| Overview of Qwen3-ASR Models |
|
|
| The Qwen3-ASR models are sophisticated automatic speech recognition (ASR) systems designed to provide high-quality transcription and language identification across a wide range of languages and dialects. They are part of the Qwen3 family, which includes two main models: Qwen3-ASR-1.7B and Qwen3-ASR-0.6B. |
| Key Features |
|
|
| Language Support: Both models support 52 languages and dialects, including 30 languages and 22 Chinese dialects. |
|
|
| Architecture: They utilize a dual-module architecture: |
| Audio Transformer (AuT) Encoder: Processes audio input to extract features. |
| Language Model (LM) Decoder: Generates text output from the processed audio features. |
|
|
| Performance: The models achieve state-of-the-art results in various challenging conditions, such as: |
| - Noisy environments |
| - Singing voice recognition |
|
|
| Model Specifications |
| Model Name Parameters Supported Languages Special Features |
| Qwen3-ASR-1.7B 1.7 billion 52 High accuracy, robust in complex settings |
| Qwen3-ASR-0.6B 0.6 billion 52 Efficient for lower resource environments |
|
|
| Inference Capabilities |
| - Streaming and Offline Inference: Both models can handle real-time audio processing and can also work with pre-recorded audio. |
| - Timestamp Prediction: The models can provide timestamps for words and sentences, which is useful for applications like subtitle generation. |
