Preprint
Lost in Transcription: How Speech-to-Text Errors Derail Code Understanding
Jayant Havare et al., arXiv, 2026
Read the originalRead on October 17, 2026
- What it is
- A January 2026 arXiv preprint on spoken queries about source code. It measures how speech recognition errors affect code search and question answering in four Indic languages and English, and it names custom identifiers as a central difficulty.
- How much weight it carries
- Peer review status not confirmed (preprint). Its experiments target code understanding benchmarks in Indic languages and English, not coding dictation for developers, so its error rates do not transfer directly to dictating prompts.
Facts taken from this source
- The 2026 arXiv preprint Lost in Transcription names custom identifiers, such as variable and function names, as a central difficulty for speech input about code.
“custom identifiers such as variable and function names” (Abstract)
- In a 2026 study of spoken questions about code, having a language model refine the speech recognizer's output significantly improved performance in both the transcription and the code understanding stages.
“demonstrated that LLM-guided refinement significantly improves performance across both transcription and code understanding stages” (Abstract)
Pages that rest on this source
- Speech to Text for AI: Why Spoken Prompts Work BetterGuide
- Voice Input for Claude Code, Cursor and AI Coding ToolsGuide
- Do AI chatbots understand prompts with dictation errors?Question
- How do you dictate a good prompt for ChatGPT or Claude?Question
- Why does speech recognition get names and jargon wrong?Question