TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #112

Transcribes speech live, committing each word as soon as it is heard

Every streaming transcriber cheats a little: it shows you a guess, then quietly rewrites it once more audio arrives. NetEase Youdao built one that refuses to, and it is fast enough not to need the trick.

MODELConfucius4-R2T2
PUBLISHEDSeptember 21, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYAUDIO

Field notes

01What it does

You feed it audio in small slices, anywhere from 80 milliseconds to two seconds long, and it returns text for each slice as it goes. The unusual promise is that the text is final: what it has printed stays printed. Most live transcribers show a draft and fix it a second later, which is fine for a person reading captions but breaks anything downstream, like a translator or an agent that already acted on the first version. This model only emits a word once it has decided that word is stable, and holds the rest back until the next slice.

The dial you control is slice size. At 160 milliseconds you get near-instant captions with a few more errors; at two seconds you get accuracy close to a batch transcription of the whole file. Youdao's own tests put the 160 ms setting close to the full offline Qwen3-ASR it was built from on English and Chinese sets, and ahead of other open streaming models at a similar delay, with typical latency of 200 to 600 milliseconds. You can hand it a list of names or jargon as a hint, and it covers some thirty languages, with Chinese and English as the strong ones.

Under the hood it is a 2B-parameter fine-tune of Qwen3-ASR trained to recognize the longest stretch of text it can safely commit to. It ships with a vLLM backend for throughput, a plain transformers backend, and a ready-made WebSocket server that streams the incremental text to any client.

02How to try it

The Hugging Face Space is the quick look: upload a clip or record from your microphone, optionally force the language or add hot words, and read the transcript. That Space runs the one-shot offline mode, so it shows the accuracy but not the streaming. For the live behaviour, the model page embeds a side-by-side video against GPT's live transcription, and Youdao hosts its own demo at r2t2.youdao.com. To run streaming yourself, clone the GitHub repo and use the example script on any wav file; it needs a GPU and vLLM, and a Docker image is provided. First test: record yourself reading a paragraph with a few names in it, run it once at 160 ms and once at 1000 ms, and watch whether any printed word ever changes. Then pass the names as context and check they come out spelled right.

03Caveat

The code is Apache 2.0, but the weights use NetEase's own license: free to use, commercially included, unless your company passes 100 million monthly users or RMB 1 billion in revenue, in which case you need a separate agreement. Streaming also needs a GPU and vLLM; the Hugging Face Space only shows the offline path.

04What you can do with it

  • Live captions for a talk or a stream that never flicker or rewrite.
  • Feed a translation model or an agent text it can act on the moment it appears.
  • Meeting notes that build up sentence by sentence, with names spelled right via hot words.
  • Voice commands where the first word matters and waiting two seconds is too long.
  • Subtitle Chinese and English speech with one model.

Try the demo

View model page

Read this issue on neuralexpedition.com →