Qwen3‑ASR vs FunASR: In‑Depth Technical Comparison

Overview

The author compares two popular open‑source speech‑recognition (ASR) projects—FunASR, an industrial‑grade toolkit, and Qwen3‑ASR, a multimodal large‑model‑based system—clarifying that they are developed by separate teams within Alibaba’s Tongyi Lab.

Teams and Leadership

FunASR is maintained by the Tongyi Speech Team (formerly DAMO Academy Speech Lab) under lead Li Xiangang, a veteran with over a decade of speech experience. The project lives in the alibaba-damo-academy GitHub organization and follows an engineering‑first, production‑ready philosophy.

Qwen3‑ASR belongs to the Qwen (千问) large‑model team, originally led by Lin Junyang, a young P10 researcher who built the world‑leading Qwen series. Its code and weights are hosted under the QwenLM GitHub organization.

Technical Approaches

FunASR uses Paraformer , a non‑autoregressive Transformer designed specifically for ASR. It decodes an entire sentence in one step, delivering inference speeds several times faster than autoregressive models, and includes a full toolchain (VAD, punctuation restoration, speaker diarization, hot‑word customization). The model is trained on 60 k hours of Chinese industrial data to excel in customer‑service, meeting, and recording scenarios.

Qwen3‑ASR treats speech as just another modality of the Qwen3‑Omni multimodal large model. Audio is tokenized and fed into the same semantic engine that handles text and images, giving it native multilingual ability (52 languages and dialects, including 22 Chinese dialects), word‑level forced alignment with < 50 ms error, and unified streaming/offline inference.

Core Differences (summarized)

Positioning : FunASR – industrial‑grade full‑stack ASR toolkit; Qwen3‑ASR – multimodal large‑model ASR foundation.

Core model : Paraformer (non‑autoregressive) vs Qwen3‑Omni (multimodal LLM).

Language coverage : FunASR focuses on Chinese with 31 languages; Qwen3‑ASR supports 52 languages and 22 Chinese dialects.

Streaming : FunASR requires a separate streaming model; Qwen3‑ASR offers streaming and offline in a single model.

Timestamp precision : FunASR provides frame‑level alignment; Qwen3‑ASR offers word‑level alignment with <50 ms error.

Hot‑word customization : FunASR gives industrial‑grade, weight‑adjustable hot‑word support; Qwen3‑ASR relies on the base model’s semantic ability.

Long‑audio handling : FunASR processes hour‑long recordings; Qwen3‑ASR caps at 20 minutes, targeting short video or dialogue.

Inference speed : FunASR’s non‑autoregressive decoding is extremely fast; Qwen3‑ASR uses hybrid decoding balancing speed and accuracy.

Deployment : FunASR offers a complete toolchain and private‑deployment friendliness; Qwen3‑ASR runs on consumer‑grade GPUs with a lightweight model.

Chinese WER : FunASR ~1.94 % on Aishell‑1; Qwen3‑ASR ~1.6 % on Wenetspeech.

Scenario Recommendations

Choose FunASR for Chinese‑centric industrial deployments, heavy customization needs, ultra‑low latency real‑time transcription, or processing hour‑long audio on edge devices (its 330 M model runs comfortably on modest hardware).

Choose Qwen3‑ASR when multilingual or dialect support is critical, precise word‑level timestamps are required for subtitles, deep integration with large language models is desired, or complex audio tasks such as singing recognition or accented speech need the broader semantic power of a multimodal LLM.

Maintenance and Stability

Both projects are independent; FunASR’s speech team has remained stable, and Qwen3‑ASR’s code and weights are already open‑sourced on GitHub, ensuring community continuity even after leadership changes.

Conclusion

The tools are complementary rather than competing: FunASR delivers a stable, fast, production‑ready ASR stack, while Qwen3‑ASR provides a versatile, multilingual foundation powered by a large multimodal model. Selecting the right one depends on whether the priority is industrial robustness or universal, model‑driven flexibility.