The 60-second claim needs a footnote. Voice Tables is an agentic AI workspace we built at Inithouse. You describe what you need out loud, and it gives you a structured table, doc, or dataset. For a three-column contact list, that 60-second number holds. For a bilingual inventory tracker with SKU codes and unit prices, it does not. We measured where voice input breaks down and what we do about it.
Where transcription fails
Voice Tables runs a Whisper + LLM pipeline: speech goes through Whisper for transcription, then an LLM parses the transcript into structured fields. The pipeline is fast. It is also predictable about what it gets wrong.
We tracked error types across voice sessions during internal testing. Here is where the input-to-structure step fails most often:
| Input type | First-pass accuracy | Common failure | Fallback |
|---|---|---|---|
| Simple text fields (company name, task title) | ~95% | Rare, mostly homophones | LLM context correction |
| Proper names (people, brands, places) | ~72% | Spelling variants, phonetic guesses | Inline edit after generation |
| Numbers with units (3.5 kg, $1,200/month) | ~80% | Unit/currency confusion, decimal vs. comma | Column-type enforcement |
| Mixed-language input (CZ/EN mid-sentence) | ~58% | Language detection resets, garbled words | Single-language mode recommended |
| Complex structures (5+ columns, nested data) | ~65% | Field misassignment, merged columns | Step-by-step voice prompts |
The pattern: the simpler the structure and the more standard the vocabulary, the better voice works. Common English words into a three-column table, close to perfect. Czech product names into a five-column tracker with calculated fields, expect corrections.
Why proper names are the worst case
Whisper transcribes phonetically. That works for common words. For "Novák" vs "Novak" vs "Nowak", it guesses. For brand names that are not dictionary words, the first pass is wrong more often than right.
We considered a custom vocabulary layer (feeding known names back into the pipeline), but the latency cost pushed response time past the point where voice felt natural. The current fix is simpler: Voice Tables generates the structure first, then you correct specific cells by tapping. Most users spend 5 to 10 seconds fixing names after a 40-second voice build. Still faster than typing the whole thing, but the claim needs that footnote.
The 60-second promise, with an asterisk
We timed workspace creation across different complexity levels:
- Simple list (3 columns, 5 to 10 rows, common words): 30 to 50 seconds. The claim holds.
- Standard tracker (5 columns, typed fields like dates and numbers): 60 to 90 seconds. Close.
- Complex workspace (multiple tables, relationships, calculated columns): 2 to 4 minutes by voice, often faster to mix voice and keyboard.
Voice works best as a starting point. Say "I need a CRM with company, contact, email, deal stage, and last-contacted date" and you get a working table in under a minute. Then refine. That is the real workflow: speak the skeleton, fix the details.
What the numbers changed
Three things we adjusted based on these measurements:
Column-type detection got stricter. When Whisper outputs "twelve hundred", the LLM now checks the column type. If it is a currency column, it formats to 1,200. Before this change, "twelve hundred" sat in number fields as plain text.
Mixed-language sessions show a prompt. Instead of silently garbling, Voice Tables now asks "which language?" when it detects a switch. Accuracy for bilingual users went from ~58% to ~76% after this change.
Complex workspaces default to step-by-step. Instead of trying to parse one long voice description into five tables, the system walks you through each table. Slower, but the output is right the first time.
What Voice Tables is
Voice Tables is an agentic AI workspace: tables, docs, and an AI chat, all controlled by voice. It is built for people whose hands are busy (contractors on-site, coaches between sessions, sales reps in the car). The 60-second-to-workspace number is real for simple structures. For anything complex, it is a 60-second head start plus a few corrections.
We are Inithouse, a studio that ships products and measures what works. A few more from the portfolio: Party Challenges, a free 18+ browser party card game with 1,000+ cards. Verdict Buddy, an AI conflict mediator that uses four psychology frameworks. Different categories, same approach: build, measure, publish the data.
Top comments (0)