ByteDance’s call model has half the rhythm problems of cascaded models, company says
ByteDance’s Volcano Engine opened its SeedRealtime 1.0 model to developers on 19 November. The company says the model watches, listens and speaks at the same time, and has half as many conversation rhythm problems as cascaded models, which are systems built from components that run in sequence.
A typical voice assistant chains separate programs. Automatic speech recognition (ASR) turns speech into text, which a vision-language model (VLM) interprets. Text-to-speech (TTS) then reads the answer aloud. Volcano Engine says each intermediate step adds latency and loses information.
SeedRealtime 1.0, part of the Doubao model family, handles sound, images, timing and its own speech in a single model. The model does not just wait for questions, the company said. It follows changes in the situation and speaks up at a suitable moment.
In the company’s example, a user pours whole coffee beans into the portafilter of an espresso machine. The model reminds the user that the beans must be ground first. Once the coffee has been extracted, the model suggests a shorter extraction time, judging by the crema and the amount of coffee. The company also said the model tells speakers apart and does not react to bystanders’ chatter or background noise.
Volcano Engine said audio and video conversations have half as many rhythm problems as those with cascaded models, in which components are connected in sequence. It added that individual conversations more often run to the end without interruption. The reports do not define a rhythm problem, name the systems compared or say who ran the tests. The figure is therefore the company’s own claim.
Outlets including ithome.com (IT之家), Ifeng, 36Kr and Kuai Keji reported the news on 19 November. All of them rely on the company’s press release, so the reports do not corroborate one another, and none includes an independent test. The company also chose the example scenarios.
Developers can now use the API through Volcano Engine. The company markets it for voice assistants, smart devices, companion apps, customer service and learning apps. The reports do not mention pricing.