Whisper-Tiny-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-Tiny-En found here.

This repository provides scripts to run Whisper-Tiny-En 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: tiny.en
- Input resolution: 80x3000 (30 seconds audio)
- Mean decoded sequence length: 112 tokens
- Number of parameters (WhisperEncoder): 9.39M
- Model size (WhisperEncoder): 35.9 MB
- Number of parameters (WhisperDecoder): 28.2M
- Model size (WhisperDecoder): 108 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
WhisperDecoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	3.837 ms	3 - 28 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	2.209 ms	12 - 14 MB	FP16	NPU	Whisper-Tiny-En.so
WhisperDecoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	5.218 ms	6 - 182 MB	FP16	NPU	Whisper-Tiny-En.onnx
WhisperDecoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	3.0 ms	1 - 75 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	1.663 ms	4 - 23 MB	FP16	NPU	Whisper-Tiny-En.so
WhisperDecoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	3.739 ms	18 - 72 MB	FP16	NPU	Whisper-Tiny-En.onnx
WhisperDecoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	2.679 ms	0 - 71 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	1.37 ms	4 - 52 MB	FP16	NPU	Use Export Script
WhisperDecoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	3.722 ms	25 - 74 MB	FP16	NPU	Whisper-Tiny-En.onnx
WhisperDecoder	SA7255P ADP	SA7255P	TFLITE	17.729 ms	3 - 69 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	SA7255P ADP	SA7255P	QNN	15.132 ms	9 - 16 MB	FP16	NPU	Use Export Script
WhisperDecoder	SA8255 (Proxy)	SA8255P Proxy	TFLITE	3.734 ms	2 - 35 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	SA8255 (Proxy)	SA8255P Proxy	QNN	2.116 ms	10 - 13 MB	FP16	NPU	Use Export Script
WhisperDecoder	SA8295P ADP	SA8295P	TFLITE	4.959 ms	3 - 65 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	SA8295P ADP	SA8295P	QNN	3.096 ms	0 - 14 MB	FP16	NPU	Use Export Script
WhisperDecoder	SA8650 (Proxy)	SA8650P Proxy	TFLITE	3.787 ms	3 - 30 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	SA8650 (Proxy)	SA8650P Proxy	QNN	2.256 ms	10 - 13 MB	FP16	NPU	Use Export Script
WhisperDecoder	SA8775P ADP	SA8775P	TFLITE	5.098 ms	3 - 70 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	SA8775P ADP	SA8775P	QNN	2.986 ms	9 - 18 MB	FP16	NPU	Use Export Script
WhisperDecoder	QCS8275 (Proxy)	QCS8275 Proxy	TFLITE	17.729 ms	3 - 69 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	QCS8275 (Proxy)	QCS8275 Proxy	QNN	15.132 ms	9 - 16 MB	FP16	NPU	Use Export Script
WhisperDecoder	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	3.862 ms	3 - 37 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	QCS8550 (Proxy)	QCS8550 Proxy	QNN	2.144 ms	10 - 13 MB	FP16	NPU	Use Export Script
WhisperDecoder	QCS9075 (Proxy)	QCS9075 Proxy	TFLITE	5.098 ms	3 - 70 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	QCS9075 (Proxy)	QCS9075 Proxy	QNN	2.986 ms	9 - 18 MB	FP16	NPU	Use Export Script
WhisperDecoder	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	4.443 ms	3 - 71 MB	FP16	NPU	Whisper-Tiny-En.tflite
WhisperDecoder	QCS8450 (Proxy)	QCS8450 Proxy	QNN	2.63 ms	4 - 50 MB	FP16	NPU	Use Export Script
WhisperDecoder	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	2.132 ms	10 - 10 MB	FP16	NPU	Use Export Script
WhisperDecoder	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	4.199 ms	75 - 75 MB	FP16	NPU	Whisper-Tiny-En.onnx
WhisperEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	105.224 ms	20 - 68 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	71.508 ms	1 - 3 MB	FP16	NPU	Whisper-Tiny-En.so
WhisperEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	71.236 ms	35 - 144 MB	FP16	NPU	Whisper-Tiny-En.onnx
WhisperEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	82.877 ms	19 - 58 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	56.974 ms	1 - 19 MB	FP16	NPU	Whisper-Tiny-En.so
WhisperEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	56.279 ms	35 - 317 MB	FP16	NPU	Whisper-Tiny-En.onnx
WhisperEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	77.269 ms	20 - 46 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	50.539 ms	1 - 202 MB	FP16	NPU	Use Export Script
WhisperEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	48.99 ms	40 - 300 MB	FP16	NPU	Whisper-Tiny-En.onnx
WhisperEncoder	SA7255P ADP	SA7255P	TFLITE	530.794 ms	20 - 43 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	SA7255P ADP	SA7255P	QNN	445.254 ms	1 - 8 MB	FP16	NPU	Use Export Script
WhisperEncoder	SA8255 (Proxy)	SA8255P Proxy	TFLITE	101.135 ms	20 - 66 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	SA8255 (Proxy)	SA8255P Proxy	QNN	73.304 ms	1 - 3 MB	FP16	NPU	Use Export Script
WhisperEncoder	SA8295P ADP	SA8295P	TFLITE	106.672 ms	20 - 50 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	SA8295P ADP	SA8295P	QNN	88.456 ms	1 - 15 MB	FP16	NPU	Use Export Script
WhisperEncoder	SA8650 (Proxy)	SA8650P Proxy	TFLITE	102.773 ms	3 - 46 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	SA8650 (Proxy)	SA8650P Proxy	QNN	72.589 ms	1 - 4 MB	FP16	NPU	Use Export Script
WhisperEncoder	SA8775P ADP	SA8775P	TFLITE	179.392 ms	20 - 42 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	SA8775P ADP	SA8775P	QNN	86.566 ms	0 - 10 MB	FP16	NPU	Use Export Script
WhisperEncoder	QCS8275 (Proxy)	QCS8275 Proxy	TFLITE	530.794 ms	20 - 43 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	QCS8275 (Proxy)	QCS8275 Proxy	QNN	445.254 ms	1 - 8 MB	FP16	NPU	Use Export Script
WhisperEncoder	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	100.167 ms	20 - 69 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	QCS8550 (Proxy)	QCS8550 Proxy	QNN	70.514 ms	1 - 4 MB	FP16	NPU	Use Export Script
WhisperEncoder	QCS9075 (Proxy)	QCS9075 Proxy	TFLITE	179.392 ms	20 - 42 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	QCS9075 (Proxy)	QCS9075 Proxy	QNN	86.566 ms	0 - 10 MB	FP16	NPU	Use Export Script
WhisperEncoder	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	130.22 ms	19 - 64 MB	FP16	GPU	Whisper-Tiny-En.tflite
WhisperEncoder	QCS8450 (Proxy)	QCS8450 Proxy	QNN	97.048 ms	1 - 229 MB	FP16	NPU	Use Export Script
WhisperEncoder	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	69.281 ms	0 - 0 MB	FP16	NPU	Use Export Script
WhisperEncoder	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	66.618 ms	67 - 67 MB	FP16	NPU	Whisper-Tiny-En.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[whisper-tiny-en]"

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_tiny_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_tiny_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.

python -m qai_hub_models.models.whisper_tiny_en.export

Profiling Results
------------------------------------------------------------
WhisperDecoder
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 3.8                    
Estimated peak memory usage (MB): [3, 28]                
Total # Ops                     : 557                    
Compute Unit(s)                 : NPU (557 ops)          

------------------------------------------------------------
WhisperEncoder
Device                          : Samsung Galaxy S23 (13)   
Runtime                         : TFLITE                    
Estimated inference time (ms)   : 105.2                     
Estimated peak memory usage (MB): [20, 68]                  
Total # Ops                     : 271                       
Compute Unit(s)                 : GPU (260 ops) CPU (11 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_tiny_en import Model

# Load the model
model = Model.from_pretrained()
decoder_model = model.decoder
encoder_model = model.encoder

# Device
device = hub.Device("Samsung Galaxy S23")

# 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()
# 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()

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.

decoder_profile_job = hub.submit_profile_job(
    model=decoder_target_model,
    device=device,
)
encoder_profile_job = hub.submit_profile_job(
    model=encoder_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.

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()
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()

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-Tiny-En's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of Whisper-Tiny-En can be found here.
The license for the compiled assets for on-device deployment 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.