The AI Giants Are Narrowing Their Focus. Huri Built Speech AI for the Languages They Leave Behind.

September 2026 was supposed to be the month the AGI era began. OpenAI launched GPT-6 Astra, its president said the company was entering an era of artificial general intelligence, and Nvidia's chief executive said AGI had arrived.

The same month, IBM released its new Granite Speech 5.0 transcription models. They are small, fast and English-only. IBM's earlier speech models handled several languages and speech translation. This generation drops that in exchange for speed. By IBM's own figures, it transcribes over three and a half hours of English audio every second on a single high-end GPU.

So in the same month, one headline promised a machine that can do everything, and another showed a technology leader narrowing to the language with the most data and the most customers. For anyone who works with Pacific communities, the second one is the more honest guide to where the industry is heading.

Independent evaluators lost no time testing Astra on translation. It topped several rankings. On a benchmark built from difficult, real-world cases, though, the best configuration passed only about 45% of them, and it remained well below human performance. Evaluators also found it comparatively weak in Polish, Japanese and Thai, and in one test its Finnish was worse than an older OpenAI model's.

Polish, Japanese and Thai have tens of millions of speakers and some of the richest digital archives in the world. If a model billed as general intelligence still stumbles there, nobody should expect it to handle a Tongan health notice, a Chuukese lease or a Tahitian radio interview.

Washington has its own fix. At the end of September, a presidential order renamed AI Super Intelligence across federal agencies. In Pacific languages, it still sounds more like superficial intelligence.

None of this is a failure of engineering. It's a matter of economics. General-purpose AI is built where data is plentiful and demand is measured in millions of users. That logic produces impressive results in English, Spanish and Mandarin. It's also why the most widely used open speech recognition model lists 100 languages and only two Polynesian ones, Māori and Hawaiian. It's why the largest multilingual speech project of recent years reached a thousand languages largely through recordings of scripture.

The tools that dominate the industry news cycle (AI dubbing for streaming platforms, enterprise voice agents, live translation in the world's major languages) are excellent at what they were built for. Pacific languages are not what they were built for.

Even in English, general-purpose speech AI still makes serious mistakes. A September 2026 benchmark tested seven leading commercial transcription engines on more than 1,400 physician dictations. Reviewers rated 191 of the errors potentially fatal, most of them garbled drug names. That's in English, the best-resourced language in speech AI.

Now picture a vaccination reminder in Samoan, an evacuation notice in Fijian or a consent form read aloud in CHamoru, handled by a model that has barely heard the language. The glottal stop that changes a Tahitian word's meaning, or the long vowel that separates two Samoan words, doesn't show up on any leaderboard. A listener in Apia or Suva notices it in the first sentence.

Huri Translations didn't wait for the giants to get around to the Pacific. We built our own speech AI for Pacific languages.

Our speech synthesis model speaks eight Pacific languages: Tahitian, Samoan, Tongan, Fijian, Rapa Nui, CHamoru, Carolinian and Chuukese. It was designed around these languages from the start. It isn't a global model with a Pacific setting. It handles Pacific spelling the way Pacific readers write it, glottal stops and long vowels included.

It also comes with the standard Huri has always applied to voice work. Every synthetic voice track is checked by a person listening against the script, and anything that doesn't match is re-recorded. We hold synthetic voice to the same bar as human voice-over.

Regulators are moving fast on synthetic voices. Under Article 50 of the EU AI Act, providers of AI systems that generate synthetic audio must mark it in a machine-readable, detectable way, starting 2 August 2026. In September, California enacted SB 1050, which requires ads that use AI-generated performers to say so. Disputes over voice cloning are already in court, and dubbing actors are writing AI training bans into their contracts.

Synthesis is only half of speech. Our next model is already in development, speech recognition for Pacific languages, built for the transcripts, captions and subtitles that general-purpose recognizers handle worst.

The giants will keep getting faster and broader in the languages that pay their bills. Huri is going deeper in the languages that matter to the Pacific.

If your organisation needs Pacific voices today, or wants an early look at Pacific speech recognition, contact Huri Translations. We'd like to hear what you're working on.