This page documents two sources: a case study published by Google DeepMind and our own Voice Agent Benchmark Landscape. Store ratings are recorded separately on the figures page.
Google DeepMind: case study on Cue
Address: deepmind.google/models/gemma/gemmaverse/cue-ai. Published in Google DeepMind's Gemmaverse; figures current as of May 2026.
What it says:
- Cue is “a voice-activated AI agent that lives on the user's desktop.”
- Cue's dictation polish step moved from a cloud model to Gemma 4 E4B running locally through Ollama; the median latency of that step fell from 876 ms to 488 ms, measured on Apple Silicon over a 227-sample real-voice benchmark covering English and mixed-language input.
- Per-user dictation rose by about 30 percent across active beta users in the four weeks before and after the switch.
- The marginal inference cost of the polish step dropped to zero, and dictation is unlimited on every tier, including the free one.
- Speech is transcribed by a cloud speech-to-text service before the local polish step; if Ollama is not running, the app routes the polish step to a cloud model.
- Cue's agent mode relies on a cloud model.
What it does not say: it is not a product review, it makes no claim about transcription accuracy, and the 227-sample benchmark it describes is private to Cue. What the page says, set against how we quote it, is in What the DeepMind case study says about Cue.
Voice Agent Benchmark Landscape
Ours. A structured map of 13 public benchmarks for voice agents, spoken assistants, speech-enabled tool use, computer action, speech recognition and meeting understanding, released under the MIT licence with a DOI. It contains no scores and does not include Cue's private benchmark. The record, versions and citation format are on its own page.
Other products
This page does not document an external evaluation of Cue Office or PPT AI. Store ratings for PDF AI, PDF Scanner Max and Evolve are listed on the figures page with their counts.