Remote Control phone voice channel: uncommon proper names transcribed inconsistently (e.g. "Jarvis" heard as "Josh," "Travis")
Summary
When speaking to a Remote Control session via the phone voice channel, an uncommon proper name used repeatedly in conversation (the assistant's configured custom name, "Jarvis") gets mis-transcribed inconsistently — different wrong words on different attempts, not one fixed substitution.
Steps to reproduce
- Start a Remote Control session on a phone, connected to a local Claude Code session where the assistant is configured with a custom name ("Jarvis").
- Speak a sentence containing that name, e.g. "Hey Jarvis, do you hear me, Jarvis?"
- Observe the transcription.
Observed
The name is heard as different common words on different attempts — "Josh" on one occasion, "Travis" on another, both within the same conversation context.
Expected
Consistent, accurate transcription of the name, especially given it's used repeatedly within the same session — repetition should reinforce recognition, not vary the result.
Supporting context
A comparable local voice pipeline (Whisper-based, via the open-source backtalk project) solves exactly this by biasing decoding toward the configured assistant name using Whisper's initial_prompt parameter — passing the name as a hint measurably fixes recognition of an otherwise uncommon word. This suggests the fix pattern is well understood; the Remote Control voice pipeline doesn't appear to expose an equivalent per-session vocabulary/name hint.
Impact
Minor but recurring friction — every voice interaction risks misaddressing the assistant, which undercuts the "just talk to it naturally" experience Remote Control's voice mode is meant to deliver.