Buy off-the-shelf speech data when an existing dataset already matches your language or dialect, your speech style and your acoustic conditions, and you need data now. Commission custom collection when the gap is domain vocabulary, a specific channel, device or environment, a speaker group the dataset doesn't cover, or a language you can't find in stock.
Many teams do both: catalog data for breadth, and a small custom set for the gap and for the test set. This guide gives you a decision table, a worked cost example, what to check before you pay for ready-made data, and three hybrid plans. A note on bias: SpeechData.ai sells both ready-made datasets and collection to spec, so the question here is which one fits.
What off-the-shelf and custom mean
Off-the-shelf (or ready-made) speech data was recorded, transcribed and priced before you arrived. You license it as it is. The languages, speakers, topics, devices and transcript conventions are fixed. Free open corpora are off the shelf too; check whether each license allows commercial use.
Custom speech data is recorded after you sign, to a spec you write: who speaks, what they say, on which devices, in which rooms, and how it is transcribed. You get fit and control, and you pay for recruitment, recording, QA and transcription done for your project alone.
Quality can be checked in a sample either way. The deciding difference is fit: how closely the recordings match the audio your model will hear in production.
Off-the-shelf vs custom: the decision table
| Criterion | Off-the-shelf fits when | Custom fits when |
|---|---|---|
| Language and dialect | A dataset exists for your exact variety, such as Latin American rather than European Spanish | Your language or dialect isn't in stock, or only a neighboring variety is |
| Speech style | The corpus style matches production, such as spontaneous conversation for a voice agent | You need scripted prompts, commands or dialogues that follow your product's flow |
| Domain vocabulary | Everyday conversation covers what your users say | Users say product names, drug names, account terms or codes that everyday talk rarely contains |
| Acoustic conditions, device, channel | The devices and channels in the metadata resemble yours | You need narrowband calls, a headset, far-field, in-car audio or a specific noise |
| Speaker demographics | The dataset's mix of age, gender and region resembles your users | You need set quotas, such as older speakers or one regional accent |
| Timeline | You need data this quarter, and it is already recorded and transcribed | You can wait for recruitment, recording, QA and transcription |
| Budget predictability | You want the price before the first call: hours times a per-hour price | You can work from a per-project quote that follows the spec |
| Exclusivity | Competitors licensing the same recordings is acceptable | Nobody else may license your recordings |
| Evaluation data | You need a general benchmark for a language, under a license that covers evaluation | You need a held-out test set that sounds like production |
If every row lands in the left column, license. If only one or two land on the right, consider a small custom set on top of catalog data before you scope a full custom corpus.
Where ready-made data runs out
Speech style and recording channel are easy to underestimate. In OpenAI's Whisper paper, two models that both scored 2.7% word error rate on LibriSpeech's clean test set, which is read English audiobook speech, scored 13.8% and 28.3% on Switchboard, a corpus of conversational telephone speech. The 28.3% belonged to a wav2vec 2.0 model trained with supervision on LibriSpeech, the 13.8% to Whisper. A score on one kind of speech says little about accuracy on another, so check the style and channel of any dataset, ready-made or custom, before anything else.
Past style and channel, four gaps send teams to custom collection:
- Domain vocabulary. Everyday conversation has common words in quantity, but your product names, medication names or account jargon may never come up. A custom scenario puts them in speakers' mouths, in context.
- Channel, device or environment. If production audio is narrowband phone calls, a drive-through speaker post or a car cabin, recordings from quiet rooms on good microphones don't show the model those conditions.
- A speaker group. A catalog dataset has the speaker mix it has. If your users skew older or come from one region, you need recruitment against quotas.
- A language that isn't in stock. Our catalog covers 60 languages and dialects. South African English is one of them, while isiZulu, isiXhosa and Sesotho are recorded to spec by native speakers.
What it costs: a worked example
Off-the-shelf costs are simple arithmetic when the price per hour is published. Ours run from $60 to $95 per hour of audio, depending on the language, and our speech data pricing page breaks them down by region. Volume and multi-language discounts are on request, so the figures below are list prices.
