A seller in Guangzhou films a phone case for forty seconds, then needs the same clip in Spanish, Portuguese and Arabic before Friday. Two years ago that meant a dubbing studio, one native speaker per language, and a week of quotes. Now it means right-clicking a clip in Filmora and picking a target language from a dropdown.
The collapse of that workflow into a menu item is real, and it is happening across the industry at once. Whether it is a good business is a different question, and Wondershare's half-year report is the right place to ask it.
Key takeaways
- Filmora's AI Translation accepts 23 source languages and outputs 14 languages for combined voice-and-caption translation, per Wondershare's documentation.
- Lip-sync dubbing is limited to clips of five minutes or less, and voice cloning requires recording a consent sample before the model will copy a voice.
- Wondershare (万兴科技, 300624.SZ) reported first-half 2026 revenue of RMB 706.1 million (≈ US$99.4 million / HK$776.7 million), down 7.09% year on year.
- Its net loss attributable to shareholders widened to RMB 69.0 million (≈ US$9.7 million / HK$75.9 million) from RMB 52.8 million (≈ US$7.4 million / HK$58.1 million).
- AI server calls passed 900 million in the half, up more than 80%, with average daily token consumption above 100 billion in June, the company said.
What one right-click buys
Filmora's translation panel is deliberately unglamorous. You choose "translate voice and caption" or subtitle-only, set a source language, set a target, and optionally enable lip sync. The output arrives as a new track on the timeline: cloned voice, translated subtitles, both aligned to the original performance. Subtitles can be exported as an SRT and edited line by line.
The language asymmetry is the interesting detail. Voice-and-caption translation reads 23 source languages but writes 14. Everything else falls back to subtitles. That gap tells you where the model is confident and where it is not — and it means a Vietnamese or Thai creator can translate a video into English, but a Chinese creator cannot yet get a convincing Thai dub out of the same button.
The compliance scaffolding is also visible in the product. Voice cloning requires the user to read a consent script on the record. Generated assets are labelled as AI-made, and exports may carry a visible or invisible watermark. These are not marketing choices; they are what Chinese labelling rules and platform policy now require from a tool that can put words in a person's mouth.
For a small merchant, that is enough. For a distributor of microdrama (短剧) shipping forty titles a month, it is a starting point, not a finish line.
The report that complicates the demo
The product demo is clean. The financials are not.
Revenue fell 7.09% in the first half while the company's AI footprint expanded. Server calls for AI features rose more than 80% year on year to over 900 million, and by June the platform was consuming more than 100 billion tokens a day on average. Those two lines point in opposite directions: more inference, less income.
There are innocent readings. Wondershare sells several products, and a decline in one line does not indict the AI features. Token consumption is an input metric, not a revenue metric, and a company pushing hard into AI may deliberately subsidise usage to build habit. The company itself frames the shift toward token-based revenue models as a future driver, which is an admission that today's subscription pricing does not yet capture what heavy AI use costs.
Still, the arithmetic is uncomfortable. If a subscriber pays a flat annual fee and then dubs twenty videos a month, the marginal cost of that subscriber rises with every minute of generated speech. A dubbing minute is not a stored file; it is compute. Flat pricing plus metered cost is the shape of a margin problem, and the half-year loss is where it surfaces.
Dubbing is a bad place to compete on price
The buyers of AI dubbing are unusually price-sensitive: cross-border sellers, microdrama exporters, training teams, small media outlets. They compare on cost per minute and switch on a ten-percent difference. That is not a market where a premium tool wins on quality alone.
It is also crowded. iFlytek sells a dedicated dubbing line aimed at media and microdrama clients. ByteDance's voice cloning sits inside its cloud API. CapCut bundles translation into an editor most users already have. HeyGen targets the high end with better lip sync and a higher price.
In that field, standalone dubbing is being priced toward zero, and the durable value sits on either side of it: on the distribution side, where the platform owns the shelf the dubbed video lands on, or on the rights side, where a licensed voice is worth more than a cloned one. Wondershare owns neither. It owns an editor and a subscription.
Honest limitations
This piece relies on Wondershare's 2026 half-year report filed with the Shenzhen Stock Exchange, half-year coverage from Chinese financial media, and Wondershare's own product documentation. The company does not break out AI-related revenue, so any claim linking token consumption to profitability is inference, not disclosure. The five-minute lip-sync limit and the 23/14 language split are current as documented and may change with releases; we did not test output quality across languages for this article. Currency conversions use RMB 7.1 = US$1 and RMB 1 = HK$1.1.
What readers can do now
- Check the target-language list before committing a workflow; a tool that reads 23 languages may only speak 14.
- Keep a consent record for every cloned voice, because the platform keeps one and a dispute will ask for yours.
- Model dubbing cost per finished minute rather than per seat, especially if your volume is seasonal.
- Treat "AI usage up" and "revenue up" as separate claims; in this sector they have been moving in opposite directions.