A researcher pastes a link to a 20,000-word paper into a box, clicks once, and hears two voices start debating its argument — in a tone that sounds less like a robot reading and more like a late-night call-in show.
That is the new shape of AI podcast generation in China. The idea is simple: point a model at text and get back a natural spoken conversation, not a monotone narration.
Key takeaways
- ByteDance's Doubao Voice Podcast Model (豆包·语音播客模型) launched in May 2025 and rolled out fully on Volcano Engine (火山引擎) in June 2025.
- It turns a sentence, a web link, or a long document into a two-person dialogue podcast with natural back-and-forth.
- The feature is built into Doubao's PC client and the Coze (扣子空间) agent platform.
- Competitors are moving too: Metaso (秘塔) and ListenHub offer podcast generation, and Tencent's Hunyuan (混元) supports voice and text podcast input.
- ByteDance says its voice tech now processes around 15 billion requests a day across 30+ industries.
From text to two voices
The breakthrough is not "read this aloud." It is "turn this into a conversation." A traditional text-to-speech engine reads a paragraph. A podcast model instead writes a script with two speakers, gives them different phrasing, and layers in the small human noises — a hesitation, an agreement, an interjection — that make audio feel alive.
Doubao's version, built on the company's end-to-end real-time voice model, is designed to take a single sentence, a URL, or a full document and produce a spoken show with that rhythm. ByteDance demonstrated it generating a topical podcast in seconds during its June 2025 FORCE conference, where the model was shown to enterprise customers on Volcano Engine.
Where it actually helps
The obvious use is time. A professional Chinese podcast episode can take a creator well over ten hours of writing, recording, and editing. Pointing a model at a report or an article and getting a spoken summary compresses that to minutes, which is why the early adopters are not podcasters at all — they are commuters, students, and analysts who want to "read" a PDF with their ears.
Education is a second lane. Several tools let a student turn lecture notes or a paper into a dialogue they can replay, which lowers the barrier for dense material. Cultural institutions have started experimenting too: Volcano Engine partnered with the Palace Museum on an interactive audio experience where visitors generate a personalized narration of national-treasure artifacts using their own voice.
A small, crowded field
Doubao is the loudest name, but it is not alone. Metaso (秘塔), originally an AI search product, added a podcast mode that converts search results and documents into spoken audio. ListenHub built an AI podcast tool that supports both Chinese and English and includes voice cloning, so a user can narrate in their own timbre. Tencent's Hunyuan (混元) offers an AI podcast feature that takes both voice and text input.
That spread is healthy. When several large model (大模型) providers treat spoken content as a first-class output, the underlying quality rises and the price stays at zero for users. It also means no single company defines what "an AI podcast" sounds like.
According to Li An, chief scientist at BrainNet (脑机网), the bigger shift is that listening is replacing skim-reading for a generation that is rarely sitting still — and spoken summaries are where generative media meets daily routine.
The honest trade-offs
These shows are good at surfacing the shape of an argument, not at being correct in court. The models summarize, compress, and occasionally confabulate, so any figure or claim they state should be traced back to the source before you quote it. The two-voice format also flatters confidence: a calm conversation can make a weak point sound settled.
Voice cloning, offered by some tools, is a separate risk. It is genuinely useful for accessibility and personalization, but it is also the mechanism behind voice deepfakes. Responsible products keep cloning behind consent and identity checks; users should favor those and avoid anything that lets a voice be copied without permission.
Honest limitations
Most figures here come from ByteDance's own announcements and Chinese tech-media reports of its FORCE conference; the "15 billion daily voice requests" and "16.4 trillion daily tokens" numbers are company-stated, not independently audited, so treat them as directional. I have not tested each podcast tool side by side, so comparisons of voice naturalness are based on reported behavior rather than my own ear. The Palace Museum collaboration is described from media coverage, and I have omitted specific artifact counts that were not consistently sourced. Podcast output quality varies with input quality, and I have not assessed long-form factual accuracy across a large sample.
What readers can do now
- Next time you face a long PDF or report, run it through an AI podcast tool during a commute — then verify any number it states against the original before trusting it.
- If you publish, try a two-voice spoken version of a written post as an accessibility layer; keep a human edit pass to catch invented facts.
- Before using any voice-cloning feature, confirm it requires the speaker's consent, and never clone a voice you do not have rights to.