- Who it helps
- Researchers and developers evaluating voice assistants and AI note-taking tools.
- Availability
- October 30 paper is public and links an artifact site. Code and demonstrations were not independently inspected.
What was evaluated
The authors tested 13 open-source speech-recognition models from four families, including digital and over-the-air conditions. Their attacks target transcript content or degrade transcription; results can propagate into downstream tools.
GhostPrefix tests reusable audio attacks on speech transcription ↗
How to read the claim
The paper reports strong results in its tested conditions. Those results do not establish that every voice product is affected, and a transcription failure is distinct from a downstream application executing an action.
GhostPrefix tests reusable audio attacks on speech transcription ↗
What you can do
- Read the tested model families and conditions before applying the result to a particular voice system. Open reference ↗
Where the evidence stops
Abstract checked November 4. No audio attack generated, demonstration watched or experiment reproduced. Real-product exposure remains unverified.
Sources read. Primary source read November 4, 2026 UTC. No software installation, model inference or complete workflow executed.
Sources & further reading
Paper · paper
GhostPrefix tests reusable audio attacks on speech transcription ↗
The reported cross-model audio attack affects how voice-pipeline evaluations should be designed. The abstract supports the tested scope, with real-product exposure and replication unresolved.
Shengding Liu, Ashwin Innuganti and Qiben Yan · page checked 2026-11-04
Coverage note: Independent attention and end-to-end execution remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
GhostPrefix studies whether a short audio segment placed before speech can alter what a speech-recognition system transcribes.