Kotoba: Developing Voice AI for a Multilingual World

Our Investment in Kotoba

East Asia represents 1.6 billion people and 20% of the global population. Roughly half of the world's knowledge workers speak an Asian language, yet the region remains deeply underserved by Western AI. With hundreds of distinct languages, each carrying its own linguistic nuances and data characteristics, building for East Asia requires far more than translation: it demands a ground-up approach to model training and market expertise.

The Founders

Noriyuki Kojima and Jungo Kasai first met as PhD students - Noriyuki at Cornell, Jungo at the University of Washington - both working on problems at the edge of natural language processing, LLMs, and speech. What kept surfacing was the same frustration: the most spoken languages in East Asia were being treated as secondary targets by the models built to serve them.

Japanese, Korean, and Chinese are structurally different from Western languages. Each carries distinct phonologies, honorific registers, and writing systems that require fundamentally different modeling decisions. Yet nearly every major speech AI on the market was built English-first and adapted outward. The performance gap was real, and neither founder could find a convincing effort to close it. So they built one.

The Investment

Today, we're thrilled to announce that we are partnering with Kotoba Technologies and our lead role in their $10 million seed round, alongside Salesforce Ventures and Sony Ventures.

The Kotoba research team brings exceptional focus and depth to two distinct problems: high-controllability speech pipelines for AI agents, and ultra-fast native speech-to-speech models for real-time communication and translation. Across recognition and synthesis, the Koto model family - TTS, ASR, and Speech-to-Speech - outperforms models from North American and European labs on Japanese, Korean, and Chinese. We were also drawn to a shared conviction: that inference will increasingly run on hybrid compute architectures, where smaller on-device models work in tandem with large cloud models rather than replacing them.

Noriyuki and Jungo are building one of the leading speech AI research labs in Asia. We're honored to support their mission to bring state-of-the-art speech models, multimodal agents, voice-centric wearables, physical AI hardware, and real-time translation to a global audience.

The Costly Assumption Baked Into Every Major Speech Model

Every major speech AI system built over the last decade started from the same place: English. Not because English speakers represent the largest market, or because English phonology is the most complex, or because English data is the richest proxy for human communication at large. English came first because the researchers, the compute, the benchmark datasets, and the venture capital were concentrated in places where English is the dominant language. The assumption was never stated explicitly. It didn't need to be.

The cost of that assumption is now visible. Japanese, Korean, and Mandarin are not dialects of English with different vocabularies. They carry distinct tonal systems, honorific registers, writing systems, and grammatical structures that require different architectural decisions from the ground up - not a fine-tuning layer applied after the fact. A speech model performing at 95% accuracy in English can fall well below enterprise threshold in Mandarin or Japanese, particularly in low-latency, on-device contexts where there is no large cloud inference stack to catch the errors.

The workarounds most enterprises deploy today - larger context windows, aggressive post-processing, custom fine-tuning - add latency and engineering cost without closing the underlying gap. They are patches on a foundation that was never designed for the languages being asked of it.

The market bearing the cost of that assumption is not small. East Asia represents some of the highest-density concentrations of enterprise AI adoption in the world. Voice agents, simultaneous translation, and on-device ASR have moved from experimental to critical infrastructure for companies operating across Japan, Korea, and China. The demand is structural, and it's accelerating. What has been missing is not appetite, but rather, it's a model family built for these languages from first principles. Kotoba was built to close exactly the gap the industry created.

What Kotoba Built

Kotoba's proprietary model, Koto, is purpose-built for real-time speech applications with industry-leading performance in Japanese, Korean, and Chinese. It handles the full suite of modern speech AI use cases: AI voice agents, smart hardware devices, multimodal models, and simultaneous speech translation.

Language, Precision, and Presence

Kindred has been watching the convergence of speech AI and East Asian markets for several years. We invested early in PlayAI, backing their work on turn-taking dialogue models and diffusion-based speech architectures well before the broader market recognized the opportunity - work that led to Meta's acquisition in 2025. We've spent considerable time with founders and operators in Japan, Korea, and across the region, and have seen firsthand how acute the need is for speech AI built to each language's actual requirements, not adapted to them after the fact.

Three things brought us to Kotoba specifically, and rarely appear together.

First, state-of-the-art ASR and TTS models demonstrably outperforming other labs on the languages Koto was built for. Kotoba's performance in Japanese, Korean, and Chinese reflects architectural choices made from first principles - not a fine-tuning layer applied to a general-purpose base - and achieves top scores on both latency and quality benchmarks.

Second, immediate customer traction with AI-native platforms and large enterprise customers following our initial introductions.

Third, a team with the research depth to develop speech-to-speech models that are faster and more contextually fluent, as well as edge models optimized for the latency and cost requirements of robotics, autonomy, and wearables. Noriyuki and Jungo bring frontier expertise in NLP and speech alongside the cultural and commercial fluency to operate across Japan, Korea, China, and the US. Technical depth combined with cross-market fluency is exactly what the longer-term global opportunity requires.

The new funding will go toward three priorities: deepening the S2S model family for AI agents and smart devices; scaling on-device deployments across automobiles, electronics, and AI wearables through partnerships in Asia and the US; and accelerating the enterprise and developer rollout globally, including the API platform, SDK, and forward-deployment efforts for global platforms.

We're excited to partner with Noriyuki, Jungo, and the Kotoba team as they build the voice AI platform for East Asia and beyond. Learn more about Kotoba.