Zipformer: Optimized for Qualcomm Devices

Zipformer streaming ASR (Automatic Speech Recognition) model is a state-of-the-art system designed for transcribing spoken language into written text streamingly. This model is based on the transformer architecture and has been optimized for edge inference by replacing 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 audios. Time to the first token is the encoder's latency, while time to each additional token is joiner's latency, where we assume a max decoded length specified below.

This is based on the implementation of Zipformer found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up 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.

For more device-specific assets and performance metrics, visit Zipformer on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models 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 Zipformer on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.speech_recognition

Model Stats:

• Input resolution: 80x71 (0.71 seconds audio)
• Max decoded sequence length: 200 tokens
• Model checkpoint: pfluo/k2fsa-zipformer-chinese-english-mixed
• Model size (decoder) (float): 13.2 MB
• Model size (encoder) (float): 242 MB
• Model size (joiner) (float): 12.2 MB
• Number of parameters (decoder): 3.47M
• Number of parameters (encoder): 63.2M
• Number of parameters (joiner): 3.21M

Performance Summary

License

• The license for the original implementation of Zipformer can be found here.

References

• Zipformer A faster and better encoder for automatic speech recognition
• 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.