Buy speech data, ready-made or built to specification

Two ways to buy: a dataset that already exists and fits your requirement, or recordings produced against your specification. Either way you get a costed quote and a consent record for every speaker.

  • A real number and a timeline, not a price range.
  • On commissioned builds, you pay for accepted hours, not hours delivered.
  • No medical or clinical data. No recorded telephone calls.

See what ships with every delivery

We reply within two business days. Your details are used only to answer this request. See our privacy policy.

  • 120 languages, low-resource ones included
  • One pilot batch before every production run
  • A signed consent record for every speaker
  • Ready-made datasets and commissioned builds, both quoted in writing
  • Audio

    Hours are not interchangeable. Twenty speakers in a treated room and two hundred in kitchens train different models, and the specification has to say which one you are buying.

  • Transcript

    Verbatim text, timestamps, speaker separation and phonetic alignment are four separate pieces of work. Agree which of them are in scope before the quote.

  • Metadata

    Speaker ID, native region, device, environment and session date. Once those columns read unknown, the dataset cannot be curated after delivery.

  • Licence

    Evaluating a model, training a production model and redistributing an output are three permissions with three different prices.

The terms that decide the price

Two quotes for the same language can differ by a factor of several. These six terms are where the difference comes from, and every one of them is settled in writing before production.

Delivered vs usable

  • You pay for hours that pass acceptance, not hours handed over.
  • Agree the rejection rule before production starts.

Speaker balance

  • Many hours from a few voices cost less per hour.
  • Diversity is a line item, not a rounding error.

Acoustic condition

  • Treated room, office, street, vehicle, call centre.
  • Same language, four different products.

Transcript depth

  • Verbatim, normalised, timestamped, diarised, aligned.
  • Each layer is priced separately.

Rights

  • Evaluation, production training, redistribution.
  • No single rate covers all three.

Exclusivity and term

  • Exclusive rights cost more than non-exclusive ones.
  • A fixed term changes the number again.

Two ways to buy speech data

Dimension Ready-made dataset Commissioned to your order
What you choose From what already exists and passes your test Every parameter in the specification
Fit to your model Close by design; you adapt around the gaps Built for the use case you stated
Languages Whatever is available now 120 languages, including low-resource ones
Consent Checked against the licence you need Collected for your stated use
Exclusivity Usually non-exclusive; it depends on the source licence Not resold, unless you take a non-exclusive licence
Delivery time After licensing; no production run After production and a pilot batch

Neither route is better in the abstract. A ready-made dataset is cheaper and faster when something existing genuinely fits. Commissioning is the only route when the rights or the recording conditions you need do not exist yet. If a ready-made set does not fit, we say so rather than sell it anyway.

Ready-made means a dataset that already exists and can be licensed as it stands. You give us the language, the recording condition and the rights you need, and we come back with what is available: a specification sheet, a sample and the licence terms. Nothing is rebuilt, so it is the faster route whenever something genuinely fits.

Commissioned means recordings produced against your written specification. The language, the speaker profile, the acoustic condition, the transcript depth and the rights are each fixed in writing and priced separately. It starts with a pilot batch, and you pay for accepted hours.

Both routes carry the same paperwork: a consent record for every speaker, the collection methodology, and a licence that names what you may train and whether you may redistribute outputs.

How the per-hour figure is quoted

Speech data is quoted per delivered audio hour, and that number only means something once the acceptance rule is fixed. An hour that fails your test is not an hour you paid for.

Two vendors can quote the same language at very different rates because they are counting different things: whether transcription sits inside the figure, whether failed recordings are replaced, and whether studio time and speaker fees are part of the line item or billed beside it.

Ask for the rate and the acceptance rule together. A rate with no rejection rule cannot be compared with anything.

The floor is zero. Free read-speech corpora such as Mozilla Common Voice exist and are in use. What buyers pay for is conversation, consent records and transcripts, which is the part the free corpora do not cover.

