Numanac is now part of the ElevenLabs Startup Grants program.
Alma, the AI agent inside our voice-first platform, turns what agronomists, growers, and field crews say into the record their operation runs on. Every report, recommendation, and task Alma produces starts as a transcript. When the transcript is wrong, everything built on it is wrong too, often in ways nobody notices until later.
The grant gives us a year of access to ElevenLabs models. We are putting it toward the part of our stack that matters most: transcribing agricultural speech accurately, in the places where it is actually spoken.
Agronomy is spoken regionally
Building Numanac has taught us how regional agronomic communication is. It varies within the same crop in a single state. Two agronomists a few hours apart can describe the same field problem with very little vocabulary in common.
The differences show up everywhere:
- Accents and speech patterns that shift from one growing region to the next
- Local names for pests, weeds, and diseases
- Product trade names, active ingredients, and varieties
- Shorthand for growth stages, practices, and equipment
General-purpose speech models hear very little of this. They are trained mostly on everyday conversation, meetings, and broadcast audio. Field speech adds wind, engines, and a speaker walking a row with the phone in a pocket. Accuracy drops fast, and it drops hardest on the words that carry the most meaning.
One wrong word travels
An agronomist says "canopy." The transcript reads "can of peas."
A person reading the note would catch it. A system extracting structure from the note may not. "Canopy" carries real information: canopy closure, canopy cover, whether a spray reached into the canopy. "Can of peas" carries none of it. At worst it gets read as a crop or a product.
On a platform like ours, that error does not stay in one sentence. Alma uses the context captured in the field to:
- Build reports and activity digests
- Inform recommendations
- Support agronomic decisions
- Coordinate the operational work that follows
All four read from the same record. One misheard term can become a wrong observation in a digest, a missing input to a recommendation, or a task that never gets created. The people who pay for the error are rarely the ones who heard the original audio.
Signal in the field is not on or off
Most software treats connectivity as a switch. Fields do not work that way. Signal drops at the end of a row, comes back near the truck, and sits at one bar for much of the day.
We learned this in the field.
Low connectivity is its own environment, and Alma treats it as one, alongside fully offline and fully online work.
A layered system
Our answer is a layered system rather than a single model. It improves transcription at the moment of capture, whether the user is offline, on one bar, or fully connected, and it quietly calibrates to each user over time.
The aim is easy to state. A transcript should hold up to agronomic language in the conditions where it is spoken, and get better the more someone uses it.
Where ElevenLabs fits
ElevenLabs models are central to how we handle the inputs that break general-purpose transcription: regional accents, agronomic vocabulary, and audio recorded in working field conditions rather than quiet rooms.
Access at volume matters here. The system improves with use, and a year of abundant access lets us run it across the full range of crops, regions, and people already using Alma, without rationing.
The ElevenLabs Startup Grants program gives early-stage companies twelve months of access to the ElevenLabs platform. We are glad to be part of it.
Thanks
Many thanks to Jonathan Chang for welcoming us into the ElevenLabs ecosystem.
What's said in the field should become the record an operation runs on, word for word. If you work in the field and there is a term your phone never gets right, send it to us. Those are the words we want to fix first. Our support team: support@numanac.com | Our CEO: daniel.lee@numanac.com
FAQ
What is Alma?
Alma is the AI agent inside Numanac, a voice-first platform for capturing agricultural field context. Agronomists, growers, and field crews speak their observations, and Alma turns them into structured records, reports, and recommendations. Ag retailers, growers, and equipment makers license Numanac as white-labeled infrastructure.
Why is agricultural speech hard to transcribe?
It is regional, technical, and recorded outdoors. Accents, local pest and product names, and growth-stage shorthand vary even within one crop in one state. General-purpose speech models see little of this vocabulary, and field noise makes it harder.
What happens when a field note is mistranscribed?
The error spreads. Alma uses transcribed context for reports, recommendations, agronomic decisions, and operational work, so one wrong term can surface in all of them.
Does Numanac work without cell service?
Yes. Numanac is built to work offline, in low connectivity, and online.
How does Numanac use ElevenLabs?
Numanac uses ElevenLabs models to improve transcription of regional, accented, and agronomic speech recorded in working field conditions. Numanac is part of the ElevenLabs Startup Grants program.