Codex Desktop Windows dictation frequently fails with "Unable to transcribe audio" despite working mic

Summary

Codex Desktop on Windows frequently shows the toast Unable to transcribe audio after dictation. This appears to be a recurring recorder/session failure rather than a microphone, Windows permission, or connectivity problem.

The user reports this happens many times per day and relies heavily on dictation for long prompts. Prior incidents include both:

• a failed dictation that left recoverable local WebM/Opus blobs, and
• a failed dictation that left no new blob/transcript at all.

The current latest failure is the no-blob variant.

Environment

• Product: Codex Desktop on Windows
• Package: OpenAI.Codex_26.527.3686.0_x64__2p2nqsd0c76g0
• Codex CLI runtime after update: 0.137.0
• OS from codex doctor: Windows 10.0.26100 / Windows 11 Core, 64-bit
• Auth/connectivity: codex doctor reports auth configured, websocket connected, state/log DB integrity OK
• Doctor summary after update: 17 ok | 1 idle | 2 notes | 0 warn | 0 fail

Latest Observed Failure

Date/time: 2026-06-04, approximately 01:16-01:26 America/New_York

Observed UI symptom:

• Codex Desktop displayed a toast saying Unable to transcribe audio.

Local checks immediately after the failure:

• transcription-history.jsonl did not receive a new successful transcript after the failure.
• The Codex blob_storage directory existed, but recursive file count was 0; no new recoverable audio blob was written for this failed attempt.
• Windows microphone permission for Codex was Allow under the relevant ConsentStore registry key.
• FFmpeg DirectShow listed the active microphones, including Microphone (JLab Jbuds Cam Microphone) and Microphone Array (Realtek(R) Audio).
• A direct FFmpeg test recorded successfully from Microphone (JLab Jbuds Cam Microphone) into a valid WAV:
• duration: about 1.97 seconds
• sample rate: 16000 Hz
• size: 63118 bytes

Prior Related Evidence

1. Previous recoverable failed dictation

• A prior failed retry dictation event left recoverable local WebM/Opus blobs.
• The reconstructed audio was about 1755.356625 seconds, proving Codex had recorded audio locally but failed later in the save/transcribe flow.
• Full-file transcription exceeded the model duration limit and had to be chunked.

1. Previous no-blob failed dictation

• Windows mic permission was allowed.
• FFmpeg could read the same JLab microphone successfully.
• No new blob or transcript appeared for the failed attempt.
• That pointed to a Codex Desktop recorder/session abort before a usable blob/transcript was created.

The latest failure matches the second pattern.

Expected Behavior

When dictation fails to transcribe, Codex Desktop should either:

1. save the transcript into the input box, or
2. preserve a local recoverable audio artifact and expose a retry/recover option.

Actual Behavior

Codex Desktop shows Unable to transcribe audio and does not reliably save a transcript or recoverable audio blob.

Working Theory

The failure appears to be inside Codex Desktop's recorder/session lifecycle. The microphone and OS permission are valid, and OpenAI connectivity is healthy. Sometimes audio is recorded locally and can be recovered from blobs; other times the recorder appears to abort before writing any audio artifact.

Local Mitigation Tried

• Verified Windows mic permission is allowed.
• Verified the microphone records outside Codex.
• Verified Codex state/log DBs are healthy via codex doctor.
• Updated global Codex CLI runtime from 0.135.0 to 0.137.0.
• Regenerated redacted codex doctor --json locally after the update.

Request

Please investigate Codex Desktop dictation failures where the app aborts before writing a local blob/transcript, despite valid Windows mic permission and successful external microphone capture.

A recoverable local audio fallback would prevent loss of long dictation when transcription fails.