| --- |
| library_name: pytorch |
| license: apache-2.0 |
| tags: |
| - foundation |
| - real_time |
| - android |
| pipeline_tag: automatic-speech-recognition |
|
|
| --- |
| |
|  |
|
|
| # DeepSpeech2: Optimized for Qualcomm Devices |
|
|
| DeepSpeech2 is an end-to-end automatic speech recognition (ASR) model. It uses convolutional layers for feature extraction followed by bidirectional recurrent layers and CTC for sequence-to-sequence learning. |
|
|
| This is based on the implementation of DeepSpeech2 found [here](https://github.com/SeanNaren/deepspeech.pytorch). |
| This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.64.0/src/qai_hub_models/models/deepspeech2) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary). |
|
|
| Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) 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 | |
| |---|---|---|---|---| |
| | TFLITE | float | Universal | | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deepspeech2/releases/v0.64.0/deepspeech2-tflite-float.zip) |
|
|
| For more device-specific assets and performance metrics, visit **[DeepSpeech2 on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/deepspeech2)**. |
|
|
|
|
| ### Option 2: Export with Custom Configurations |
|
|
| Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.64.0/src/qai_hub_models/models/deepspeech2) 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 [DeepSpeech2 on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.64.0/src/qai_hub_models/models/deepspeech2) for usage instructions. |
|
|
| ## Model Details |
|
|
| **Model Type:** Model_use_case.speech_recognition |
| |
| **Model Stats:** |
| - Input resolution: Spectrogram (800 frames x 161 features) |
| - Model size: 330.48MB |
| - Number of parameters: 94.6M |
| |
| ## Performance Summary |
| | Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
| |---|---|---|---|---|---|--- |
| | DeepSpeech2 | TFLITE | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 3206.956 ms | 111 - 125 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 3172.903 ms | 39 - 52 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 4147.87 ms | 196 - 209 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 5833.182 ms | 200 - 215 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 11021.262 ms | 68 - 550 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 4139.576 ms | 81 - 83 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® SA8775P | 14871.745 ms | 112 - 122 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® SA8650P | 14871.745 ms | 112 - 122 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® SA8255P | 14871.745 ms | 112 - 122 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® QCS8450 | 5833.182 ms | 200 - 215 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 5249.532 ms | 68 - 550 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 3172.903 ms | 39 - 52 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® SA7255P | 18286.913 ms | 107 - 117 MB | CPU |
| | DeepSpeech2 | TFLITE | float | Qualcomm® SA8295P | 9335.031 ms | 100 - 110 MB | CPU |
| |
| ## License |
| * The license for the original implementation of DeepSpeech2 can be found |
| [here](https://github.com/SeanNaren/deepspeech.pytorch/blob/master/LICENCE). |
| |
| ## References |
| * [Deep Speech 2: End-to-End Speech Recognition in English and Mandarin](https://arxiv.org/abs/1512.02595) |
| * [Source Model Implementation](https://github.com/SeanNaren/deepspeech.pytorch) |
| |
| ## 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). |
| |
