What I built
Turning text into speech is roughly two stages: an acoustic model that produces sound features from the text, and a vocoder that turns those features into an actual waveform. The usual setup uses an off-the-shelf part for one of them. At Livetoon we designed and trained both from scratch — because I didn't want the naturalness, or the latency, tied to someone else's choices.
Only a handful of companies in Japan, the big names included, can hold their own from-scratch speech synthesis. We took it on with a small four-person team. I worked as the lead developer, hands on everything from the architecture to how training was run. Image, speech, language — getting to experience the major generative modalities from the "training" side owes a lot to this period.
Used in the field
In October 2025, this TTS was used for the read-aloud in Fujitsu's healthcare AI-agent proof-of-concept. Someone on their side said the read-aloud "had surprisingly little of the awkwardness you usually get from synthesized speech." More than any self-reported number, that one line from someone who actually used it made the work feel worth it.
Where it was covered
- AWS Startup Blog. A technical interview about building Japanese speech synthesis from scratch.
- NVIDIA-hosted webinar. I spoke on "Realizing digital health with agentic AI" (November 2025).