A language the world skipped
Hmong is spoken by millions of people — yet no major platform offers a real Hmong voice or a Hmong speech recognizer. Text translation exists; spoken Hmong does not. For a Hmong elder who cannot read a screen, that gap is the whole barrier.
Big tech builds for the largest markets first. Small and heritage languages wait — often forever. We decided not to wait. If the tools our community needed didn't exist, we would build them, and we would own every piece.
Own the whole stack
Rather than renting a black box from a platform that could change or disappear, Silver Inc built and self-hosts its own translation, speech-synthesis, and speech-recognition models. We rent only commodity compute — never the intelligence itself. The models, the training data, and the pipeline are ours.
iSpeak — the engine
Real-time translation and natural speech across 100+ languages, with a dedicated premium path for Hmong. Every product in the ecosystem runs on it.
Paj Hlub — the voice
A neural Hmong text-to-speech voice, trained from a hand-curated corpus of Hmong speech. Named after our founder's daughter, the human translator behind it.
The Hmong ear — ASR
A custom speech-recognition model fine-tuned to understand spoken Hmong, so the app can listen as well as speak — self-hosted on our own servers.
Pronunciation & tone
Hmong is tonal and phonetically rich. We built tone-normalization and pronunciation rules so the voice says words the way a native speaker actually would.
From "impossible" to shipping
Through relentless iteration and data augmentation, our Hmong speech recognizer's word-error rate fell from double digits to low single digits — accurate enough to power live conversation. The Hmong voice went from unintelligible early attempts to a voice families recognize and trust, now live on Google Play and the App Store.
The road here was not short
None of this came from a single breakthrough. It came from showing up, again and again, when the easy answer was to quit.
- Curating the data by hand. No Hmong speech dataset existed, so we built one — thousands of sentences recorded, cleaned, and aligned, one at a time.
- Dozens of model iterations. Training runs that failed, overfit, or crashed mid-way. We rebuilt the pipeline to be crash-resilient and kept going.
- Teaching the machine to hear. Speech recognition for a tonal language with little data is genuinely hard. Augmentation and careful tuning pulled the error rate down, run after run.
- Owning it end to end. We moved everything onto our own infrastructure so no outside platform could ever take the voice away from the people who need it.
- Still going. Next: more underserved languages, and a Hmong singing voice — because a language deserves to sing, not just speak.
What the giants won't build for a small language, we build for our own — then open it to the world.