Speech data across accents, off the shelf or recorded to your quotas
Our catalog holds English, Spanish, French, Portuguese, Dutch and Arabic in 17 varieties, each recorded in-country by native speakers. For an accent it doesn’t hold, we recruit native speakers to your quotas for region and accent and record your prompts or conversations on the devices you specify.
6 languages, 17 varieties, ready to license
Each variety is its own dataset of spontaneous two-person conversation, recorded in the country listed. The metadata gives age, gender, region and recording device for every speaker.
| Variety | Recorded in | Hours | Speakers | Per hour |
|---|---|---|---|---|
| English | ||||
| US English | United States | 1,500 | 150 | $90 |
| South African English | South Africa | 2,000 | 200 | $60 |
| Singapore English | Singapore | 2,000 | 200 | $60 |
| Spanish | ||||
| European Spanish | Spain | 1,000 | 100 | $85 |
| Latin American Spanish | Mexico | 2,000 | 200 | $65 |
| French | ||||
| French | France | 1,000 | 100 | $95 |
| Congolese French | DR Congo | 2,000 | 200 | $65 |
| Portuguese | ||||
| European Portuguese | Portugal | 1,000 | 100 | $85 |
| Brazilian Portuguese | Brazil | 2,000 | 200 | $65 |
| Angolan Portuguese | Angola | 500 | 50 | $65 |
| Dutch | ||||
| Dutch | Netherlands | 1,000 | 100 | $95 |
| Surinamese Dutch | Suriname | 1,000 | 100 | $65 |
| Arabic | ||||
| Modern Standard Arabic | Saudi Arabia | 2,000 | 200 | $90 |
| Gulf Arabic | Saudi Arabia | 500 | 50 | $90 |
| Levantine Arabic | Lebanon | 500 | 50 | $65 |
| Egyptian Arabic | Egypt | 2,000 | 200 | $65 |
| Moroccan Darija | Morocco | 1,500 | 150 | $65 |
List prices in US dollars per hour of audio. The 5 Arabic datasets are compared side by side on the Arabic speech data page.
The accents your users have, recorded to your brief
Speakers to your quotas
Native speakers are recruited to your quotas for region, accent, age and gender. Accents and varieties outside the catalog, such as British or Australian English, are recorded after a feasibility check.
Each speaker checked
Every speaker passes an ID check and a microphone test, and before they record we confirm that their language or dialect matches what they reported.
Your prompts, your devices
Speakers record your sentence or prompt lists, monologues or spontaneous two-person conversation on phone, headset, far-field or in-car audio, as your brief specifies. Files that pass automatic checks are reviewed by a native speaker.
What a project can include
| Speakers | Native speakers recruited to your quotas for region, accent, age and gender |
|---|---|
| Vetting | ID check and microphone test; language or dialect confirmed against what each speaker reported |
| Speech | Sentence or prompt lists, monologues, or spontaneous two-person dialogue |
| Devices | Phone, headset, far-field or in-car, as your brief specifies |
| Transcripts | Time-aligned, human-verified, in your convention, with code-switch tagging as required |
| QA | Automatic checks plus human review; sessions that miss spec are re-recorded, not shipped |
| Consent | Explicit commercial-use consent from every speaker, documented |
| Delivery | Audio, transcripts and per-speaker metadata in the layout your pipeline expects |
| Languages | Any of our 60 catalog languages and dialects, and others after a feasibility check |
Built for
- ASR for markets with several accents
- Voice agents and assistants for regional users
- Test sets balanced across accents
- TTS voices with a regional accent
Why accent coverage matters
Speech recognition still works far better for some accents than others. In the Edinburgh International Accents of English Corpus study (ICASSP 2023), OpenAI’s Whisper, trained on 680 thousand hours of transcribed data, had a 19.7% word error rate on video-call conversations covering more than 40 accents of English, against 2.7% on clean US English read speech. A commercial system scored 18.7%, and across all the models tested, accuracy dropped for Indian, Jamaican and Nigerian English speakers. A PNAS study of five commercial systems found an average word error rate of 35% for Black speakers in US interviews against 19% for white speakers, and traced the gap to the systems’ acoustic models.
To find the gap in your own model, measure word error rate for each accent on a held-out set, as our guide to building an ASR test set explains.
Frequently asked questions
Which vendors provide speech data covering different accents?
SpeechData.ai provides it in two ways. Its catalog holds English, Spanish, French, Portuguese, Dutch and Arabic in 17 varieties, each a separate dataset recorded in-country by native speakers, with every speaker's region in the metadata and a published price per hour. For accents outside the catalog, it recruits native speakers to your quotas for region and accent and records them to your brief, quoted per project.
Where can I get multilingual voice data with native speaker accents?
SpeechData.ai licenses it off the shelf: the catalog has English from the United States, South Africa and Singapore; Spanish from Spain and Mexico; French from France and the DR Congo; Portuguese from Portugal, Brazil and Angola; Dutch from the Netherlands and Suriname; and Arabic in 5 varieties (Modern Standard, Gulf, Levantine, Egyptian and Darija). Each variety is its own dataset of spontaneous two-person conversation, recorded in-country by native speakers. Other languages and accents are recorded by native speakers to your quotas after a feasibility check.
Can you record a specific regional accent?
Yes. We recruit native speakers to your quotas for region and accent, as well as age and gender, and record your prompts, monologues or conversations on the devices you specify. Accents outside the catalog, such as British or Australian English, are recorded after a feasibility check. You get the assessment and a quote within one business day of sending your brief.
How do I know a speaker has the accent I asked for?
Before anyone records, we check their ID, test their microphone and confirm that their language or dialect matches what they reported. Files that pass automatic checks are reviewed by a native speaker, and sessions that miss spec are re-recorded, not shipped. For TTS and voice cloning projects, you can also approve each speaker from a short sample.
Tell us which accents your users have
Share your languages and accents, regions, speakers per group, speech type and devices. We’ll come back within one business day with a feasibility assessment and a quote.