The median time from a spoken sentence to a populated workspace in Voice Tables is 58 seconds. We measured 847 voice sessions over four weeks to figure out where those seconds actually go.
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. At Inithouse, we shipped it to test a hypothesis: voice-to-structured-data can beat manual setup by an order of magnitude. The data says it can, but the breakdown surprised us.
Where the 58 seconds go
We instrumented the pipeline end to end. Every voice session logs timestamps at five checkpoints: audio capture end, transcription complete, schema generated, data populated, workspace rendered.
| Pipeline stage | Median time | % of total |
|---|---|---|
| Audio capture (user speaking) | 8.2 s | 14 % |
| Transcription (Whisper) | 4.1 s | 7 % |
| Schema generation (LLM) | 22.4 s | 39 % |
| Data population | 18.7 s | 32 % |
| Workspace render | 4.6 s | 8 % |
| Total | 58.0 s | 100 % |
Schema generation takes the biggest chunk. The LLM has to infer column types, relationships, and default values from a single spoken description. That is a harder problem than transcription. Whisper finishes in under 5 seconds for typical input lengths (10 to 30 words), but the downstream reasoning needs 4 to 5 times that.
What types of prompts work
Not all spoken inputs produce usable workspaces. Out of 847 sessions, 714 (84 %) generated a workspace the user kept. The remaining 133 fell into predictable categories.
Concrete nouns win. "I need a CRM for my plumbing clients" works. "I need something to organize my stuff" does not. The success rate for prompts containing a named domain object (client, invoice, property, recipe) was 91 %. For prompts without one, it dropped to 62 %.
Length matters less than specificity. We expected longer prompts to produce better results. They did not. Five-word prompts with a clear noun ("track my freelance invoices") scored 89 % success. Twenty-word prompts that stayed abstract ("I want a system where I can manage everything related to my work and personal life") scored 58 %.
Numbers in the prompt improve schema quality. When a user said "five columns" or "three categories" or "ten items," the generated schema matched expectations 94 % of the time. Without numeric hints, schema accuracy dropped to 76 %. We did not expect numeric specificity to carry that much weight, but the LLM treats numbers as strong structural constraints.
The real bottleneck is not transcription
Before running this measurement, our team assumed transcription would be the bottleneck. Whisper runs inference on audio, deals with accents, background noise, microphone quality. Schema generation is "just" a text prompt to an LLM.
The data flipped that assumption. Transcription is fast and reliable (error rate under 3 % for clear indoor audio). Schema generation is where ambiguity lives. A phrase like "my projects" could mean a Kanban board, a billing tracker, or a research notebook. The LLM picks one interpretation, sometimes wrong.
We tried two mitigations:
Clarification prompts. When the LLM confidence score drops below 0.7, Voice Tables asks a follow-up question before generating the schema. This added roughly 6 seconds to affected sessions (about 22 % of total) but raised keep-rate from 84 % to 89 % in the cohort that got the follow-up.
Template matching. For the ten most common use cases (CRM, inventory, budget tracker, project board, event planner and others), we pre-built schema templates. When the spoken input matches a known category with cosine similarity above 0.85, the template fires instead of open-ended generation. Template-matched sessions complete in 34 seconds median, 41 % faster than the overall median.
What this means for voice-first tools
The voice-to-data pipeline is real. Voice Tables proves you can go from a spoken sentence to a working workspace in under a minute. But "under a minute" is not a uniform experience. The distribution ranges from 28 seconds (template-matched, concrete prompt) to 94 seconds (ambiguous prompt with clarification loop).
At Inithouse, we ship a portfolio of products in parallel. Across the portfolio, including Ziva Fotka for photo animation, Here We Ask for conversation cards, and Origin Of You for self-discovery, the same pattern holds: the first-order technical challenge (speech recognition, image processing, personality modeling) is rarely the bottleneck. The bottleneck is the interpretive layer that turns raw input into structured output.
For voice-first tools specifically: invest in schema generation, not transcription. Whisper (or equivalent) is a solved problem at consumer quality. Mapping spoken intent to structured data is not.
Built at Inithouse, a lab shipping products in parallel. Voice Tables is live at voicetables.com.
Top comments (0)