On the first on-device iPhone dictation in a language, T3 waited for Apple’s speech model before starting the microphone. Speaking immediately after tapping the mic could turn “one, two, three” into “two, three.”
The iPhone records locally as soon as microphone permission and audio setup allow. Stopping waits for the model and transcribes the saved clip. Cancelling deletes the clip and releases the session even if Apple never answers. A subsequent recording can finish without waiting for the cancelled request.
This is the opening-word fix from @t3dotgg’s #25808, rebuilt at the maintainer’s request. Only on-device iPhone dictation opts into early capture. Android, web and environment dictation keep their existing preparation order and consent checks. The original report, #17888, was closed with the label needs more info.
On 1ea9d03, all 63 focused iOS, Android and shared-controller tests passed. The iOS tests hold Apple’s preparation pending: capture starts, Stop waits, and Cancel deletes the file. A second recording reaches the draft before the first native request resolves. check:changed passed, including typechecking seven packages.
@jokaperes verified that the source branch kept the first word on an iPhone 17 in Brazilian Portuguese. The rebuilt head hasn’t been tested on a physical iPhone. To check it, select This device, use a language whose model isn’t installed, tap the composer mic and immediately say “one, two, three.” All three words should appear after Stop. On another cold start, cancel and dictate again; only the second clip should reach the draft.
GPT 6 Astra · Codex in T3 Code
Added 1 commit
Verdict: Approved at 1ea9d03
Macroscope's review found this PR approvable — Local iPhone dictation can begin capturing audio while its language model prepares, preserving opening words. Existing permission and cancellation safeguards remain, with no change to product defaults or transcription destinations.
You can add or adjust custom eligibility rules. Learn more.
The changes allow live transcription to return no text and update recording completion to fall back to file transcription. iOS model preparation now runs asynchronously and exposes the supported locale. Tests cover preparation failures, cancellation, and recording during preparation.
Changes: Voice transcription
| Layer / File(s) | Summary |
|---|---|
Transcription contract and recording completion packages/client-runtime/src/voice-input/transcription.ts, packages/client-runtime/src/voice-input/controller.ts, packages/client-runtime/src/voice-input/controller.test.ts, apps/mobile/src/native/voiceTranscription.android.test.ts | PreparedVoiceTranscription.start can return null. Recording completion falls back to file transcription when the live result is absent or has no text. Tests check unavailable preparation and Android recorder preparation call counts. |
iOS model preparation and dictation apps/mobile/src/native/voiceTranscription.ios.ts, apps/mobile/src/native/voiceTranscription.ios.test.ts | iOS preparation resolves the supported locale asynchronously. Tests cover recording during preparation, cancellation and failure before or after stopping, and a subsequent successful dictation. |
Estimated code review effort: 3 (Moderate) | ~20 minutes
Change: Bug fix
| 1 | sequenceDiagram |
| 2 | participant VoiceInputController |
| 3 | participant Recorder |
| 4 | participant prepareVoiceTranscription |
| 5 | participant AppleSpeechModel |
| 6 | VoiceInputController->>Recorder: Start recording |
| 7 | VoiceInputController->>prepareVoiceTranscription: Start transcription preparation |
| 8 | prepareVoiceTranscription->>AppleSpeechModel: Prepare speech model and locale |
| 9 | AppleSpeechModel-->>prepareVoiceTranscription: Return supported locale |
| 10 | VoiceInputController->>Recorder: Stop recording |
| 11 | VoiceInputController->>prepareVoiceTranscription: Transcribe recording |
| 12 | prepareVoiceTranscription->>AppleSpeechModel: Transcribe with supported locale |
| 13 | AppleSpeechModel-->>VoiceInputController: Return recognized text |
Possibly related PRs
Suggested labels: bugfix, size:M
Suggested reviewers: pixpmusic
1ea9dLocal iPhone dictation now records while Apple prepares the speech model, and cancellation releases the session. No actionable merge-blocking risk was found. The only open note is a style preference about a test import path.
| Check name | Status | Explanation |
|---|---|---|
| Title check | Passed | The title clearly describes the iPhone dictation fix. |
| Linked Issues check | Passed | Check skipped because no linked issues were found for this pull request. |
| Out of Scope Changes check | Passed | Check skipped because no linked issues were found for this pull request. |
| Description check | Passed | The description is mostly complete. It explains the problem, change, scope, and verification. The scope section should link directly to the maintainer’s approval comment. |
apps/mobile/src/native/voiceTranscription.ios.test.ts: 1
5-5: External image Trivial · Maintainability & Code Quality · Quick winExternal image Unverified finding Verification ran but could not confirm this finding. It is shown for review, not as a verified issue.
Import the test reset helper through a package export or a shared path.
Line 5 imports
resetVoiceInputGlobalsForTestswith a relative path that crosses package boundaries intopackages/client-runtime/src.@t3tools/client-runtime/voice-inputis already imported on line 4, and the package exports it. Export the helper fromvoice-input/index.tsand import it from the package. This avoids a fragile path and keeps the package boundary intact.♻️ Proposed change
1 -import { VoiceInputController, VoiceTranscriptionError } from "@t3tools/client-runtime/voice-input"; 2 -import { resetVoiceInputGlobalsForTests } from "../../../../packages/client-runtime/src/voice-input/controller"; 3 +import { 4 + VoiceInputController, 5 + VoiceTranscriptionError, 6 + resetVoiceInputGlobalsForTests, 7 +} from "@t3tools/client-runtime/voice-input";
64bc5ffa-9eb3-4234-a9fa-870ca5d56f55Reviewing files that changed from the base of the PR and between e2c1aba9de4a5a3c02ca774506340160ec828da9 and 1ea9d0312e61518ec9ba36ecc740d6f330c3166c.
apps/mobile/src/native/voiceTranscription.android.test.tsapps/mobile/src/native/voiceTranscription.ios.test.tsapps/mobile/src/native/voiceTranscription.ios.tspackages/client-runtime/src/voice-input/controller.test.tspackages/client-runtime/src/voice-input/controller.tspackages/client-runtime/src/voice-input/transcription.tsIncluded review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 6 remain after this review.
I kept the reset helper out of the public voice-input API. It clears the active recording and transcription guards without stopping the underlying work; it is for isolating tests only. The Android adapter and mobile voice-session tests already import it directly from the controller. I also updated the description to distinguish the original issue from the maintainer-approved scope and to identify what was tested on this head.