MediaMotionAfrica's media archives, made training-ready.
MediaMotion partners with South Africa's leading media networks to license, process and deliver thousands of hours of multilingual audio and video — Afrikaans, isiZulu, English, and the code-switching in between — for frontier labs and AI startups.
Speech the way it is actually spoken.
Not read-aloud sentences in a quiet room. Real programmes, real arguments, real accents — with the noise, overlap and language-switching that models fail on in production.
Talk radio & phone-ins
Spontaneous, unscripted callers across provinces. Heavy code-switching, regional accents, real background noise.
Soapies & studio drama
Scripted dialogue, multiple camera angles, clean face tracks. Emotional range that read-speech corpora never capture.
Podcast conversations
Two to four speakers, 40–120 minutes, near-field mics. Interruptions, laughter, back-channels — the texture of real dialogue.
News & current affairs
Read anchor speech paired with field interviews and vox pops. Named entities, place names and dates that generic corpora get wrong.
Sports commentary
Rugby, football and cricket. Rapid, high-energy speech with crowd noise and two commentators talking over each other.
Panel & debate shows
Four to six voices switching between isiZulu, Afrikaans and English mid-sentence. The hardest diarisation problem we have — and the most valuable.
From tape vault to training set, without the grey areas.
Partner with the network
We sign directly with broadcasters, radio groups and studios. Every catalogue is audited for chain-of-title before a single frame is ingested.
- Licence scoped to AI training only
- Talent & guild clearances reviewed
- Excluded-content list honoured
Make it legible to a model
Our pipeline runs language identification, speaker diarisation, verbatim transcription and word-level alignment — then native speakers check it.
- Code-switch boundaries tagged per span
- Speaker IDs consistent across episodes
- Human QA on a stratified sample
Ship exactly what you asked for
Clips, manifests and metadata in your bucket, with a provenance record for every file. Start with a five-hour sample; scale to the full catalogue.
- S3 / GCS delivery, signed URLs
- JSON, RTTM, TextGrid, WebVTT
- Custom curation briefs welcome
The unglamorous work that makes data usable.
Language ID & code-switch tagging
Per-span language labels, including within-utterance switches. isiZulu, Afrikaans and English at launch.
Speaker diarisation
Who spoke when, with speaker identities kept consistent across a series. Overlap regions marked, not dropped.
Verbatim transcription & translation
Disfluencies kept. Orthography checked by native speakers. Parallel English translation available on request.
Word-level alignment
Timestamps for every token, so you can cut clean clips, build TTS pairs or train streaming ASR.
Rights, consent & provenance
A signed licence, a clearance summary and a hash-linked provenance record shipped with every file.
Redaction & compliance
PII in phone-ins, minors on camera, and POPIA obligations handled before delivery — not after a complaint.
Built with the networks, not around them.
South Africa's broadcasters have decades of programming in eleven official languages. Almost none of it has ever been on the open web. We are the licensed route in — starting with the three languages that cover most airtime, and expanding partner by partner.
Your archive is an asset. Treat it like one.
Broadcasters, radio groups, studios and podcast networks: licence content you already own into a new revenue line, on your terms.
- Non-exclusive or exclusive licensing, training-use only
- You keep copyright, distribution and the veto
- Free catalogue audit and valuation
- Provenance ledger and usage reporting
Speech your model has never heard.
Frontier labs and AI startups: pre-training corpora, fine-tuning sets and uncontaminated evaluation data for African languages, delivered clean.
- ASR, TTS, speech LLMs and multimodal video
- Five-hour sample before any commitment
- Custom curation briefs — by dialect, domain, speaker mix
- Licence terms written for training, not for broadcast
The models everyone is shipping were trained on a tiny slice of how the world actually talks. Roughly a fifth of a billion Africans speak the languages in our first three collections. That gap is a product, not a footnote.— Founding note, MediaMotion
Tell us what you're building — or what you're sitting on.
We reply to every message personally. Model builders get a sample set; data providers get a free catalogue audit.
- MODEL BUILDERS & PARTNERSpartnerships@mediamotion.ai
- DATA PROVIDERSsales@mediamotion.ai