The speech data your models have never heard

Proprietary speech corpora for low-resource languages: native audio, transcription, 14 annotation layers, commercial rights included. Produced in studio with native speakers.

Fon
audio_file
fon_04217.wav
language
fon-BJ
speaker_id
BEN-FON-083
gender
F
age_band
30-39
region
Littoral
speech_type
Spontaneous
split
train
domain
Culture
consent_id
BEN-FON-083-C1
transcription
Mi ɖo gbɛ na mì...
translation_en
We have something to tell you...
duration
6.42 s
snr
41 dB
sampling_rate
48 kHz
30 languages & local varieties650M+ speakers covered14 layers of annotationGold standard multi-annotatorRights secured at the source

Catalog

Languages the web barely has

They are missing from scraped corpora, and therefore from your training data. We record them in studio, with native speakers under contract.

650M+ speakers across these languages

  • Beninese French Beninfr-BJ
  • Fon Beninfon-BJ
  • Wolof Senegalwo-SN
  • Senegalese French Senegalfr-SN
  • Ewe Beninee-BJ
  • Yoruba Beninyo-BJ
  • Hausa Beninha-BJ
  • Pulaar Senegalfuc-SN
  • Bambara Senegalbm-SN
  • Maninka Senegalmlq-SN
  • Darija Moroccoary-MA
  • Moroccan French Moroccofr-MA
  • Swahili Tanzaniasw-TZ
  • Tanzanian English Tanzaniaen-TZ
  • Dioula Côte d'Ivoiredyu-CI
  • Baoulé Côte d'Ivoirebci-CI
  • Ivorian French Côte d'Ivoirefr-CI
  • Fulfulde Cameroonfub-CM
  • Pidgin Cameroonwes-CM
  • Cameroonian French Cameroonfr-CM
  • Twi (Akan) Ghanatw-GH
  • Ewe Ghanaee-GH
  • Ga Ghanagaa-GH
  • Dagbani Ghanadag-GH
  • Ghanaian English Ghanaen-GH
  • Hausa Nigeriaha-NG
  • Yoruba Nigeriayo-NG
  • Igbo Nigeriaig-NG
  • Nigerian Pidgin Nigeriapcm-NG
  • Nigerian English Nigeriaen-NG

Annotation

Every hour is annotated across 14 layers. One corpus, many voice systems

ASR, TTS, speech translation, voice agents, voice servers: the same corpus powers the training, fine-tuning and evaluation of your models and machines.

Text

VerbatimNormalizedITN

Language & pronunciation

Code-switchingTones & geminationG2P

Time & speakers

AlignmentDiarization

Meaning

TranslationSemanticsProsody & emotion

Context & usage

Acoustic eventsIndexingDialogue acts

Gold standard, measured quality

Multi-expert annotation3 independent native speakers, blind, arbitration by a senior linguist
Reliability measured and publishedKrippendorff α, WER, CER, delivered in the datasheet
Strict linguistic complianceofficial orthography, tones and gemination respected
Legal compliancecommercial and AI consent traced per speaker
Linguistic signaturetonal error rate, which no competitor publishes
ASRTTSSpeech translationVoice agentsVoice serversAudio searchEvaluation

Production

This data cannot be scraped. We produce it, end to end

Local studios and mobile units, native narrators under contract, trained transcribers: we control every link, from speaker to delivered dataset.

  1. Narrator contract, GDPR consent
  2. Recording studios, mobile units
  3. Annotation 14 layers, linguist arbitration
  4. Quality + rights Gold validation, IP assignment
  5. Catalog delivery by secure export
  6. Your model training, fine-tuning, evaluation

Our field presence: Labari.io, our free listening app, streams stories and folk tales in local languages and sustains our narrator network. labari.io

Licensing

Three ways to access the data

From a one-off test to language exclusivity: same quality standard, same rights, in every format.

Catalog license

Annual subscription access to one or more languages: search, export and delivery of structured datasets.

Custom dataset

Dedicated production for your needs: language, domain, register, speech type, volume. You specify, we produce in studio.

Temporary exclusivity

An exclusivity window on a language, domain or corpus, before it joins the shared catalog.

Data ready for production

GDPR consenttraced per recording
Intellectual propertycomplete chain of rights, from speaker to dataset
Clear commercial licenseno ambiguity for AI use

Research

What our speech data changes, measured

Whisper Tiny adapted to Beninese French with our data, tested offline on a phone, with a public test set and a reproducible evaluation protocol.

Read the research