Implementation

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.