Your online real-time transcribe demo with is not working
Last spoken word is not rendered to the live transcription on your demo site: https://r2t2.youdao.com/demo
Thanks for pointing this out! This is actually intentional behavior rather than a bug.
To ensure higher recognition quality, the most recently predicted token (which may be a word or punctuation mark) is kept in a provisional state and is therefore not immediately shown to the user. This is part of R2T2’s true streaming, append-only design: once a token is committed and displayed, it should never need to be revised.
From an engineering perspective, this final provisional token can be committed based on different criteria—for example, after detecting a certain period of silence. We simply haven’t added such a finalization rule to the current online demo.
If you’d like different behavior, feel free to build your own demo using our codebase:
https://github.com/netease-youdao/Confucius4-R2T2
You can then decide when to display or commit the final provisional token according to the requirements of your application.
we will fix it soon !
Ok, I thought it was a bug - so I didn't bother trying the model.
Maybe something to consider, make it render every word.
@kipdg Thanks for the feedback — we agree that seeing the final word is important from a user-experience perspective.
There is a small trade-off here: immediately committing every predicted word would lead to a slight decrease in recognition accuracy — on average, less than 1 WER point across the 14 datasets we evaluated. This is quite natural with a 160 ms streaming chunk size, especially in ambiguous situations where a little more acoustic context can help the model make a better decision. You can see this trade-off in our evaluation results under Quality Priority vs. Latency Priority:
https://github.com/netease-youdao/Confucius4-R2T2#Evaluation
That said, based on this feedback, we’ve updated the demo at https://r2t2.youdao.com/demo with a simple finalization mechanism: when no new text is produced for a certain period of time, the remaining provisional content is automatically committed and displayed. This allows us to preserve the append-only streaming behavior while making the experience more natural when speech ends. Feel free to give the demo another try — we’d be very happy to hear any further feedback!
The online demo is intentionally just a simple example for visualizing R2T2 in action and getting a first impression of its capabilities, and we strongly encourage users to adapt the streaming behavior to their own requirements — for example, by changing the chunk size, the number of tokens kept in the provisional state, or the finalization strategy — and would be very happy to see people build their own downstream applications on top of R2T2 and explore different latency/quality trade-offs.