A ready-made licence is priced differently: usually a one-time fee for the dataset as it stands, sometimes with a term or a user limit, not a per-hour rate. When both routes are open to you, ask for the two figures side by side.

How to check a sample before you commit

  1. Same speaker, two sessions

    Ask for two sessions from one speaker and confirm both carry the same identifier. If identity changes, diarisation downstream is already broken.

  2. Re-transcribe twenty clips

    Compare your own transcript against the delivered one. Convention defects surface immediately; acoustic ones do not.

  3. Open the metadata file

    Count populated fields across the whole sample, not the first page. Empty columns are the usual defect.

  4. Read one consent record

    It has to name machine-learning training, not only publication or research.

  5. Load it in your pipeline

    Use the format that was specified in writing, not a copy someone converted for you.

A sample is not a preview of the audio quality. It is a test of the three things that go wrong after delivery: identity, metadata and convention.

What ships with the delivery

  • A consent record for every speaker, with the permitted use spelled out in explicit terms.

  • The collection methodology: device, environment and session protocol.

  • The licence, naming whether the permitted use is evaluation, production training or redistribution of outputs.

  • A delivery manifest: usable hours, speaker count, and what quality control rejected.

Every delivery, ready-made or commissioned, ships with the same four documents. Missing any one of them turns the dataset into a legal question at review, whatever its audio quality.

What we do not sell

Four kinds of audio are refused regardless of how good the recording is. The list is short, and it is not negotiable.

  • Call recordings of private individuals.
  • Medical and clinical recordings.
  • Children's voice data without documented guardian consent.
  • Audio whose consent does not name the use it is being sold for.

Every category below is recorded against a written brief, not pulled from a shared pool. Open the one that matches your project and the page states what gets recorded, how it is annotated, and what a sample looks like.

Speech categories you can order

Each card is a full specification page. If your project matches two of them, that usually means the brief needs both, e.g. accented English recorded under office noise.

  • Call Center Speech

    Customer-service calls, simulated rather than intercepted, capturing both the agent and the caller side, used for call center QC, intent recognition, and dialogue systems.

    • Category
  • Conversational Speech

    Speech from natural conversation between two or more people, on open or semi-structured topics, used for conversational AI, voice assistants, and small talk models.

    • Category
  • Singing Voice

    Vocal recordings with melody, in both a cappella and accompanied form, used for singing voice synthesis, music information retrieval, and lyric alignment.

    • Category
  • Speech Commands

    Targeted recordings of short command words or phrases, usually with many speakers reading each entry several times over, used for wake words and on-device command recognition.

    • Category
  • Read Speech

    Recordings of speakers reading specified text, with clear pronunciation and known text, the base material for TTS and ASR.

    • Category
  • Podcast Speech

    Long-form podcast and interview audio, either solo monologue or two-person conversation, used for long-form speech recognition and speaker modeling.

    • Category
  • Multilingual Speech

    Speech data covering multiple languages within one project, used for multilingual ASR, cross-lingual transfer, and language identification.

    • Category
  • Noisy Speech

    Speech collected under background noise (street, in-car, restaurant, office and other real environments), used for noise-robust models and speech enhancement.

    • Category
  • Code-Switching Speech

    Speech that mixes two or more languages inside a single utterance (Hinglish, Spanglish, Taglish), used for recognition in real spoken settings.

    • Category
  • Children Speech

    Speech data from child speakers, grouped by age band, used for children's speech recognition and children's education products.

    • Category
  • Audiobook

    Continuous long-form narration in audiobook style, steady in intonation and generous in duration, the main source material for high-quality TTS.

    • Category
  • Short Video Speech

    Speech in the talking-head style of short video (fast, emotionally strong, colloquial), used for short video subtitles and content understanding.

    • Category
  • Live Stream Speech

    Long-session spoken content from live streams (host monologue and responses to audience interaction), unscripted and spoken quickly.

    • Category
  • Voice Assistant

    Command and dialogue data for voice assistants, usually with intent annotation, covering the full chain from wake word to understanding to response.

