MediaMotionSpeech your model has never heard.
Web-scale corpora are thin on African languages and thinner still on how they are really spoken — fast, overlapping, switching mid-sentence. We deliver licensed broadcast, radio and podcast data with the annotations to make it useful on day one.
Built for the models people are actually training.
ASR for low-resource languages
Verbatim transcripts with word timestamps across telephone, studio and broadcast conditions. Spontaneous speech, not read prompts.
TTS & voice models
Clean studio speech from consented presenters and performers, segmented into utterance-level pairs with normalised text.
Speech LLMs & audio-native models
Long-form multi-speaker conversation with diarisation, so models learn turn-taking, prosody and code-switching, not just words.
Dubbing & speech translation
Parallel isiZulu / Afrikaans ↔ English translations aligned at the segment level, with speaker and emotion tags.
Multimodal video understanding
Broadcast video with face tracks, shot boundaries, on-screen text and aligned dialogue. Multi-camera drama and studio formats.
Evaluation & benchmarks
Held-out sets that have never touched the open web. Stratified by language, accent, noise and speaker count, with confidence intervals reported.
Formats your data loader already understands.
Media
Annotations
Different stages, different data.
Volume and variety
Thousands of hours of lightly-labelled broadcast audio and video across every collection — for self-supervised and semi-supervised pre-training.
- Language ID + diarisation only
- Priced per hour, volume tiers
Precision and coverage
Fully transcribed, aligned, QA'd subsets curated to your brief — a dialect, a domain, a speaker demographic, an acoustic condition.
- Word-level alignment, translation
- Custom curation from a brief
Clean and held-out
Sets we never license for training, stratified so you can report per-language, per-condition numbers with confidence.
- Guaranteed off-web
- Eval-only licence
Terms written for training, not broadcast.
- Perpetual training licence for models you build
- Model outputs unrestricted; redistribution of the data prohibited
- Indemnity on rights — we hold the upstream licences
- Provenance record for every file, for your own audits
- Eval sets on a separate, non-training licence
Brief. Sample. Curate. Deliver.
Tell us the languages, hours, conditions and annotation depth you need. We send a five-hour annotated sample in your formats, iterate on the spec, then deliver to your bucket.
- Sample turnaround: 5 working days
- Delivery via S3 / GCS with signed URLs
- Ongoing drops as new catalogues clear
The honest answers.
How do you handle code-switching in transcripts?
Every transcript carries per-span language tags, down to individual words where a switch happens mid-sentence. We use the language's own orthography for each span rather than romanising or "translating" it, and native-speaker QA specifically checks switch boundaries.
Can we get audio-only from video sources?
Yes. Many video collections are delivered as audio stems with the video available as an add-on. Where a programme has separate dialogue and music/effects stems, we deliver the dialogue stem.
What is the speaker demographic coverage?
We report gender, approximate age band and home-language where it is known or inferable, per speaker. Phone-in collections skew toward broad public demographics; studio collections toward professional presenters. We can curate to a target mix.
Is there a minimum order?
No. The five-hour sample is free; after that you can license as little as 50 hours. Pre-training volume tiers start at 1,000 hours.
Which languages next?
isiXhosa and Sesotho are in diligence with partners now. The full list of eleven official languages is on the roadmap, sequenced by partner catalogue availability. Tell us what you need and it will move up the list.