Offline development update · 14 October 2026

The latest development app generates Gemma 4 E2B minutes on the device after an explicit model import, and encodes MP3 locally in project and batch exports. These operations do not upload audio or transcripts or switch to a cloud service. Offline speech also supports vocabulary hints. The latest source now includes local anonymous speaker suggestions for transcribed recordings up to three minutes, using a bundled, verified model. Suggested labels require review and can be undone; mixed or overlapping passages remain unassigned. No reusable voice profiles or real-person identification are provided. The imported Gemma model can be removed without deleting projects and is excluded from device backup; bundled speaker/separation models are removed with the app; opening an external model download link is a separate network action.

The latest development app also separates voice/vocals and background on the device using a bundled verified Spleeter model, without uploading a recording or using processing credits. It previews, balances and exports real stems as AAC/M4A, MP3, WAV, FLAC and AIFF; approved redactions are applied before processing and to both saved stems. The private original is preserved. Tests cover selected recordings and simulators, not guaranteed quality: separation may remove wanted speech, leave leakage or introduce artifacts, and vocal bandwidth is limited. It cannot reliably isolate each overlapping speaker. Physical-device and broader listening checks remain open.

The latest development source asks questions across user-selected transcripts locally with Gemma and displays exact cited quotations for review. Private saved question checkpoints contain selected quotations and project references, are excluded from backup and bounded to32MiB, and can be cleared in Storage. Deleting a project discards these disposable checkpoints. Removing the model preserves saved checkpoints, so complete unchanged answers can still reopen without inference. Offline search/questions are included in Local Tools or Pro after model import. Phone/iPad saved-answer replay and original source playback pass real simulator UI checks; Phone missing-answer replay and real-request cancellation/relaunch checks also pass with no invented citation. Fresh-inference UI, interruption at every stage, natural-recording accuracy and physical performance remain release gates. current cross-meeting continuity still has a legacy explicitly approved service path. Stem exports now support verified AAC/M4A, MP3, WAV, FLAC and AIFF settings; converting a lossy recording to a lossless file does not restore lost quality.

Prepared exports in the latest development source recheck the complete original file and reviewed edits before sharing, using incremental SHA256 reads. Simulator checks verify an approved privacy AAC/SRT copy, complete audio decoding, preserved original bytes, and actual local Save to Files on iPhone and iPad with matching destination bytes. The export-ready screen also offers a direct full-screen Files picker for accessible iPad use. This does not guarantee removal of every sensitive paraphrase or compatibility with every destination; review share copies before sending them.

Simulator and real-file checks are documented; they do not establish physical-device performance. The uploaded Build 21 predates these changes and still has the legacy approved service paths described below. Full offline conversion, AI accuracy, summary length and long-recording support, and physical-device verification remain incomplete. SoundBrief is not yet released to clients.

Chinese script copies · offline development tool

Project tools → Translate transcript → Chinese script creates a separate Simplified or Traditional copy on iOS 17 or later, using native character conversion. No account, AI model, language download or recording upload is needed. This changes characters; it does not translate Cantonese into Mandarin or correct a transcription. Review names and ambiguous characters. The private original, speaker labels and source passage times are preserved.

Saved copies stay in the private, backup-excluded translation history. Delete a chosen copy or its project to remove it. Shared SRT, WebVTT, TXT and JSON copies exclude approved protected passages and recheck the original and current edits before export. Files you deliberately save or share follow the destination’s retention settings. Phone and iPad simulator checks verify conversion, relaunch recovery, actual Files saves and unchanged original bytes; physical-device checks remain open. Apple language translation is a separate iOS 18 or later operation with device-specific availability. The uploaded Build 21 predates this feature; the app is not released.

Speech activity · offline development tool

Project tools → Find speech locates suggested speech regions using a small model included in the app. No account, model download or recording upload is needed. Selecting a region selects the corresponding retained audio; it does not automatically cut sound or infer words or speaker identities. Quiet speech, overlap and singing can be misclassified, so listen before editing.

Saved activity timings stay in private app storage, excluded from backup, and can be cleared in the tool or removed with the project. Temporary analysis audio is deleted after completion or cancellation; interrupted owned working files are recovered on the next analysis. Originals are preserved. Phone and iPad simulator checks cover real inference, replay and selection; physical performance, manually labelled boundary accuracy and long-file stress remain release checks.