# Off-the-shelf or custom speech data: how to choose

Source: https://indecalai.com/blog/ots-vs-custom-speech-data/
Published: 2026-10-27 · By the IndecalAI team · 6 minute read · Topic: Datasets

You can license speech data that's already recorded, or pay to have it recorded for you. Both are reasonable, and the wrong pick costs months or money. Here's how to tell which one your project needs, and why many teams end up using both.

## Key takeaways

- Off-the-shelf (OTS) speech data is already recorded, so you can test a sample this week. Custom data is recorded to your brief, so it fits better but takes longer.
- Choose OTS when your users sound like the general population of a language. Choose custom when they don't: a specific accent, a noisy channel, children, or your own product's vocabulary.
- You can do both: an OTS base for breadth, then a smaller custom set aimed at the gaps the test set shows.
- Whatever you buy, read the licence: what it lets you train and ship, whether speakers consented to commercial AI use, and who else can license the same audio.

## What each one actually is

**Off-the-shelf speech data** has already been recorded, transcribed and checked. A vendor collected it once and licenses it to more than one customer, so you're buying access to something that exists today. You can usually hear sample clips and read a spec sheet before you sign anything.

**Custom speech data** is recorded for you. You decide who speaks, what they say, where and on what device, and the vendor recruits speakers and records to that brief. Nothing exists until the project starts, so it takes longer, but every hour is aimed at your model.

Neither is better in general. They answer different questions: OTS answers "can we get good data fast?", custom answers "can we get exactly the data our users need?"

## Side by side

The trade-offs, line by line. The details vary by vendor, so treat the right-hand columns as what's typical and ask about the specifics.

|  | Off-the-shelf | Custom collection |
|---|---|---|
| Time to first data | Fast: sample clips on request, the full set once the licence is agreed | Longer: speakers are recruited, recorded and checked first |
| Fit to your users | General population of the language, in the conditions it was recorded in | Exactly the speakers, accents and conditions you specify |
| Your own vocabulary | Only if it happens to be in the audio | Covered by scripted prompts you approve |
| Metadata | Whatever was captured at recording time | Agreed in the brief before anyone records |
| Exclusivity | Usually non-exclusive: check who else can license it | Scoped to your engagement |
| Cost | Usually lower per hour, since the recording cost is shared | Higher per hour, but no hours spent on audio you don't need |
| Main risk | Close but not quite your users | Takes longer, and the brief has to be right |

## Five questions that decide it

Work through these in order. If you answer "no" early, you can usually stop there.

1. **Is your language already in stock?** **Yes:** Keep going. **No:** Custom collection is the only route, or a public corpus if the licence allows commercial use.
2. **Do your users sound like the general population?** **Yes:** OTS covers you. **No:** If they're mostly one accent, region, age group or children, plan a custom set for that group, or a custom top-up on an OTS base.
3. **Do your recording conditions match?** **Yes:** Good. **No:** Narrowband phone lines, cars, far-field microphones and noisy shop floors are hard to fake. Record in them.
4. **Does the model need your own words?** **Yes:** Product names, commands and domain terms need scripted prompts, which means custom. **No:** General speech is fine, and OTS stays in the running.
5. **How soon do you need to train?** **This month:** Start with OTS and measure. **No rush:** Custom fits if the earlier answers point to it. Either way, test on your own held-out audio before you scale up.

## You can do both

Sometimes the five questions don't land cleanly on one side. The language is in stock, but the product runs on phone calls. The general population fits, but there are 300 product names to learn. That's where a mix works well.

> Buy breadth off the shelf. Pay for precision where your test set says you need it.
>
> (Our rule of thumb)

License an off-the-shelf base so the model hears a wide range of voices early. Fine-tune, score it on audio from your real users, and read the errors. Then commission a smaller custom set aimed at exactly what's still failing: the accent it keeps missing, the channel it can't handle, the words it gets wrong. You spend custom money only where it's needed, and you start training weeks sooner. If you're still working out the volume, our post on [how many hours of speech data you need](https://indecalai.com/blog/how-many-hours-of-speech-data-to-fine-tune-whisper/) covers how to size it.

**See what's already in stock** 71 languages, about 500 hours each and roughly 1,000 for Hindi, with sample clips and a spec sheet on request. [Browse OTS datasets](https://indecalai.com/datasets/ots/)

## What to check before you license any speech dataset

A dataset is only as useful as the licence that comes with it. Whether it's ours or anyone else's, get clear answers to these before you train:

- **What you may do with it.** Training is the obvious use, but check that commercial products and models trained on the data are allowed, and for how long.
- **Speaker consent.** Did every speaker agree to their voice being used to train commercial AI? Ask how consent was collected, not just whether it was.
- **Who else can license it.** Off-the-shelf data is usually non-exclusive. That's often fine, but know it going in.
- **Metadata.** Speaker ID at minimum, plus gender, age band, region or accent, and recording device if you need results by group.
- **Transcription style.** Verbatim or clean, how numbers and filler words are written, and how mixed-language speech is handled.
- **A sample first.** Listen to clips and run them through your pipeline before you sign.
- **Handling and deletion.** Where the data is stored, who can access it, and what happens to it when the licence ends.

> **About public datasets:** Free corpora vary a lot. Mozilla Common Voice is released under CC0 (newer downloads add terms such as not re-sharing the data) and LibriSpeech is CC BY 4.0, but some research corpora are licensed for non-commercial use only. Check before you train a product model on them.

## How this works with us

We do both, so we don't need to push you toward either one.

- **Off-the-shelf:** about 35,000 hours of speech, prompts and conversations across [71 languages](https://indecalai.com/datasets/ots/), already recorded. Every speaker consented to commercial AI training, and we send sample clips and a spec sheet on request.
- **Custom:** scripted prompts and natural conversations recorded to your brief from a network of 30,000+ people worldwide, using our own recording tools. We sign the NDA before any data moves, and record a small sample free before you commit. Start with the [custom speech data](https://indecalai.com/speech-data.html) page.

Not sure which you need? Send the language, who your users are and how they'll be recorded, and we'll tell you honestly which route fits, or whether a mix of the two does.

## Common questions

### What is an off-the-shelf speech dataset?

Speech audio that has already been recorded and transcribed, and is licensed to more than one customer. You can usually hear sample clips and read a spec sheet before you buy.

### Is off-the-shelf speech data cheaper than custom collection?

Usually, per hour, because the recording cost is shared across customers. Custom data costs more per hour because speakers are recruited and recorded for you, but none of it is spent on audio you don't need.

### Can I get exclusive rights to an off-the-shelf dataset?

Sometimes. It depends on the vendor and the licence. If exclusivity matters, ask before you test samples, and compare the cost with a custom collection scoped to you.

### Can I combine off-the-shelf and custom speech data?

Yes: an off-the-shelf base for breadth, then a smaller custom set aimed at the accents, conditions or vocabulary your test set shows the model still gets wrong.

### What should a speech dataset licence cover?

What you may train and ship with the data, whether every speaker consented to commercial AI training, who else can license it, how long the licence runs, and what happens to the data when it ends.

## Sources

- https://commonvoice.mozilla.org/en/datasets
- https://www.openslr.org/12

## Related

- [How many hours of speech data do you need to fine-tune Whisper?](https://indecalai.com/blog/how-many-hours-of-speech-data-to-fine-tune-whisper/)
- [How to choose a data annotation vendor: the questions worth asking](https://indecalai.com/guides/choosing-a-data-annotation-vendor.html)
- [Speech dataset specification: what to put in your collection brief](https://indecalai.com/guides/speech-data-collection-specs.html)

Talk to our team: https://indecalai.com/contact.html