SignalMatrix · Pre-seed 2026
The internet is silent.
The airwaves aren't.
We turn live TV and radio into quality speech data for the big languages AI still can't hear.
Broadcast in. AI-ready data out.
One pipeline. Two products.
2,000 channels
Live TV and radio, captured as it airs.
SignalMatrix
Intelligence feed
Training data
AI ran out of internet.
Quality data is now the bottleneck.
“We've achieved peak data and there'll be no more.”
Ilya Sutskever, OpenAI co-founder · NeurIPS, December 2024
$14.3B
Meta paid for 49% of Scale AI, a data company, in June 2025.
$1.2B
Surge AI's estimated 2024 revenue, selling human data to frontier labs.
10 months
David AI went from a $5M seed to a $50M Series B, selling audio data alone.
Sources: The Verge / NeurIPS 2024; Meta–Scale deal reporting, June 2025; Forbes via Sacra (Surge, press estimate); David AI and Bloomberg, 2025.
Big languages. Almost no data.
Hours of transcribed training audio in OpenAI's Whisper, the reference speech model.
Sources: Whisper (arXiv 2212.04356); Ethnologue. Bars on a log scale. Kurdish, Oromo, Igbo, Zulu, Tigrinya and Wolof: also 0 hours.
These languages live on air.
Not online. On the radio and TV people actually use.
59%
of Africans get news from radio at least a few times a week, across 38 countries. Radio is still the continent's top news source.
Real speech
Talk shows and call-in radio are spontaneous, multi-speaker and dialect-switching. Exactly the speech clean benchmarks miss.
71–99%
of the time, Whisper large-v3 writes Pashto audio in the wrong script. Models fail where real speech lives.
Sources: Afrobarometer Dispatch AD1173 (Round 10, 2024/2025); Pashto ASR benchmark, arXiv 2604.04598 (2026).
Three things have to be true at once.
Nobody has all three.
Capture at scale
Thousands of channels, 24/7, with scheduling and failure recovery.
Who has it today: media monitors, English-first.
Dialect-level ASR
Multi-speaker, code-switching, and cheap enough per hour to run at scale.
Who has it today: labs, trained on clean audio.
Verification
Cross-references, entity linking and provenance for every segment.
Who has it today: nobody.
Heard. Transcribed. Verified.
How one broadcast segment becomes data.
Heard
Live TV and radio captured as it airs, with station, country and timestamp on every segment.
→ recording + metadataTranscribed
Speech-to-text tuned for dialects, with speaker, entity and topic tags.
→ time-coded segmentsVerified
Every segment quality-checked, cross-referenced against independent sources and linked into a knowledge graph.
→ claims with sourcesWhat one segment looks like.
Illustrative output, and the signals pulled from it.
الحكومة قررت رفع أسعار الوقود ابتداءً من الشهر الجاي.
“The government decided to raise fuel prices starting next month.”
- Speakers, up to 8 separated at once
- Entities and topics, linked in a graph
- Claims, each tied to a speaker
- Sentiment, events, contradictions
- Cross-references to other sources
- On-screen figures, via vision
The hard engineering is already done.
Five breakthroughs SignalMatrix has built, where off-the-shelf models fall short.
−60%
CostCheaper per hour than Whisper-class transcription.
5,000+
ScaleChannels our ingestion pipeline is built to carry.
40+
ReachLanguages our transcription stack already handles.
8
Real conversationSpeakers separated at once. Single-speaker models break.
Vision
Beyond audioOn-screen figures tie each claim to who said it.
What we've built.
585,996 hours of real speech, and counting.
Live counts from the production warehouse.
2,000 channels. 27 countries. 8 languages.
Live today, deepest in Arabic, Farsi, Dari and Kurdish.
~2,000
TV channels
27
Countries
8
Languages live
Shaded: countries with recorded broadcast, from the production warehouse. Outlined: next-round countries.
One asset. Two buyers.
Revenue now, scale later.
Intelligence feed
Verified, dialect-tagged broadcast speech, delivered as it airs.
- Risk and narrative intelligence platforms
- Hedge funds tracking emerging markets
- Newsrooms and media groups
Training and evaluation data
Real-world speech for the languages models can't hear yet.
- Frontier AI labs
- Voice AI companies
- Sovereign AI programs
Three markets. One dataset.
Who pays for real-world speech.
$2.8B spent by investment managers on alternative data in 2025, up 17% in a year.
Billions spent each year by AI labs on training data, rising as public data runs out.
Sovereign AI programs in MENA, Africa and South Asia, with almost no native speech data.
Serviceable market: non-English broadcast speech across 50 languages. Estimate to add.
Source: Neudata, “The state of the alternative data market in 2026,” February 2026.
Both markets are already testing the data.
Design partners working with the current corpus today.
A voice-AI company
- Testing dialect-tagged broadcast speech
- As training and evaluation data
A risk-intelligence platform
- Testing the verified broadcast feed
- For narrative tracking
Only SignalMatrix checks all four boxes.
Investors just funded this category. We own the half nobody's taking.
The loop that compounds.
Every hour we collect makes the next one cheaper.
Can't be recaptured
Past broadcasts are gone once they air. Competitors can't get them back.
Slow to copy
Broadcaster licenses take time to sign.
Built by the people who built Arabic AI.
Dialect-level ASR and data you can trust are both hard. This team has done each before.
Abdelrahman Mansour
- 15+ years building Arabic-language media and information platforms.
- Co-founder of Matsadaash, a leading Arabic information platform.
- Grew funding for information initiatives from $0 to $8M.
Abu Bakr Soliman
- Built SILMA 1.0, an early Arabic LLM that beat Gulf sovereign models on the Open Arabic LLM Leaderboard.
- Author of AraVec, among the most-cited Arabic NLP work.
- Led AI at Unifonic, including a leading MENA social listening platform.
The road to 50 languages.
Every new language is a new corpus on the same pipeline, not a rebuild.
8
Live todayArabic and its major dialects, Farsi, Dari, Kurdish and four more.
20
This roundPashto, Urdu, Hausa, Swahili and more high-speaker, low-data languages. First paying contracts.
50
NextAcross Africa and Asia: Amharic, Yoruba, Somali, Bengali and beyond.
$3M for 20 languages and first revenue.
Pre-seed on a SAFE, room to $4M. About 24 months of runway.
$3M
Use of funds- 40% capture, new languages and broadcaster rights
- 30% machine learning and data quality
- 20% go-to-market and partnerships
- 10% infrastructure and reserve
By the end of the round
- 20 languages live in production
- First paying contracts across feeds and data
- Rights-cleared track signed with first broadcasters
Vision
Every language AI can hear.
The default supplier of real-world speech for every language AI can't hear yet.
50
languages
5,000+
channels
24/7
captured and verified
The world's speech disappears as it airs. Help us keep it.