Voice dictation in the desktop app drops words throughout the recording — output is fragments, not a transcript (macOS, v2.1.237)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

Environment

• Claude Code desktop app, v2.1.237
• macOS (Darwin 25.6.0)
• Microphone: built-in MacBook Pro microphone, default input device, 48 kHz, input level ~57%
• Dictation language: reproduces in both en and de
• Reproduces in both hold and tap mode — switching to tap makes no observable difference

What's Wrong?

Voice dictation loses speech continuously throughout a recording — at the beginning, in the middle, and at the end. The result is not a degraded transcript, it is fragments: isolated recognized chunks with the connecting speech missing entirely. Every dictated prompt has to be rewritten by hand, which removes the entire point of the feature.

This is not a microphone or OS-level problem, and not a language-setting problem:

• Same machine, same microphone, same spoken sentence, back to back: ChatGPT's dictation transcribes it essentially verbatim. Claude Code returns fragments.
• Setting the dictation language to de vs en makes no difference — both produce fragments when speaking the matching language.
• Switching from hold mode to tap mode makes no difference, so this is not only the documented hold-mode warmup dropping the first words.

A plausible cause is the streaming transcription design itself. ChatGPT records the full utterance and transcribes it as one batch; Claude Code streams and finalizes segments live. Each segment boundary is a place where audio can be dropped, and the observed symptom — loss spread evenly across the whole recording rather than concentrated at one end — fits that.

What Should Happen?

A dictated passage should come back as a complete transcript, accurate enough to send without manual repair — comparable to what batch dictation in other assistants delivers on the same hardware.

Steps to Reproduce

1. Desktop app, voice dictation enabled (either mode)
2. Speak a continuous passage at normal pace
3. Compare the inserted transcript against what was actually said

Observed: recognized fragments separated by gaps where speech was simply not transcribed. I can supply a verbatim spoken-vs-transcribed sample pair on request if that helps triage.

Additional context

The recording limits compound this — the 15s silence auto-stop and the 2 minute maximum both cut in well before a real dictated prompt is finished. Tracked separately in #74534.

This overlaps a cluster of reports from spring 2026 that were auto-closed as duplicates or as inactive, never fixed, and are now locked: #43240 (umbrella), #40528 (start of speech dropped), #46154 (cuts out after 2-5 words mid-hold), #47592 (truncated after ~4 sentences). Filing fresh against 2.1.237 as the lock message instructs. The only comparable open issue, #80277, is scoped to the VS Code extension.