Skip to content

Releases

TalkTalk 1.0.21 (build 88)

Choose a tag to compare

@aaajiao aaajiao released this 14 Nov 15:18

TalkTalk 1.0.21 fixes unnecessary language-resource warnings when German or French are not selected.

Changes

  • Chinese and English are the required Apple speech resources. German and French are optional and start unselected; your choices are saved.
  • The existing language-resource step downloads only selected missing resources and checks readiness against that selection. Unselected German and French resources no longer trigger startup, General settings or menu warnings.
  • Apple translation-resource checks also follow the selected languages. Deselecting a language clears its optional-resource warnings and keeps installed resources intact.
  • Change your selection through Settings > General > Onboarding, in the existing language-resource step. The existing page structure and controls are retained, with checkboxes beneath the language badges.
  • Resource preparation choices do not restrict Qwen's built-in language support or the scene language menus. If you actually use an Apple recognition direction whose resources are missing, TalkTalk still offers the existing download prompt for that direction.

Verification

Resource selection, persistence, startup checks, onboarding state and menu behavior passed 77 focused regression checks. The complete functional checks passed: 1,760 Swift tests, 130 Python engine tests (one environment-dependent skip), and 80 isolated installer checks. The signed release build and bilingual smoke tests also passed.

The new checkboxes and actual macOS resource downloads have not been manually exercised on screen. This release validation did not download or delete language resources, change user preferences, record audio or make cloud translation requests.

Requires Apple Silicon and macOS 26.4 or later. The app remains ad-hoc signed and is not notarized. Users on 1.0.18 or later can update from the app; earlier versions require a manual download.

TalkTalk 1.0.20 (build 86)

Choose a tag to compare

@aaajiao aaajiao released this 14 Nov 15:00

TalkTalk 1.0.20 adds French listening, subtitles, translated speech and French as your spoken language.

Changes

  • Listening subtitles support all 12 translation directions between Chinese, English, German and French in Meeting, Phone and Video modes.
  • In Meeting and Phone modes, speak Chinese, English or French and translate into any of the other three languages: nine speaking directions.
  • French works with local Qwen recognition, Qwen and Ming synthesis, Apple translation, Apple recognition fallback and macOS system voices. Compatible installed models and saved voice profiles are reused.
  • French numbers, amounts, dates, percentages, units, punctuation and common spoken number forms are handled throughout translation checks and speech preparation. Existing language settings and glossary files remain compatible.
  • French joins the existing language menus, previews, glossary fields and voice previews. The interface remains English and Chinese, with the same page structure, layout components and interaction flows.

Verification

The complete functional checks passed: 1,755 Swift tests, 130 Python engine tests (one environment-dependent skip), and 80 isolated installer checks. The signed release build and bilingual smoke tests also passed. Local synthetic French speech was recognized exactly by Qwen3-ASR 0.6B. Qwen 1.7B and Ming 0.5B generated French samples that also passed recognition readback. Apple high-fidelity translation preserved the budget amount in all six directions between French and Chinese, English or German.

Apple French recognition requires its speech resources, available through the existing speech-resource setup step. Qwen French recognition does not depend on that download. French voice likeness, long speech, remote audio in real meetings and phone calls, and real-time latency still need practical listening checks; short synthetic samples do not establish those results.

Requires Apple Silicon and macOS 26.4 or later. The app remains ad-hoc signed and is not notarized. Users on 1.0.18 or later can update from the app; earlier versions require a manual download.

TalkTalk 1.0.19 (build 84)

Choose a tag to compare

@aaajiao aaajiao released this 14 Nov 13:58

TalkTalk 1.0.19 fixes video-only setup so Qwen recognition is prepared before you start using subtitles.

