Kotoba
Frontier speech AI · Tokyo & SF

Ever worked with someone who doesn’t speak your language natively?

Kotoba ships AI speech translation and more

English 日本語 한국어 中文
QR code — submit a question
Scan to ask
Kotoba
Products

We ship our own frontier speech models

For teams
Meeting tool
Zoom, Google Meet, etc.
For everyone
Mobile App
Live translator in your pocket
iOS · Android
For builders
Koto models
Frontier ASR, TTS, and STS translation
QR code — submit a question
Scan to ask
2 / 9
Kotoba
Some of the team

Research to app

Yuta Sasaki
Yuta Sasaki
Founding ML Engineer
Qian Cheng
Qian Cheng
Machine Learning Engineer
Jaeeon Park
Jaeeon Park
NLP Researcher
Justin Garcia
Justin Garcia
Full Stack Application Engineer
QR code — submit a question
Scan to ask
3 / 9
Kotoba
How we train · Qian Cheng

How we train Koto:

1
Data
Human speech audio
2
Pretrain
Model learns speech patterns
3
Fine-tune
Paired examples teach translation
4
Align & evaluate
Check accuracy and latency
QR code — submit a question
Scan to ask
4 / 9
Kotoba
How inference works · Yuta Sasaki

Conventional speech translation

Step 1
Audio processing
Clean up the input audio
→
Step 2
Speech recognition
Speech → text
→
Step 3
Text translation
Text → text
→
Step 4
Speech synthesis
Text → speech
Four models in a chain: latency adds up and each module’s errors propagate downstream
QR code — submit a question
Scan to ask
5 / 9
Kotoba
How inference works · Yuta Sasaki

How Koto works: full-duplex

Input
Audio processing
Listens while the speaker is still talking
→
One model
Output
Translated speech
Speaks while it is still listening
Low latency and high accuracy
QR code — submit a question
Scan to ask
6 / 9
Kotoba
Building products • Justin Garcia

Customer feedback to shipping

Gather feedback from customers
Prioritize
Build, test and ship
QR code — submit a question
Scan to ask
7 / 9
Kotoba
Working at Kotoba · Jaeeon Park

Real conversations shape our research

Look beyond benchmark scores
Translation quality, latency, and conversational flow.
Discover failure modes through dogfooding
Use the product ourselves to see where conversations break down.
Turn observations into testable hypotheses
Design controlled experiments and use the evidence to choose the next step.
QR code — submit a question
Scan to ask
8 / 9
Kotoba
Careers

We’re hiring

AI Researcher San Francisco
Speech Edge AI Researcher Tokyo & SF
Full-Stack Engineer Tokyo
Infrastructure Engineer Tokyo & SF
Technical Account Manager Tokyo & SF
More GTM · Comms · Partnerships
site.kotoba.tech/company#careers Talk to us after ⟶
QR code — submit a question
Scan to ask
9 / 9