Qwen3 Omni Jp Vllm

Qwen3 Omni Jp Vllm is a Japanese-language model for automatic speech recognition (ASR) and speaker diarization, utilizing the Qwen3-Omni-30B-A3B-Instruct weights and operating with a vLLM server runtime. It is designed to transcribe Japanese audio while distinguishing between speakers in a single pass, assigning speaker labels such as [spk_0] and [spk_1] accurately. The model demonstrates strong performance in recognizing proper nouns, including customer names, company names, and branch names, as well as numerical data like stock codes, in Japanese telephone conversations.

The tool distributes a setup and server implementation that leverages the original Qwen3-Omni-30B-A3B-Instruct model weights without additional training. It is delivered as a server-based solution using vLLM, which incorporates fused MoE kernel, PagedAttention, and CUDA graph technologies. This setup achieves significant speed improvements over standard transformers implementations, with reported performance of transcribing 111 seconds of telephone audio and generating 200 tokens in 2.1 seconds (approximately 95 tokens per second), making it about 6.5 times faster than the standard approach in the measured environment (NVIDIA A100 80GB, bf16).

Users can interact with the model by uploading audio via a browser or by running transcription from the command line for individual files. Qwen3 Omni Jp Vllm is intended for applications requiring Japanese ASR with speaker separation, particularly in scenarios involving telephone conversations where accurate identification of speakers and specialized terms is important.