    • Category
  • Elderly Voice

    Speech data from elderly speakers, usually stratified by age band and health status, used for voice products built for older users.

    • Category
  • Accented English

    English speech from non-native speakers or specific regional accents, grouped by accent origin, used to improve accent robustness in ASR.

    • Category
  • Speech Translation

    Parallel data pairing source-language audio with target-language translation, used for speech-to-text and speech-to-speech translation models.

    • Category
  • In-the-Wild Speech

    Speech collected under fully natural conditions (no studio, no topic constraints), as close to real usage as collection gets.

    • Category

Scenario tags

  • Call centre
  • IVR prompts
  • Live agent dialogue
  • Two-person conversation
  • Group discussion
  • Read speech
  • Scripted prompts
  • Podcast monologue
  • Podcast interview
  • Audiobook narration
  • Singing voice
  • A cappella
  • Voice commands
  • Wake words
  • Voice assistant turns
  • Intent annotation
  • Noisy environments
  • In-car audio
  • Restaurant noise
  • Code-switching
  • Hinglish
  • Spanglish
  • Multilingual projects
  • Accented English
  • Regional accents
  • Non-native speakers
  • Child speakers
  • Age-banded children
  • Elderly speakers
  • Age and health bands
  • Short video speech
  • Live stream speech
  • In-the-wild speech
  • Speech translation pairs
  • Parallel audio and text

Questions we get asked

Do you sell ready-made speech datasets?

Yes, alongside commissioned builds. A ready-made dataset is licensed as it stands: you see the specification sheet, a sample and the licence terms before you commit. Availability changes with what is in the network, so we confirm it against your requirement rather than publish a stock list. If nothing ready-made fits, we quote a commissioned build instead.

What does a speech data quote depend on?

Six things: how many delivered hours survive acceptance, how many distinct speakers they come from, the acoustic conditions, the transcript depth, the rights you need, and whether the licence is exclusive and for how long.

Which languages can you deliver?

120 languages, described in 468 language-by-category dataset briefs, with low-resource languages included. The language pages state what is hard about collecting each one.

What licence do I get, and can I train a commercial model?

The licence names the permitted use, and production training for a commercial model is available. Redistributing the raw audio is a separate permission. We do not retain the right to resell a dataset produced for you.

Can I check quality before committing?

Yes. A pilot batch is produced before full production. Speaker identity, noise floor and transcript convention are settled on that batch, where problems are cheaper to fix than after delivery.

Do you handle call recordings or medical audio?

No. Calls between private individuals, and medical or clinical recordings, are outside what we handle whatever the consent position. The compliance entries explain the reasoning behind each exclusion.

Keep reading

  • AI Data Brokerage

    A data broker sits between the people who need training data and the people who can produce it. Datatoorder does that as a platform: we collect and process data in-house, licence ready-made datasets, and commission what does not exist yet: no recycled data, no self-serve stock catalogue.

    • Buying
  • AI Training Data Providers

    What to check before you sign with a training data provider, and how we compare on each point.

    • Buying
  • AI Training Data Marketplace

    Marketplaces sell what already exists. We collect, process and trade what you actually need. Here is when each model makes sense.

    • Buying
  • Full data catalog

    39 data categories across 120 languages, and how to specify each one.

    • Buying

Submit a sourcing request

Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.

  • Pilot batch before the full run, so problems surface early.
  • Consent documentation delivered with the data.
  • No medical or clinical data. No recorded telephone calls.

We reply within two business days. Your details are used only to answer this request. See our privacy policy.

  1. You send the spec Language, hours, recording conditions, and the deadline.
  2. We reply within two business days A real number and a real timeline, not a range.
  3. You approve a pilot batch A small run first, so problems surface before the full order.
  4. Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.

Prefer email? [email protected]

Contact

Talk to a human

Send a specification and we will come back with a real number and timeline.

Submit a sourcing request

Or use the contact form