Speech data collection companies recruit speakers and record new audio to your specification. Companies that do this include Appen, DataForce (part of TransPerfect), LXT, TELUS Digital, Defined.ai, Shaip, FutureBeeAI, Alconost and David AI, plus SpeechData.ai, which we run. They differ less in what they promise than in how they record: managed crowds who record remotely on their own phones and computers, supervised sessions in studios, cars or a client's site, and catalog-first vendors that collect new data to extend datasets they already sell.
The one question that sorts them is where your audio has to be recorded. If speakers can record where they are, on ordinary phones, headsets or laptops, nearly every company here can do it, and the choice comes down to languages, quotas and QA. If the recording has to happen in a studio, a car or on your premises under supervision, the list gets much shorter.
A note on bias: SpeechData.ai is one of the companies below. Everything about the others comes from their own service pages, case studies and homepages, checked in October 2026. Every contributor count and language count is the vendor's own claim, and we left out prices. If you want speech that already exists rather than new recordings, our comparison of speech data providers covers catalogs and open datasets.
Three ways collection companies record speech
Vendors sell this as speech, voice or audio data collection services, but the work is the same: find the right speakers, record them, check the audio, transcribe it and deliver it with consent records. How they record is where they part ways.
Managed crowd, recorded remotely. Contributors sign up through the vendor's app or web tool and record at home or out in the world on their own devices. This scales across countries and accents quickly, and the acoustic variety is real, which helps a model that will hear real users. The vendor controls less: rooms, microphones and noise vary, so speaker screening and QA carry the weight.
Supervised sessions. A moderator runs each session in a studio, a rented home, a car or the client's site, on equipment the vendor or the client provides. This is the route for prototype hardware, a fixed microphone setup, in-car audio under controlled conditions, or a studio voice for text-to-speech. Fewer companies offer it, and every session needs a booked space and staff.
Catalog first, collection on top. Some vendors are built around ready-made datasets and record new data in the same format when a language or domain is missing. That suits teams who already like a catalog's conventions and want the gaps filled to match.
Most large vendors run the first two side by side. The table shows what each company says on its own site.
Speech data collection companies compared
| Company | How it records, per its site | Languages, in its own words | Ready-made datasets too | Source |
|---|---|---|---|---|
| Appen | Remote on its multi-device platform; supervised sessions at its facilities, customer sites, studios, rented homes or in cars | 500+ locales overall; ready-made datasets in 80+ languages | Yes | Data collection |
| DataForce (TransPerfect) | Crowd mobile app; supervised sessions at its offices, your facility or a rented location; remote voice talent in a cloud recording tool; TransPerfect studios | Over 250 languages and dialects | None found | Data collection |
| LXT | From home by PC or phone app; at a nearby location; moderated sessions in a studio, at client premises or in LXT centers | 150+ countries and 1,000+ locales | Unclear: its datasets pages describe custom builds | Audio data collection |
| TELUS Digital | Contributors' own devices; in-cabin automotive programs | 500+ languages and dialects for annotation | Yes | Data for AI training |
| Defined.ai | Global crowd; studio-quality emotional recordings; voice talent for TTS | 500+ languages and dialects | Yes | Data collection |
| Shaip | Mobile app or desktop; recordings in restaurants, offices and homes | Over 100 languages | Yes | Speech data collection |
| FutureBeeAI | Contributor community; on-site at your location; specific microphones; studio TTS with voice talent | 100+ languages | Yes | Audio data collection |
| Alconost | Scripted reading in a studio or remotely; two-speaker dialogue on separate channels; phone channel; mobile app | Any language and dialect, including rare ones | Yes, some sets | Voice and speech data |
| David AI | Not described; its datasets are channel-separated conversations | 15+ languages in its multilingual set | Yes | Homepage |
| SpeechData.ai | Native speakers on phone, headset, far-field or in-car audio, as the brief specifies | 60 languages and dialects, others after a feasibility check | Yes, 60 conversational datasets | This site |
"None found" and "Unclear" mean we found no list of named ready-made speech datasets on that company's site, not that none exist.
