Parakeet-TDT-0.6B-v3: Optimized for Qualcomm Devices

Parakeet-TDT-0.6B-v3 is a 600M-parameter ASR model by NVIDIA using TDT architecture with a FastConformer encoder and 2-layer LSTM predictor, split into Encoder, Decoder, and Joint components for efficient on-device inference across English and 25 European languages.

This is based on the implementation of Parakeet-TDT-0.6B-v3 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.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.50, ONNX Runtime 1.30.0	Download
ONNX	mixed_with_float	Universal	QAIRT 2.50, ONNX Runtime 1.30.0	Download
QNN_DLC	float	Universal	QAIRT 2.50	Download
QNN_DLC	mixed_with_float	Universal	QAIRT 2.50	Download

For more device-specific assets and performance metrics, visit Parakeet-TDT-0.6B-v3 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 Parakeet-TDT-0.6B-v3 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.speech_recognition

Model Stats:

• Input resolution: 128 mel bins x variable time (16kHz mono audio)
• Model checkpoint: nvidia/parakeet-tdt-0.6b-v3
• Model size (decoder) (float): 47 MB
• Model size (encoder) (float): 2270 MB
• Model size (joint) (float): 20 MB
• Number of parameters (decoder): 12.2M
• Number of parameters (encoder): 609.5M
• Number of parameters (joint): 5.3M
• Supported languages: English + 25 European languages
• TDT durations: 5 (0,1,2,3,4 frames)
• Total parameters: 627.0M
• Vocabulary size: 8193 (8192 BPE + 1 blank)
• WER (LibriSpeech test-clean): 1.93%
• WER (LibriSpeech test-other): 3.59%

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
decoder	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.254 ms	0 - 28 MB	NPU
decoder	ONNX	float	Qualcomm® Dragonwing™ Q-8750	0.254 ms	0 - 28 MB	NPU
decoder	ONNX	mixed_with_float	Snapdragon® 8 Elite For Galaxy Mobile	0.172 ms	0 - 36 MB	NPU
decoder	ONNX	mixed_with_float	Qualcomm® Dragonwing™ Q-8750	0.172 ms	0 - 36 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.237 ms	0 - 28 MB	NPU
decoder	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	0.237 ms	0 - 28 MB	NPU
decoder	QNN_DLC	mixed_with_float	Snapdragon® 8 Elite For Galaxy Mobile	0.157 ms	0 - 32 MB	NPU
decoder	QNN_DLC	mixed_with_float	Qualcomm® Dragonwing™ Q-8750	0.157 ms	0 - 32 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	41.336 ms	2 - 1149 MB	NPU
encoder	ONNX	float	Qualcomm® Dragonwing™ Q-8750	41.336 ms	2 - 1149 MB	NPU
encoder	ONNX	mixed_with_float	Snapdragon® 8 Elite For Galaxy Mobile	40.611 ms	2 - 1148 MB	NPU
encoder	ONNX	mixed_with_float	Qualcomm® Dragonwing™ Q-8750	40.611 ms	2 - 1148 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	38.004 ms	1 - 948 MB	NPU
encoder	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	38.004 ms	1 - 948 MB	NPU
encoder	QNN_DLC	mixed_with_float	Snapdragon® 8 Elite For Galaxy Mobile	38.305 ms	1 - 948 MB	NPU
encoder	QNN_DLC	mixed_with_float	Qualcomm® Dragonwing™ Q-8750	38.305 ms	1 - 948 MB	NPU
joint	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.196 ms	0 - 43 MB	NPU
joint	ONNX	float	Qualcomm® Dragonwing™ Q-8750	0.196 ms	0 - 43 MB	NPU
joint	ONNX	mixed_with_float	Snapdragon® 8 Elite For Galaxy Mobile	0.122 ms	0 - 32 MB	NPU
joint	ONNX	mixed_with_float	Qualcomm® Dragonwing™ Q-8750	0.122 ms	0 - 32 MB	NPU
joint	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.224 ms	0 - 42 MB	NPU
joint	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	0.224 ms	0 - 42 MB	NPU
joint	QNN_DLC	mixed_with_float	Snapdragon® 8 Elite For Galaxy Mobile	0.12 ms	0 - 27 MB	NPU
joint	QNN_DLC	mixed_with_float	Qualcomm® Dragonwing™ Q-8750	0.12 ms	0 - 27 MB	NPU

License

• The license for the original implementation of Parakeet-TDT-0.6B-v3 can be found here.

References

• Canary-1B-v2 & Parakeet-TDT-0.6B-v3: Efficient and High-Performance Models for Multilingual ASR and AST
• 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.