Fun-ASR-Realtime: real-time ASR is finally being tested where accents, noise, and product pressure live

The easy demo is a clean Mandarin sentence spoken close to a microphone. The product test is messier: a meeting room, two people talking over each other, a project codename, a regional accent, a keyboard in the background, and a user staring at captions that need to appear now, not after the recording ends.

That is the lens through which Fun-ASR-Realtime should be read. Alibaba’s Model Studio documentation exposes fun-asr-realtime as a WebSocket-based real-time speech recognition service. Audio streams in. Punctuated text streams out. The service is positioned for live captions, online meetings, voice chat, voice assistants, and other “while the user is still speaking” scenarios.

The current stable model is listed as equivalent to fun-asr-realtime-2025-11-07, while the Beijing region also lists fun-asr-realtime-2026-02-28 as a newer snapshot. The model-selection page recommends Fun-ASR for real-time recognition when hotwords and dialect support matter.

What changed from the earlier Fun-ASR realtime line

The cleanest comparison in the public docs is against fun-asr-realtime-2025-09-15. That earlier snapshot is listed for Mandarin and English. The current main Fun-ASR-Realtime line expands the practical target: Chinese, English, Japanese, and Chinese dialects or accents. The Chinese coverage named in the docs includes Mandarin, Cantonese, Wu, Minnan, Hakka, Gan, Xiang, Jin, plus regional Mandarin accent belts such as Zhongyuan, Southwest, Jilu, Jianghuai, Lan-Yin, Jiaoliao, Northeast, Beijing, Hong Kong, and Taiwan.

The public announcement numbers go further: 16 dialects, 30 languages, average character accuracy of 88.62% across the dialect test, Shanghai dialect at 92.41%, Wenzhou dialect at 82.74%, and about a 20% accuracy lift for selected East and Southeast Asian language scenarios such as Thai. Those numbers matter because accent and dialect are where many “good enough” ASR systems stop being good enough.

The other upgrade is control. The real-time documentation says Fun-ASR supports hotwords, and that fun-asr-realtime plus fun-asr-realtime-2025-11-07 support context enhancement. In plain product language: you can feed domain terms, conversation history, or expected vocabulary into the recognizer so that names, drugs, car models, school terms, and internal project names are less likely to be mangled.

Compared with SenseVoice, it is stronger in a different place

SenseVoiceSmall remains one of the more interesting open models in the FunASR ecosystem. Its model card is very clear about the pitch: multilingual ASR, language identification, speech emotion recognition, and audio event detection. It claims support for more than 50 languages, uses a non-autoregressive framework, and reports about 70ms to process 10 seconds of audio. The GGUF version is especially practical: roughly 235MB for Q8, 470MB for F16, and 936MB for the F32 reference file.

Fun-ASR-Realtime is not trying to win the same contest. Its advantage is the production realtime path: WebSocket streaming, interim and final results, timestamps, hotwords, multiple audio formats, client reconnection, heartbeat support, and SDK access. The official model-selection page also says Fun-ASR models support Android and iOS SDK access, which matters when the phone is the microphone and the recognizer is a cloud service.

So the choice is not “which model is better” in the abstract. If you need local speech understanding with emotion and event tags, SenseVoiceSmall is still a strong candidate. If you are building live dictation, customer-service transcription, classroom captions, meeting assistants, or in-car speech input, Fun-ASR-Realtime is closer to a product API than a research checkpoint.

How large is the model file?

For the cloud model, there is no public downloadable checkpoint named Fun-ASR-Realtime on Hugging Face from the API search I checked. Alibaba exposes it as a hosted Model Studio service. That means the strict answer is: the cloud model’s actual serving weight size is not public.

There are useful nearby numbers from the open FunASR family. FunAudioLLM/Fun-ASR-Nano-2512 is about 1.897GB in the Hugging Face repository, with model.pt at about 1.880GB. Fun-ASR-MLT-Nano-2512 is almost the same size. The GGUF path splits the payload into a 447.59MB encoder F16 file plus a Qwen3 0.6B file: about 461.79MB for Q4_K_M, 525.83MB for Q5_K_M, or 767.47MB for Q8_0.

Those numbers are only the model payload. A shipping app still needs audio capture, VAD, decoding, tokenization, network logic, caching, error handling, and product UI. For mobile, that difference matters.

Can it run on a computer or a phone?

On a computer, yes, in two different ways. The hosted service can be called from macOS, Windows, or Linux as long as the app can open a WebSocket connection and authenticate with a Model Studio API key. If you want local inference, the open Fun-ASR-Nano and SenseVoiceSmall paths are more relevant. GPU Python is the straightforward route; GGUF plus the FunASR llama.cpp runtime is the lighter CPU and edge route.

On a phone, the practical answer depends on what “run” means. As a client, yes: the official docs mention Android and iOS SDK access for Fun-ASR models. The phone records audio, streams it to the service, and renders the transcript. As a fully offline local model, it is possible in principle through smaller GGUF or converted packages, but it is not a casual consumer-app first-install decision. A 1GB-class ASR component changes download size, storage, thermals, battery life, and device support. It makes more sense as an optional offline pack, an enterprise module, or a dedicated device feature.

The real signal

The important shift is not that another ASR model exists. It is that real-time ASR is being pulled out of the clean-demo world and pushed toward dialects, noisy rooms, hotwords, long-lived connections, and SDK integration. That is where speech recognition becomes infrastructure. The best version of this model will be the one users barely notice: they speak, the text appears, and the system does not panic when the accent changes.

Sources

• Alibaba Cloud Model Studio: Real-time speech recognition user guide, updated July 6, 2026.
• Alibaba Cloud Model Studio: ASR model-selection page, updated July 1, 2026.
• Hugging Face: FunAudioLLM/Fun-ASR-Nano-2512, Fun-ASR-MLT-Nano-2512, Fun-ASR-Nano-GGUF.
• Hugging Face: FunAudioLLM/SenseVoiceSmall and SenseVoiceSmall-GGUF.