The companies, one by one
Appen
Appen's data collection page describes two setups: remote collections, in which its workforce records on Appen's multi-device platform at home or in public places, and fully supervised on-site sessions at Appen facilities, customer sites, professional recording studios, rented homes or in cars. In an automotive case study, it recruits native speakers to a carmaker's target demographics in more than 20 languages and records utterances in various environments, including a recording booth that simulates driving. It also lists more than 360 ready-made datasets in 80+ languages, with speech among the data types.
DataForce (TransPerfect)
DataForce, part of TransPerfect, lists scripted or conversational speech, wake words, multi-speaker dialogues and in-car recordings under controlled conditions on its data collection page. It records four ways: crowd contributors through its mobile app; supervised sessions in its offices, at your facility or at a rented location; trained voice talent working remotely in its StudioNEXT cloud recording interface; and sessions in TransPerfect's studios, which it uses for studio-quality TTS data.
It says its talent database holds more than 1,000,000 collaborators covering over 250 languages and dialects. In one case study, more than 500 participants per language in Mandarin and Korean each completed ten sessions in its phone app, at different locations and times of day.
LXT
LXT's audio data collection page covers scripted prompts (text lines, keywords and command sets, with wake words as one use), spontaneous or guided dialogue between two or more speakers, emotional speech, and speech captured in cars, offices, streets and homes. Contributors record from home on a PC or a phone app, at a nearby location by phone, or in a moderated session in a studio, at the client's premises or in LXT's own data collection centers.
It says it has over 10 million contributors across 150+ countries and 1,000+ locales. Its site menu has an off-the-shelf datasets section, but its speech recognition page there describes datasets built to order, backed by a supplier network with rights-cleared conversational speech in 100+ languages, and we found no list of named ready-made speech datasets.
TELUS Digital
TELUS Digital runs large managed data programs in which speech is one part. In a voice assistant case study, contributors recorded prompts on their own devices, and TELUS Digital used IP address tracking to validate their local addresses, because the client needed local native speakers. Over five years the program delivered more than 4 million audio prompts and utterances from more than 5,000 contributors in 50 geographic regions, and the client's assistant gained 35 languages. Its data for AI training page cites a community of more than 1 million contributors, 500+ annotation languages and dialects, and ready-made datasets that include audio and speech recognition.
Defined.ai
Defined.ai runs a marketplace of ready-made datasets and also collects to order. Its speech recognition page cites more than 2 million hours of speech data, including scripted, podcast and call center recordings, and 500+ languages and dialects covered, and offers custom TTS collection with voice talent reading phonetically balanced scripts. Its data collection page lists real-world conversations, domain-specific voice prompts for call centers and automated systems, and studio-quality emotional recordings, from a crowd it puts at 1.6 million contributors in 150+ countries.
Shaip
Shaip licenses ready-made speech datasets and collects to order, according to its speech data collection page. It lists scripted single-speaker monologues, two-person dialogue on dual-channel files, multi-person discussions, wake words and key phrases, and recordings in acoustic settings such as restaurants, offices and homes. Its dataset listings name a mobile app or desktop as the recording platform. It says it works in over 100 languages with 30,000+ collaborators.
FutureBeeAI
FutureBeeAI's audio data collection page lists multi-person and agent-customer conversations, scripted prompts and monologues, unscripted monologues, wake words and commands (including in a car), emotional speech, and studio-recorded TTS speech from professional voice talents. Besides its contributor community, it offers on-site collection at a location you choose and recording on specific microphones or devices. It says it works in 100+ languages with more than 20,000 contributors in 50+ countries, and it also sells ready-made speech datasets.
Alconost
Alconost, which also sells translation and localization, describes scripted monologues recorded in a studio or remotely, two-speaker dialogue on separate left and right channels, phone-channel recordings, emotional speech and capture through a mobile app on its voice and speech data page. For voice cloning, it offers certified-studio recordings with an exclusive buyout that includes cloning rights for each voice. It says it records any language and dialect, including rare ones, and it sells some ready-made sets, such as 70 hours of Arabic emotional speech in five dialects.
