This website uses cookies to ensure you get the best experience.

VALSEA and our selected partners use cookies and similar technologies (together “cookies”) that are necessary to present this website, and to ensure you get the best experience of it. If you consent to it, we will also use cookies for analytics purposes.

See our Cookie Policy to read more about the cookies we set.

You can withdraw and manage your consent at any time, by clicking “Manage cookies” at the bottom of each website page.

Select which cookies you accept

On this site, we always set cookies that are strictly necessary, meaning they are necessary for the site to function properly.

If you consent to it, we will also set other types of cookies. You can provide or withdraw your consent to the different types of cookies using the toggles below. You can change or withdraw your consent at any time, by clicking the link “Manage Cookies”, that is always available at the bottom of the site.

To learn more about what the different types of cookies do, how your data is used when they are set etc, see our Cookie Policy.

These cookies are necessary to make the site work properly, and are always set when you visit the site.

Vendors Teamtailor

These cookies collect information to help us understand how the site is being used.

Vendors Teamtailor
Skip to main content

We're Building the Voice Layer for Southeast Asia!

Connect Job openings
valsea-demo

The Problem Is Bigger Than Transcription

600 million people in Southeast Asia speak in ways that existing voice AI systems fail to understand. Not because the technology is broken. Because it was never built for them. Singlish. Manglish. Bahasa campur. Code-switched Vietnamese. Thai with English inserts. These are not edge cases. They are the default. Every enterprise, contact centre, clinic, and logistics operator in the region runs on speech that today's models treat as noise. We're fixing that — at the infrastructure level.

The Scale of the Problem

Million people speaking in ways no current AI understands
Languages and dialects in our training and evaluation set
Core markets where we're deployed and expanding

Open Roles — If You're the Right Fit, You'll Know

All jobs

How We Work

Ownership is assumed, not assigned. When something needs to exist, someone builds it. You don't wait for a spec. You don't wait to be asked. Speed is discipline — we ship to learn, iterate fast, and don't confuse motion with progress. We operate in ambiguity. Most of what we're doing hasn't been done before, especially at the intersection of low-resource languages, speech semantics, and real-time systems. Trust is default. We hire for character and give people room to work. If you need to be managed, this isn't the right place. We care about depth. We'd rather go deep on one hard problem than ship five shallow features. We build what should already exist — not because it's impressive, but because it matters.

Before VALSEA vs. After VALSEA

  • val

    BEFORE → Generic ASR output: "Can you lah pass me the file or not ah I send you tru whatsapp" — raw text, no structure, no action. AFTER → VALSEA output: Intent: File transfer request | Action: Send via WhatsApp | Assignee: [Recipient] | Status: Pending | Priority: Normal. This is what speech-to-meaning looks like.

    Real transcript sample

    Singlish contact centre audio, Singapore

What You'll Actually Build

  • Code-switching ASR: Build models that handle mid-sentence language shifts (e.g. English → Bahasa → back to English) without losing semantic continuity

  • Semantic correction layer: Design a post-ASR correction pipeline that fixes dialect-specific mishearing, fills in dropped words, and resolves contextual ambiguity

  • Evaluation pipeline: Build automated evals for noisy, real-world speech — including WER on code-switched audio and intent-match scoring on structured outputs

  • Transcript → workflow engine: Design the structured output schema and rule-based + ML-based extraction system that converts conversational speech into CRM entries, task tickets, and meeting summaries

  • Low-latency streaming ASR: Optimize the real-time inference pipeline to achieve sub-300ms word-level outputs on consumer hardware across 5+ SEA language variants

  • Slang normalisation: Map regional colloquialisms, contractions, and informal expressions to their formal equivalents without destroying conversational tone or speaker intent

  • Data flywheel architecture: Design the feedback loop between production outputs, human correction, and model retraining — so the system gets better the more it's used

This Is NOT for Everyone

  • If you need a detailed brief before starting — you won't thrive here. We write specs to document decisions, not to initiate them.

    We write specs to document decisions, not to initiate them.

  • If you optimise for optics over outcomes — wrong place.

    No one here cares what your role title is. Only what you shipped.

  • If ambiguity makes you anxious — this isn't the role.

    We're in early, hard territory. The map doesn't exist yet. You'd be drawing it.

  • If you're here to pad a resume — we'll both waste time. The problems we're solving are hard and consequential. They require people who actually care.

  • If you need hand-holding on technical decisions — this isn't for you. We expect you to go deeper on your domain than anyone else in the room.

Career site by Teamtailor