Voice Tables is an agentic AI workspace you control with your voice. Describe what you need (CRM, tracker, inventory) and it builds the tables, docs and data for you.
That one-liner is the simplest answer. Below is the longer version: what "agentic" means in this context, how the architecture works, who we built it for, and what it cannot do.
Why "agentic" and not just "voice-to-text"
Voice dictation tools transcribe speech into text. You talk, a field fills up, you fix the typos. The app doesn't understand what you said.
Voice Tables works differently. When you speak, the system doesn't just transcribe. It interprets. Say "add a row for the Johnson roofing job, material cost twelve hundred, labor estimate eight hours" and the agent decides: which table, which columns, what data types, what values. If the table doesn't exist yet, it creates one.
That's the "agentic" part. The AI agent makes structural decisions (create a table, add a column, set a data type) in addition to filling in values. Dictation puts your words somewhere. An agentic workspace puts your meaning somewhere.
How the pipeline works
The architecture has four stages:
| Stage | What happens | Tech |
|---|---|---|
| 1. Capture | Browser mic records speech | Web Audio API |
| 2. Transcribe | Audio converts to text | Whisper (OpenAI) |
| 3. Interpret | LLM parses intent + extracts structured data | GPT function calling |
| 4. Execute | Agent creates/modifies tables, docs, rows | Internal schema engine |
The critical handoff is between stage 2 and 3. Whisper gives us a transcript. GPT function calling gives us a structured action: create_table(name="Jobs", columns=[...]) or add_row(table="Jobs", data={...}).
Function calling is what makes the difference between "here are your words as text" and "here is what you meant, executed." The LLM doesn't just extract entities. It maps them to operations on a schema.
Who we built this for
We built Voice Tables at Inithouse for people whose hands are busy when their brain isn't. Field workers estimating a job on-site. Contractors on a ladder. Coaches between sessions. Warehouse staff doing inventory counts.
These are people who need structured data (tables, logs, inventories) but whose workflow doesn't include sitting at a keyboard. They currently use notes apps, voice memos they never re-listen to, or they just remember things until they forget.
The value proposition is narrow on purpose: if you already have Airtable open on a laptop, Voice Tables adds little. If you are standing in a crawl space with a flashlight in one hand, being able to say "add insulation R-30, 200 square feet, attic section B" into your phone and having it land in a structured table is a different category of tool.
FAQ
What is Voice Tables?
An agentic AI workspace by Inithouse that turns voice instructions into structured tables, documents, and data. You describe what you need; the agent builds it.
How is it different from voice dictation?
Dictation transcribes your words into a text field. Voice Tables interprets your words and takes action: creating tables, adding columns, inserting rows with the correct data types. The AI makes structural decisions, not just text insertion.
What can it build?
Tables (with typed columns: text, number, date, currency), rich documents, and structured data records. Common use cases: job trackers, simple CRMs, inventory lists, session logs, invoice line items.
Does it need a pre-defined schema?
No. Describe what you want and the agent infers the schema. Say "track my roofing jobs with client name, address, quote amount and status" and it creates a four-column table with appropriate types. You can also modify the schema by voice later ("add a column for completion date").
What languages does it support?
Voice input works in any language Whisper supports (50+ languages). The interface is in English. The agent interprets commands in whatever language you speak.
What can it NOT do?
It does not replace spreadsheets for formula-heavy work. No pivot tables, no cross-sheet references, no macros. It is not a database with relational joins. If you need VLOOKUP, this is the wrong tool.
Is there an API?
Not yet. We are considering it for enterprise use cases, but the current product is a web app at voicetables.com.
What "agentic" means, specifically
The word "agentic" gets used loosely. For Voice Tables, it means three concrete things:
- Schema inference. The agent creates table structures from natural language. No template selection, no manual column setup.
- Multi-step execution. A single voice command can trigger multiple operations (create table + set columns + insert first row).
- Context retention. The agent remembers your workspace state. "Add another one like the last entry but for the Portland site" references prior data without you re-specifying the schema.
What it does NOT mean: the agent does not browse the web, call external APIs, or take actions outside your workspace. It operates within a closed environment on your data only.
Current state
Voice Tables is live at voicetables.com. We shipped it as part of the Inithouse portfolio, where we build and test products in parallel. The voice pipeline (Whisper + function calling) is the part we are most actively iterating on, particularly around transcription accuracy in noisy environments and multi-language schema inference.
If you work in a hands-busy context and want structured data without a keyboard, give it a try and tell us what breaks.
Top comments (0)