Say you are building a voice agent for a bank in Mexico, with Brazil and the US next:
| Dataset | Hours | Price per hour | Cost |
|---|---|---|---|
| Latin American Spanish, recorded in Mexico | 500 | $65 | $32,500 |
| Brazilian Portuguese | 500 | $65 | $32,500 |
| US English | 500 | $90 | $45,000 |
| Total | 1,500 | $110,000 |
The Latin American Spanish dataset holds 2,000 hours from 200 speakers, so 500 hours is a quarter of it; all of it would be 2,000 × $65 = $130,000. When you buy part of a dataset, ask how many speakers the hours come from. 500 hours from many speakers covers more voices than 500 hours from a few.
The custom line can't be priced from a list. Collection is quoted per project because the cost follows the spec: how hard the speakers are to recruit, how many you need and how tight the quotas are, scripted or spontaneous speech, devices and environments, transcription depth and labels, and the deadline. For our custom collection, a feasibility assessment and a quote come back within one business day of receiving a brief.
The practical point: in a hybrid plan, the catalog part of the budget is fixed the day you choose the datasets, and only the custom part waits on a quote. A small, targeted custom set keeps most of the budget predictable.
What to check before you pay for ready-made data
Ready-made doesn't mean unseen. Ask for a sample and check four things: listen to a few files and run them through your current model; confirm every speaker has the metadata you need (ours lists age, gender, region and recording device for every speaker); read the transcript conventions, from fillers and overlaps to numbers and timestamps, so the new data lines up with what you have; and read the consent wording, which should cover commercial AI training. We send free sample audio, matching transcripts and the metadata schema to a company email.
Three hybrid plans
"Use both" is not a plan. These are.
1. Catalog conversation first, in-domain calls second
The team building the Mexican banking agent licenses 500 hours of Latin American Spanish conversation and fine-tunes a pretrained model on it, for the variety and for unscripted dialogue. The catalog doesn't contain the bank's call flows, its product names or its phone channel.
So the team commissions 20 to 50 hours of role-played banking calls as call center speech data, with agent and customer on separate channels, captured as narrowband telephony if that is what the phone line carries. The names and account details in each scenario are fictional, so no real customer data is involved.
Treat 20 to 50 hours as a planning starting point, not a measured threshold: big enough to split, small enough to learn from before you buy more. Split it by speaker. One part goes to adaptation, and the other becomes a held-out test set that no training run ever sees, as our ASR test set guide explains.
How much the in-domain hours will help is something to measure, not assume. The Whisper authors measured only zero-shot accuracy, and wrote that for many domains where high-quality supervised speech data exists, results can likely be improved further by fine-tuning.
2. Test set first, then decide
A more cautious variant reverses the order. Get a small in-domain test set first, then score your current model, and the same model after fine-tuning on catalog data. If the catalog fine-tune reaches your target, you may not need custom training data at all. If it doesn't, the errors show what the custom set must contain: which words, which speakers, which conditions.
If you already have production recordings you have the right to use, the test set can come from them instead: our audio annotation service transcribes your own audio under a written guideline.
3. Catalog for the languages it has, custom for the rest
A South African contact-center product shows the pattern. South African English is licensed off the shelf, 2,000 hours from 200 speakers, and isiZulu, isiXhosa and Sesotho are recorded to spec. Speakers also switch into other South African languages mid-conversation, so the custom brief should say whether code-switched speech is wanted and how the transcripts should mark it.
Exclusivity
Off-the-shelf data is normally non-exclusive: a catalog exists to license the same recordings to many buyers. Assume a competitor can license what you license. For general language coverage that may be fine, since what sets your product apart is more likely your in-domain data and your model than everyday conversation anyone can buy.
Custom collection can be exclusive, but only if the contract says so. If it matters, ask in writing: may the vendor reuse or resell the recordings or transcripts, does exclusivity end after a set period, does it cover the scenarios and scripts you supplied, and what does it add to the price.
Where to start
If a ready-made dataset may fit, compare its sample with your production audio: the catalog lists all 60 datasets with hours, speakers and price per hour. If the gap is domain, channel, speakers or language, send a brief for custom speech data collection and get feasibility and a quote within one business day.
