We ship Voice Tables at Inithouse. It's an agentic AI workspace you control with your voice. Describe what you need (a CRM, an inventory tracker, a project board) and it builds the tables, documents, and data structure for you.
The headline numbers: 30+ supported voice command types and roughly 60 seconds from a spoken idea to a working workspace. We measured both in production. Here's what those numbers actually mean.
What happens in those 60 seconds
The pipeline runs four stages:
- You open a workspace and tap the mic icon.
- Whisper transcribes your speech to text.
- An LLM parses intent and maps it to a function call (create table, add column, insert row, run query).
- The workspace executes the command and updates the UI.
The table-creation path, something like "I need a client tracker with name, email, status, and last contact date," consistently lands under 60 seconds including the signup. The voice-to-structured-data conversion is the piece that makes this a voice-first tool rather than another spreadsheet with a microphone bolted on.
The 30+ command types, mapped
We categorized every supported voice command by what it produces and where the pipeline tends to break.
| Command type | Example | What it creates | Pipeline reliability |
|---|---|---|---|
| Create workspace | "Build me a CRM for freelance clients" | Full table with columns + sample rows | High |
| Add row | "Add a new client: Acme Corp, active, today" | Single row with parsed field values | High |
| Add/edit column | "Add a priority column with high/medium/low" | Column definition with type inference | Medium |
| Query/filter | "Show me all clients from last week" | Filtered table view | Medium |
| Create document | "Write a meeting summary for today" | Linked document in workspace | High |
| Bulk edit | "Mark all overdue tasks as urgent" | Batch row updates | Low |
| Delete/undo | "Remove the last three rows" | Deletion + undo state | Medium |
The pattern: commands with clear structure (create this schema, add this row) succeed reliably. Commands requiring contextual reasoning (filter by relative dates, batch-edit with conditions) fail more often because errors compound across the Whisper transcription and the LLM parsing stages.
Where the pipeline breaks
We tracked voice_recording_started, voice_transcribed, and voice_transcription_error events in production. Early data from a small user sample showed transcription completing successfully in roughly one out of four attempts. The rest hit an error before the LLM even saw the text.
The root causes split into two categories.
Whisper-side:
- Background noise degrades transcript quality. Our target users (field workers, tradespeople) are often in loud environments.
- Non-English accents produce garbled output the LLM can't parse.
- Short commands ("add row") sometimes get swallowed or misheard.
LLM-side:
- Ambiguous intent: "update the table" without specifying which table or what to change.
- Schema drift: the model proposes a column type that doesn't match existing data.
- Compound commands: "create a table and add five clients from my last project" gets partially executed.
The honest takeaway is that voice-first data entry at production quality is harder than voice-first search or voice-first note-taking. The structured output requirement (specific columns, typed values, relational data) is what makes it brittle.
What we're learning across the portfolio
We see a version of this problem in other tools we build at Inithouse. Verdict Buddy, which takes unstructured conflict scenarios and produces framework-based verdicts, taught us that the parsing layer needs explicit guardrails rather than just a larger model. Tarotas, our tarot reflection app, takes a much simpler input (draw a card) and produces a grounded interpretation, which is partly why its pipeline is more stable.
For Voice Tables, the next iteration focuses on running Whisper with noise suppression for field environments and adding confirmation steps before the LLM executes destructive commands like deletes and bulk edits.
Try it
Voice Tables is live. The workspace builder works. The voice pipeline is where we're putting the engineering hours now, and we're publishing the numbers because that's how we work at Inithouse.
If you're building voice-to-structured-data systems, we'd be curious what pipeline architecture is working for you.
Top comments (0)