Repository navigation
fix(mobile): dictated speech is no longer lost on iOS - #60
Merged
Merged
Conversation
… iOS (pingdotgg#30139) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: maria-rcks <maria@kuuro.net> (cherry picked from commit bd30817)
… holds Co-Authored-By: Claude Code <noreply@anthropic.com>
…a headset The filter judged every stretch of speech against the voice profile alone. The profile is recorded once, somewhere quiet, so noise or another microphone pushed the user's own speech past the limit and it was cut as someone else's. A short opening stretch has a looser limit, so in the worst case only the first words of a dictation survived. A stretch the profile rejects is now also compared with the stretches it accepted in the same recording, and kept when it matches them. Measured on 8 Mini LibriSpeech speakers, half-minute recordings, own words lost with the filter on: noise at 5 dB 9.2% before, 1.8% after; narrowband 3.4% to 0%; echo with noise 2.7% to 0%. Strangers' words kept are unchanged (11.5%, 1.0%, 6.4%), and the clean mixed-voice numbers are unchanged (2.4% own words lost, 4.8% strangers' words kept, 19 speakers). Also adds a test that half a minute of speech in the phone's recording format (AAC, 44.1 kHz, stereo) is transcribed in full by all 3 models. Co-Authored-By: Claude Code <noreply@anthropic.com>
In a long dictation the filter cut stretches of the user's own speech whenever their voiceprint drifted from the profile. A voiceprint alone now never cuts anything. A stretch goes only when it also is more than 6 dB quieter than the user (further from the phone) and the user was heard for at least 1.25 times as long as everything cut. A voice as loud as the user's stays. Measured on 19 LibriSpeech readers: the speaker's words lost when speaking alone are 0 in noise, narrowband and echo (was up to 41.5% for one reader), and a stranger 12 dB under the speaker keeps 4.9% of their words. Co-Authored-By: Claude Code <noreply@anthropic.com>
The microphone waited for Apple to prepare its model before it recorded, so words spoken in that gap were never captured. Preparation now runs while the user talks and is awaited when the recording is transcribed. A language Apple cannot take is reported then; a downloaded model that was still loading takes the recording instead. Co-Authored-By: Claude Code <noreply@anthropic.com>
A call, Siri, a recorder error or leaving the app threw the whole recording away. expo-audio also pauses the recorder for an audio interruption without an event and never resumes it, so the app kept showing a recording that captured nothing more. An interrupted recording is now transcribed and its words are added, with a notice that it stopped early. The metering loop notices a recorder that stopped capturing and finishes the same way. Co-Authored-By: Claude Code <noreply@anthropic.com>
…overed Every recording was deleted the moment its dictation ended, so a failed, empty or wrong transcript could not be recovered, and a draft that changed during the dictation dropped the whole transcript. - The recorder's file is moved to a kept folder as soon as recording stops, before transcription. The last 5 stay, canceled ones included. - A transcription that throws offers "Transcribe again", which runs the kept file and adds it to the draft. - Settings > Microphone > Model lists the recent recordings. Each runs through the model in use without the voice filter, with a copy button. - A transcript for an edited draft goes at its end. It is dropped only when the draft is gone. - The 5-minute limit says so, and a notice after a commit offers no retry at an old caret. Co-Authored-By: Claude Code <noreply@anthropic.com>
github-actions
Bot
added
vouch:trusted
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
On TestFlight build 49, dictated speech was lost on iOS. A 30 second dictation came out as 3 words, text typed between dictations did not reach the draft, and once a dictation ended its recording was deleted, so nothing could be recovered.
Fix
Caches/dictation-recordingswhen the recorder stops, before transcription. The last 5 stay, canceled ones included.Fork only. Android keeps recordings and "Transcribe again" through the shared controller; it has no Recent recordings list.
Verification
vp run check:changedexited 0 on the head commit.Not verified: real speech through a microphone, a physical iPhone, an interruption or the 5 minute limit on a device, AirPods route changes, Android. Screenshots from the simulator pass are kept locally and are not attached.
Known limits
Model: Claude Fable 5. Harness: Claude Code in PseudoCode.
🤖 Generated with Claude Code