Changes

  • Video-only onboarding installs the private Python runtime, shared dependencies and Qwen3-ASR, without requiring speech synthesis models, TalkTalk Mic, a voice recording or microphone permission.
  • Existing models and Apple/Qwen recognition choices are kept. Adding speech synthesis later preserves the selected recognition model.
  • Cancelled or failed installations can be retried. Repairing an interrupted recognition-only setup does not add synthesis models.
  • Setup verifies actual Qwen recognition. The sample follows the selected recognition engine and clearly shows Apple fallback when needed.
  • The existing interface components and layout are retained.

Verification

The complete functional checks passed: 1,738 Swift tests, 126 Python engine tests (one environment-dependent skip), and 80 isolated installer checks. Real local Qwen3-ASR 0.6B checks passed using only the recognition model, with no Apple fallback or speech synthesis.

Fresh installation paths are covered by isolated fixtures. Permissions, downloads and live video capture have not been re-tested together on another clean Mac.

Requires Apple Silicon and macOS 26.4 or later. The app remains ad-hoc signed and is not notarized. Users on 1.0.18 can update from the app; earlier versions require a manual download.

TalkTalk 1.0.18 (build 82)

Choose a tag to compare

@aaajiao aaajiao released this 14 Nov 13:30

TalkTalk 1.0.18 adds in-app updates from GitHub Releases.

Changes

  • Choose Check for Updates… in the menu bar menu or Settings › General › App Updates. Sparkle shows the release notes and lets you download and restart when you choose.
  • Updates are manual: no automatic checks, background downloads or silent installation. Updates wait for active interpretation, recording, preview, diagnostics and maintenance. If work starts while an update downloads, finish that work and explicitly choose Restart and Update.
  • Both the update archive and feed carry Ed25519 signatures. The archive is checked before extraction; unsigned or modified feeds are rejected.
  • A standard publishing command prepares the signed assets, verifies the source/build relationship, uploads to a draft, compares GitHub file hashes and sizes, then publishes the complete release.

First installation

Version 1.0.17 and earlier do not have an updater. Download this ZIP once and move TalkTalk.app to Applications; later releases can be installed from within the app. Existing models, private runtime, voice profiles and settings are reused.

Requires an Apple Silicon Mac running macOS 26.4 or later. The app remains ad-hoc signed, without Developer ID signing or Apple notarization.

Verification

The full scripts/check.sh --skip-stress run passed, including 1,729 Swift tests, Python engine tests, and isolated packaging/publication/provenance fixtures. Release packaging checks framework signing, both interface languages and the build-source receipt.

An isolated application using the bundled Sparkle version completed a real signed download, deferred installation, replacement and relaunch after an explicit resume. Modified archive and feed tests were rejected while leaving the old application intact. This was an isolated updater test, not a real meeting test.

TalkTalk 1.0.17 (build 79)

Choose a tag to compare

@aaajiao aaajiao released this 14 Nov 12:57

TalkTalk 1.0.17 (build 79) explains why meeting speech is unavailable when the meeting app has not connected to TalkTalk Mic. Waiting for that connection no longer points you toward unrelated Input Monitoring permissions or headphones.

Changes

  • Distinguishes a meeting using another microphone, a meeting not reading TalkTalk Mic, an unknown input device, and a connection awaiting confirmation.
  • Shows actionable meeting-input guidance even while physical microphone capture is intentionally closed. Actual key-listener and physical-microphone failures retain their own messages.
  • Keeps Phone-specific instructions in Phone sessions.
  • Adds connection-state events to the local session log only when the state changes. They distinguish Feed readiness, meeting-reader confirmation, physical capture and system-input management without recording device names, device UIDs or meeting content.

The existing audio safety gates, reader-confirmation interval, recovery polling, driver visibility rules and speech generation settings are unchanged.

Validation and limits

scripts/check.sh --skip-stress passes: 1,723 Swift tests without failures, plus the Python engine and installer checks. The release build, English/Chinese smoke checks, strict signature verification and ZIP extraction/binary comparison also pass.