David AI
David AI describes itself as an audio data research company. Its homepage lists ready-made conversation datasets: Converse (channel-separated, natural two-speaker English conversations), Atlas (the same format in 15+ languages, with dialect and accent metadata), Chorus (three or more speakers) and Dialog (expert conversations). It also partners with research teams to design new kinds of data, and its stated process runs a targeted collection first and scales it to thousands of hours only after quality has been measured and tuned. The site does not say whether speakers record remotely or in a studio.
SpeechData.ai
SpeechData.ai, a Spirelight company, is ours. It licenses a catalog of 60 conversational datasets, 73,000 hours of spontaneous two-person conversation from 7,250 native speakers on dual-channel audio, and records new data to spec.
For custom work, native speakers are recruited to quotas for age, gender, region and accent, pass an ID check and a microphone test, and record sentence or prompt lists, monologues or spontaneous dialogue on phone, headset, far-field or in-car audio, as the brief specifies. Sessions that miss spec are re-recorded rather than shipped, and every speaker's commercial-use consent is documented and delivered with the data. It covers the catalog's 60 languages and dialects, plus others such as isiZulu, isiXhosa and Sesotho after a recruitment feasibility check.
None of its published services cover studio or on-site sessions, so if you need a supervised studio or a prototype device recorded on your premises, one of the vendors above fits better.
Which setup your project needs
The recording setup follows from what the model will hear in production. A few common cases, matched to what the companies above describe:
- A wake word for a phone app can be recorded remotely on the kinds of phones your users own, as long as the brief sets accent quotas and includes phrases that sound like the wake word. DataForce, LXT, Shaip, FutureBeeAI and SpeechData.ai all list wake word recording.
- In-car voice control needs audio from inside a car. Appen, DataForce, FutureBeeAI, LXT and SpeechData.ai describe in-car recording, and Appen and DataForce describe simulated or controlled driving conditions.
- A text-to-speech voice needs a few chosen speakers recorded consistently. DataForce, Defined.ai, FutureBeeAI and Alconost describe studio recording or professional voice talent, and at SpeechData.ai you can approve each voice from a short sample.
- Conversational ASR across accents needs two-person dialogue, ideally with one speaker per channel. Shaip, Alconost and David AI describe separate channels for each speaker, and SpeechData.ai's catalog is recorded that way, with several varieties of the same language off the shelf.
- Prototype hardware or a fixed microphone setup usually means on-site sessions. Appen, DataForce, LXT and FutureBeeAI describe recording at the client's site.
How to choose a speech data collection company
Custom collection goes wrong in the details, so ask every company on your shortlist the same questions and compare the answers in writing.
- Recruitment and accent quotas. How do you find speakers for each dialect, how do you confirm they are native speakers from that region, and will you commit to quotas per accent, age band and gender? What happens when a quota can't be filled?
- Device and environment control. Which microphones and rooms will speakers use, who checks the background noise before recording, and can you match the devices and acoustic conditions your product will meet?
- QA and re-recording. Which automatic checks run on every file, how much does a person listen to, and when a session misses spec, is it re-recorded at the vendor's cost or simply dropped from the count?
- Transcription conventions. Will transcripts follow your guideline for fillers, false starts, numbers, overlap and code-switching, verbatim or clean? Ask for the guideline and a transcribed sample.
- Consent documentation. What exactly did each speaker sign, does it cover commercial training and deployment, and do the signed records come with the data?
- Ownership. Do you own the recordings or license them, and can the vendor reuse or resell the same audio?
- Sample or pilot terms. Can a small batch be recorded to your spec before you commit to the full order, and on what terms?
- A test set. Can a separate group of speakers be recorded to the same spec as a held-out test set, so you measure the model on voices it never trained on? Our guide to ASR test sets explains how to build one.
Our speech data RFP template turns these questions into a document you can send to every vendor at once.
Where to start
To have native speakers recorded to your quotas, in one of our 60 languages and dialects or another language after a feasibility check, send your brief through our custom collection page for a feasibility assessment and a quote within one business day, or start from the wake word page if that is your project.
