Most speech data is sold in corpora priced for teams training from scratch. Most teams are not doing that. If you are adapting a model that already knows the language, the useful order is tens of hours, and the thing standing in the way is usually the vendor's minimum rather than the price per hour. This guide covers what a small order can actually accomplish, what 10, 25, and 100 hours cost, and when buying small is the wrong call.

What counts as a small speech dataset

In speech, "small" is anything from roughly ten to a couple of hundred hours. That sounds tiny next to the corpora vendors advertise, and for training a model from scratch it would be. But almost nobody trains from scratch any more. If you are adapting Whisper, a wav2vec variant, or a commercial ASR API to your own conditions, the base model already knows the language. What it does not know is your microphones, your vocabulary, your speakers, and your background noise, and that gap closes with tens of hours of the right audio rather than thousands of hours of the wrong audio.

The obstacle has rarely been the per-hour price. It is the minimum order. A vendor that only licenses whole corpora puts the entry point somewhere in five or six figures, which quietly rules out every evaluation, benchmark, and first fine-tune a small team wants to run. That is a packaging decision, not a cost of production.

What 10, 25, and 100 hours each buy you

Spirelight sells hours out of the same datasets rather than requiring the whole set, from a 10 hour minimum. Prices below span the catalogue's cheapest to dearest languages, and the per-hour rate falls as the order grows.

  • 10 hours, $1,000 to $1,550. Enough for a held-out evaluation set or a dialect benchmark. This is the size that answers "how badly does our current model actually do on Egyptian Arabic", which is usually the question worth answering before spending anything larger.
  • 25 hours, $2,000 to $3,250. A first fine-tuning pass in a new language or accent, with audio left over for a test split. Expect it to move word error meaningfully where the base model was weak, and barely at all where it was already strong.
  • 50 hours, $4,000 to $6,500. Domain adaptation with room for speaker diversity: call centre audio, in-car speech, or a specific regional accent, with enough speakers that the model does not overfit to a handful of voices.
  • 100 hours, $6,000 to $9,500. A production-grade adaptation for one language and domain, and the point where the full-scale per-hour rate applies.

Worked example: benchmarking before you buy

A team building a voice agent for the Egyptian market wants to know whether their ASR is good enough to ship. They buy 10 hours of Egyptian Arabic conversational audio with verified transcripts, hold all of it out as a test set, and measure word error on their current stack. Two outcomes, both useful. If word error is acceptable, they have spent four figures to avoid a six-figure collection they did not need. If it is not, they now know exactly which conditions fail, and they can commission the next order against evidence rather than a guess.

This is the sequence we would recommend to anyone unsure of scope: buy the evaluation set first, measure, then size the training order from what the measurement shows. It is cheaper than the alternative and it produces a number you can put in front of whoever approves the budget.

Worked example: a first fine-tune in a new language

A speech analytics company adding Marathi has no in-house speakers and no way to judge output quality. They order 25 hours of spontaneous conversational Marathi, split it 20 hours for fine-tuning and 5 for evaluation, and adapt an existing multilingual model. The point of the 5 held-out hours is that they come from the same recording conditions as the training audio, which is what makes the before-and-after comparison honest. A test set scraped from elsewhere measures a different problem.

What to check before buying small

Small orders are where sloppy sourcing hurts most, because there is no volume to average out a bad batch. Four things are worth confirming in writing:

  1. Domain match over raw hours. Twenty hours recorded in conditions like yours will beat two hundred hours of clean read speech for almost any production use. Ask what the recording setup actually was.
  2. A test split from the same conditions. Ask whether you can take your evaluation hours from the same collection as your training hours. If they come from different sources, your before-and-after numbers measure the source difference as much as your fine-tune.
  3. License scope at small volume. Check that a small order carries the same commercial rights as a large one, including the right to keep and deploy the model you train. Some vendors restrict this below a volume threshold.
  4. Consent that survives scrutiny. The provenance question does not get smaller with the order. Speakers should have consented to AI training use specifically, and the vendor should be able to show that per speaker.

When a small order is the wrong tool

Being straight about the limits: if you are training from scratch, building a foundation model, or working in a language with no usable base model, tens of hours will not get you there and no amount of careful curation changes that. Small orders also make poor sense when your requirement is genuinely unusual, a rare dialect in a specific acoustic environment, say, because the recruiting and setup cost dominates and a commissioned collection is the honest route. In both cases the sizing question is worth working through properly before spending anything.