Paxa Labs launches Bangkok AI lab serving Thai speech, translation and OCR through one API

What launched

Paxa Labs, a new AI research lab based in Bangkok, has announced its first products: a set of proprietary models covering Thai and English text-to-speech, document OCR and translation, all delivered through a single API. The announcement was made in a post on dev.to by the company itself, which frames the lab as purpose-built for a language that general-purpose models tend to mishandle.

Three services are live today. Text-to-speech is handled by paxa-tts-flash-v1, which offers 26 voices, including regional Thai accents from Isan, Northern and Southern Thailand alongside English-first voices. Output can be MP3, Opus or WAV, buffered or streamed, with optional timestamps at the word, sentence or utterance level. A WebSocket endpoint synthesises audio as text arrives, aimed at agent applications and live captioning, and an OpenAI-compatible alias at /v1/audio/speech means existing clients can switch over by changing the base URL.

Document OCR comes from paxa-ocr-lite-v1, which accepts PDF, PNG, JPEG and WebP files and returns per-page Markdown or typed blocks such as headings, paragraphs, lists, tables and figures. Translation is handled by paxa-translation-lite-v1, which translates fourteen source languages into Thai with controls for formality, borrowed-word handling, glossaries, do-not-translate terms and reference context. It also offers Markdown and HTML modes that translate prose while leaving markup untouched, which the company notes matters because a tag shifted by a single character can move a Thai word boundary.

Speech recognition is not among the launches. According to the post, it remains a research project and will ship only once it holds up against real Thai audio.

Why Thai needs its own models

The dev.to post lays out the lab's reasoning: Thai text is written without spaces between words, so some component has to decide where one word ends and the next begins. Tone changes word meaning, speakers routinely mix Thai and English within a single sentence, and real business documents stack both scripts together with tables, stamps and handwriting on one page.

The company's position is that general-purpose models treat these conditions as edge cases, while anyone building for Thai users faces them as ordinary input. Its development loop reflects that: the lab starts from failures observed in real Thai speech, text and documents, turns each failure into a data example and an evaluation, and lets repeated patterns shape the model, the training method and the serving stack. The improved system goes back into the product, and the next difficult input sets the research agenda.

Pricing and API design

All three products are billed in credits, at 1,000 credits to one US dollar, available pay-as-you-go or on monthly plans. According to the announcement, text-to-speech costs $15 per million characters, translation $25 per million characters of text, and OCR $0.0065 per page. New accounts start with free credits.

One unusual design choice is the billing order: requests are charged before inference runs and refunded if inference fails, so a failed call never costs anything. The company has published a separate write-up defending that sequence.

Onboarding is lightweight. Users can sign in with Google, GitHub or Hugging Face, generate a key, and try all three products in a browser playground before writing code. The company has also made its documentation machine-friendly: every docs page is served as Markdown by appending .md to its path, /llms.txt acts as a curated index, /llms-full.txt exposes the entire corpus, and an installable agent skill is available via npx.

Why it matters

Most frontier AI effort concentrates on high-resource languages, leaving languages like Thai served by models that were never optimised for its structure. Paxa Labs is betting that a lab focused on one language's specific failure modes, no word spacing, tonal meaning, dense code-switching and messy real-world documents, can outperform general models where it counts for local developers.

The OpenAI-compatible endpoint lowers switching costs, metered per-character and per-page pricing is easy to reason about, and the refund-on-failure policy removes a common billing irritation. If the quality holds up, the launch signals a broader pattern: regional labs packaging deep, language-specific expertise as ordinary APIs rather than chasing frontier-scale models.