3,200 Hours of Indonesian and Almost Nothing Else Anywhere: A Low-Resource Speech Inventory Report

Ask any speech team what’s hard about Indonesian, Thai, or Malay ASR and you’ll get the same answer: not the modeling — the data. The major public corpora thin out sharply outside high-resource languages, and the model card that boasts 95% on English benchmarks says nothing about a Jakarta call center. This is an inventory report on where low-resource speech data actually comes from in 2026, what “in stock” means, and the questions that separate real inventory from a landing page.

The State of the Shelf

Low-resource ASR has an inventory problem disguised as a research problem. The pattern, repeated across every team we’ve watched source this material:

The gap that matters is the last row: models deployed in Southeast Asian markets fail not on architecture but on exposure. A speech recognition dataset with transcripts for Indonesian at commercial scale simply didn’t exist as a purchasable item for most of the past decade — teams recorded their own, negotiated with universities, or shipped half-trained.

That’s the context in which Thordata’s datamall inventory reads as it does: 3,200 hours of Indonesian speech, 800 hours of Thai, 500 hours of Malay, all with transcripts, sold as ready datasets with a sample-first evaluation flow. In a market where a “low resource language ASR corpus” request historically returned a consulting quote, in-stock hours are the product.

What “In Stock” Should Mean — The Buyer’s Inspection List

The phrase “Indonesian speech corpus for sale” appears on more landing pages than have hours to sell. The inspection list that separates inventory from brochure:

1. Hours are transcribed hours. Ask specifically: is the 3,200 figure raw audio or transcript-aligned audio? The difference is a labeling project you didn’t budget for.
2. Speaker diversity is stated. Hours from 40 speakers train a model that hears 40 people. Look for speaker counts and demographic spread before accent claims.
3. Domain mix is disclosed. Call-center Indonesian differs from broadcast Indonesian differs from conversational home audio. The corpus should tell you which world it recorded.
4. License scope is a field, not a paragraph. “Commercial model training permitted” in writing, per dataset — the single question procurement always asks and brochures always dodge.
5. A sample ships before the invoice. Any vendor confident in transcript alignment will let you hear and read a sample. Thordata’s sample-first flow — download sample fields, inspect, then scope — is the standard the category should hold everyone to.

How Teams Actually Combine the Sources

The sourcing pattern that repeatedly ships for low-resource ASR:

Base layer:     Licensed in-stock corpus (Indonesian 3,200h / Thai 800h / Malay 500h)
                → covers breadth, transcripts, commercial license

Domain layer:   Targeted custom collection for your channel
                (call-center audio, telephony codec, regional accents)
                → the vendor's collection infrastructure, scoped as a project

Eval layer:     Held-out recordings from your actual deployment
                → never trained on, refreshed quarterly

The base layer is what changed: it’s now rentable at dataset prices rather than buildable only at lab prices. General catalog pricing runs from roughly $0.25 per 1,000 records for standard datasets, with speech corpora quoted by hours and scope — and custom collection rides the same infrastructure (residential IPs across 190+ countries, $2.00/GB down to $0.65/GB at volume) when the domain layer needs fresh audio or the market you serve isn’t on the shelf yet.

FAQ

How much does 1,000 hours of Indonesian speech data cost?
It’s the right question and the honest answer is “quoted by scope” — speaker count, domain mix, transcript quality, and license width move the number. The structural point stands: with 3,200 hours in stock, you’re pricing a purchase, not commissioning a recording program, and the sample-first flow means the quote follows your inspection rather than preceding it.

Is 500 hours of Malay actually enough?
For fine-tuning a multilingual base model, yes — low-resource ASR in 2026 is a fine-tuning discipline, and hundreds of well-aligned hours move accuracy more than thousands of noisy ones. For from-scratch training, no; but nobody should be training from scratch in a market with strong multilingual bases.

Where do ASR corpora and LLM corpora overlap?
At the transcript: a speech dataset with transcripts is simultaneously a domain-language corpus for text models — medical Indonesian, legal Thai, whatever the domain mix holds. Teams increasingly buy one corpus and use it twice; the record-level structure (audio reference + aligned transcript + language metadata) is what makes the second use free.

What about the demand side — how do we know a market is worth entering?
Search data, as usual: the queries a market types are the cheapest leading indicator of product-market fit. Running SERP monitoring — about $0.70 per 1,000 structured responses — across your candidate markets’ voice-related queries shows you where demand is thick before you commit training budget. The same continuous SERP data crawling feed, extended to your app-store keywords, becomes a quarterly market report you didn’t have to commission.

Our language isn’t in stock anywhere. Now what?
Then you’re the domain layer: custom collection scoped against your deployment — a project the same vendor’s infrastructure runs, with the base-layer economics as your benchmark. Bring query-side intelligence to the scoping meeting; knowing how your market searches tells the collection team what “representative” should mean.

The report’s one-line conclusion: low-resource speech stopped being a research problem the day the hours became inventory — and the teams still treating it as a research problem are paying lab prices for shelf items.