Anyone who has watched a badly dubbed film knows the feeling. The words are technically right, but the timing is off, the emotion lands in the wrong place and the characters suddenly sound like strangers. Translation is hard. Translating the way people actually speak is much harder.
Japanese voice AI is software that can understand, speak and translate spoken Japanese, and one of the biggest names in AI audio just made a serious bet on it. On November 16, 2026, ElevenLabs announced that it is acquiring Kotoba Technologies, a Japanese voice AI lab founded in 2023. The deal is expected to close later in the fourth quarter of 2026, subject to regulatory approvals, and the price was not disclosed.
For podcasters, authors and narrators, this matters even if you have never recorded a word of Japanese. ElevenLabs' voices already sit behind a lot of AI narration, dubbing and voice tools. When a company of that size buys a team to improve one of the world's most nuanced languages, it tells us where translated audio is heading and how quickly it is getting better.
What Is ElevenLabs Buying?
ElevenLabs is buying a small research team and its voice model, Koto. According to the announcement, Koto is strong in three areas: turning Japanese text into speech, turning Japanese speech into text, and speech-to-speech interpretation that translates a speaker in close to real time.
ElevenLabs CTO and co-founder Piotr Dąbkowski says bringing Kotoba's research together with ElevenLabs' own will improve Japanese across everything the company builds. That covers more than one product. It points to better Japanese in text-to-speech voices, transcription, conversational agents and translation tools alike.
The company also laid out its next steps in Japan. ElevenLabs plans to hire local sales leaders in the coming months, keep investing in infrastructure in the country and work more closely with Japanese businesses and global companies operating there. This builds on a move it made last year, when ElevenLabs opened ElevenLabs G.K. in Tokyo on April 14, 2025, as its first international hub.
Who Is Kotoba and What Does Koto Do?
Kotoba was founded by two researchers with deep roots in language AI. CEO Noriyuki Kojima holds a PhD in computer science from Cornell, worked as a research scientist at Hugging Face and helped build Fugaku-LLM, an open-source model trained from scratch on Japan's RIKEN Fugaku supercomputer. CTO Jungo Kasai holds a PhD from the University of Washington and was an assistant professor at the Toyota Technological Institute at Chicago.
The company has been focused on live interpretation from the start. When it raised USD 11.83 million in seed funding in August 2025, Kotoba said its speech-to-speech model was approaching an average delay of under three seconds between Japanese and English. It described that quality as close to a human interpreter. Those are the company's own claims, and the acquisition announcement does not include new benchmarks.
Koto is already in commercial use. ElevenLabs says Salesforce uses it to power Japanese-language conversations in Agentforce Voice, its AI voice agent product. Salesforce Ventures has invested in both ElevenLabs and Kotoba, which helps explain how the two companies came together.
Why Is Japanese So Hard for Voice AI?
Japanese is hard for voice AI because so much of the meaning sits outside the words. ElevenLabs' announcement points to indirect speech, pauses that show someone is thinking, polite apologies that do not admit fault and praise that people deflect rather than accept. It also describes the expectation to understand what is left unsaid, often called reading the air.
A voice model can pronounce every syllable correctly and still get those moments wrong. Rush a thoughtful pause and a speaker sounds abrupt. Read a polite refusal too literally and the listener hears agreement. Flatten the tone of a humble reply and the speaker sounds cold or even rude.
That is why ElevenLabs is buying a specialist rather than simply adding more Japanese training data. Kotoba was built around Japanese from day one. For anyone translating audio into or out of Japanese, a model that understands these habits could sound far more natural than one that treats Japanese as English with different words.
What Could This Mean for Dubbed Podcasts and Translated Audiobooks?
The announcement focuses on businesses and voice agents, and it does not mention podcasts, audiobooks or dubbing. Still, improvements to a core voice stack tend to show up across a company's products over time. Creators who already use AI tools to translate their shows or narrate their books into other languages could eventually see better Japanese results.
Japan is a large and attractive audience for audio. A podcaster with a loyal following there, or an author whose book has a Japanese publisher, could one day release a translated version far faster and more cheaply than with a full recording team. As we covered in our look at how AI narration tools deliver distribution-ready files, the technology is already reshaping how audiobooks get made.
Real-time interpretation opens another door. Live events, interviews with guests who do not share a language and bilingual shows could all become easier to produce. A host in Toronto could interview a guest in Tokyo with each person speaking their own language, then release the episode in both.
Where Do Human Narrators and Interpreters Fit?
Human narrators, translators and interpreters still bring things AI struggles with, especially in a language as context-heavy as Japanese. A skilled narrator decides how a character sounds, when to slow down and where the emotion of a scene sits. A good translator decides how a joke, an idiom or a moment of politeness should land for a new audience.
Not every company is choosing AI for this work. As we explained in our post on Audible Originals translations, Audible U.K. has grown its translated editions with human translators. Fiction, memoir and comedy depend on voice and timing, and those are the places where listeners notice mistakes first.
There are also questions of consent and credit. If a narrator's voice is recreated in Japanese, or a host's interview is interpreted by AI, the people involved should know and agree. Our post on Audible's AI characters and copyright looked at how unsettled many of these rights still are.
What Should You Check Before Releasing an AI Translation?
Start with a fluent listener. Before any translated episode or audiobook goes out, have someone who speaks the language natively listen to a full chapter or segment. Ask them to flag names, numbers, honorifics and any line that sounds stiff, rude or oddly formal.
Next, check the timing. Translated speech often runs longer or shorter than the original, which can throw off music cues, sound effects and chapter markers. Listen for pauses that have been squeezed out and for moments where the translated voice talks over a sound cue that used to sit in silence.
Then check the sound itself. AI dubs can drift in tone and loudness between sections, and a voice can change slightly from one chapter to the next. Match the loudness of every language version to the original and listen on earbuds, a phone speaker and a car stereo, since that is how most people will hear it.
Finally, keep your source files clean and organized. Translation works best when the original dialogue is clean and each speaker sits on a separate track, with music and effects kept apart. Strong source audio gives every translated version a better starting point.
What This Means for Producers and Creators
As an audio producer, I see this as another sign that translated audio is moving from a big-studio project to something independent creators can realistically attempt. That is exciting. It also means more AI dubs and translated narrations will land on producers' desks with problems that only a trained ear will catch.
When I review an AI translation, I listen for the human parts of speech first. I check where the pauses fall, whether the emotion matches the scene and whether the voice keeps the same personality from start to finish. A translated file can pass every technical check and still sound wrong to a native listener, and that is where human direction makes or breaks a release.
For authors and podcasters, my advice is to treat AI translation as a draft that still needs a producer, a native listener and a final quality check. Budget for that review from the start. A careful foreign-language release can open a whole new audience, and a careless one can close the door just as fast.
What Comes Next for Japanese Voice AI?
The deal still has to close, and ElevenLabs has not said when Kotoba's research will show up in specific products. The announcement shares no new performance figures, and the latency numbers we have are Kotoba's own claims from 2025. Creators should judge the results by listening, not by press releases.
The direction is clear, though. One of the most widely used AI voice platforms is investing in a language whose meaning often lives between the words. If it works, translated podcasts and audiobooks will sound more natural to Japanese listeners, and other languages with their own unspoken rules may follow.
The fundamentals still decide whether any of it succeeds. Clean source audio, thoughtful direction and a listener who truly understands the language are what make a translation feel like it belongs. Speak clearly, listen closely and let great audio travel further!