
A black studio microphone and a dark pair of headphones inside a wood-lined recording booth. AI-generated editorial illustration.
ElevenLabs is acquiring Kotoba, a Japanese voice AI lab. Kotoba’s Koto model can make speech from text, turn speech into text, and translate live speech. ElevenLabs says Salesforce already uses it for Japanese talks in Agentforce Voice.
The plan brings a team that studies speech in Japan into a larger voice AI company. ElevenLabs announced the deal on November 16 and updated its account on November 18. The announcement gives no price and sets no date for new features in ElevenLabs products.
What Kotoba built
Noriyuki Kojima and Jungo Kasai started Kotoba in 2023. The two founders had worked on language AI and set out to build a voice system for more than one language. Kotoba describes Koto as a real-time voice model with options to run in the cloud, on a device, or on a customer’s own systems.
Those uses have different needs. A tool that reads a line of text aloud has the full line before it starts. A live interpreter may need to begin while the speaker is still talking. An agent on a call also has to know when to wait and what to do if someone starts to speak again.
Kotoba says its own app offers live translation in voice and text. In a business call, a transcript can give teams a way to check an important name or number later. In a live chat, a quick voice reply can keep the talk moving. A voice model can serve both needs.
Why Japanese voice is more than a language option
The ElevenLabs announcement points to traits such as changes in word choice based on the people speaking. It also gives the example of subjects that a speaker can leave unstated when the listener knows the context.
That can make a simple call hard to handle. A voice system may have to work out which order a caller means from what they said earlier. A pause can be a sign that the person is working out how to put their next thought. A system that fills such gaps too soon can make the call harder to follow.
For that reason, a sample of a model reading prepared lines tells a buyer only part of the story. It helps to ask how it deals with names, breaks in speech, and a person who changes their mind halfway through a request.
The quality of the reply also depends on what the system heard. When a caller says a place name or moves between languages, it is useful to see if the words in the transcript still match the spoken request.
What ElevenLabs plans to do in Japan
ElevenLabs says it plans to build out its local sales team and keep investing in infrastructure in Japan in the coming months. It names Salesforce as a current user of Koto. Those plans make the deal as much about a market and its needs as the sound of a voice.
There are other ways AI tools use speech. In our NotebookLM Audio Overview guide, a model can take time to make a show from source material. Google’s Gemini Omni work puts audio in a wider set of media. Kotoba’s main job is to keep a live exchange moving as people talk.
Read also
Popular news

AWS lets AI agents buy select tools on the developer’s cloud bill

Human clues helped AI agents mimic supply-chain decisions

Cline-Bench tests coding agents three times on real software tasks

Claude Code vs Cursor: Try the Same Fix in Both

US House lawmakers launch a bipartisan caucus on AI safety