The input-guidance changes have automated regression coverage, including confirmation recovery, Phone preparation and onboarding. This release has not been validated in a new real meeting and does not claim to eliminate Google Meet's need to reselect TalkTalk Mic. Other supported meeting apps use the same detection and guidance path; cross-app call testing remains pending.

Install

Requires an Apple Silicon Mac running macOS 26.4 or later. Download the ZIP, unzip it and move TalkTalk.app to Applications. Existing runtimes, models and voice profiles are reused.

The app is ad-hoc signed, without a Developer ID signature or Apple notarization. SHA256SUMS.txt is attached.

TalkTalk 1.0.16 (build 78)

Choose a tag to compare

@aaajiao aaajiao released this 11 Nov 00:22

TalkTalk 1.0.16 (build 78) is a round of interface and flow fixes on top of 1.0.16 (build 77). Recognition, translation and speech are unchanged. The app version is still 1.0.16; Settings › General › About shows 1.0.16 (78).

What's new

Menu bar and sessions

  • The icon follows the session. It shows an hourglass while a session is preparing or ending, and an orange triangle when recognition is on the Apple fallback, when recognition failed, or when the last session ended with an error.
  • "End" only ends a running session. A menu left open after a session ended by itself no longer starts a new one, and the menu rebuilds when the session changes state.
  • Fewer false warnings. "Components not ready" no longer appears during engine maintenance or on a video-only setup. "Apple resources need setup" is checked again when you return to the app or finish a download.
  • New menu lines. An idle menu says when the last session ended automatically. During video subtitles, a line tells you the subtitles are hidden and brings them back.
  • Quick start explains itself. When a quick start needs you to choose a source, the scene page opens and the source chip flashes once.

Speaking

  • A failed speech chain stays visible. In a meeting, the problem marker stays in the title bar until the meeting ends or a restart works, instead of a 4-second notice.
  • Clearer playback warning. When the playback device is unknown, the warning says so instead of claiming speakers are in use. Echo protection is unchanged.
  • Permission buttons open the right pane. Key-listening problems open Input Monitoring; an unconfirmed system-audio grant opens Screen & System Audio Recording.
  • Phone sessions name Phone, not "the meeting app", in their warnings.

Diagnostic recording and privacy

  • Consent comes first. In the "record the other side" dialog, Start Recording stays disabled until you tick the box confirming everyone agreed.
  • A video subtitle session no longer uses up a "record my side" choice made for the next session.
  • A track that reaches its 512 MiB cap stops and says so, instead of still showing as recording.
  • If a recording cannot start, its switch turns off.
  • Full uninstall is disabled while a session is running.

Engine and settings

  • Your chosen bundle is kept after a cancelled or failed engine install, so you can retry it.
  • Repair reuses the bundle of the unfinished install, and asks you to choose one when the install record cannot be read.
  • The glossary shows Unsaved while it has changes you have not saved.
  • Flow diagrams size their nodes to the real column width, and source chips, language codes and the synthesis node look the same on every page.
  • When the app you chose is open but silent, the page and the start alert say it is silent instead of telling you to open it.

Subtitle window, onboarding and wording

  • Long English scene titles truncate in a narrow title bar instead of overlapping the controls.
  • Onboarding step titles that wrap are centred. A setup you ran again comes back after the relaunch for Input Monitoring.
  • With onboarding postponed, the scene page's Start still works; the menu's quick start waits until setup is finished.
  • English wording is consistent across pages, and several labels were shortened to fit.

Install

Requires an Apple Silicon Mac running macOS 26.4 or later. Download the ZIP, unzip it and move TalkTalk.app to Applications. Existing runtimes, models and voice profiles are reused; this release does not change them.

The app is ad-hoc signed, without a Developer ID signature or Apple notarization, so macOS may warn you or block the first launch. SHA256SUMS.txt is attached.

Verification

Automated checks pass (scripts/check.sh --skip-stress, 1,701 Swift tests plus the Python engine suites), along with the release build, strict code-signature verification and the bundled smoke test. The build was installed and opened once without problems. The individual fixes have passed automated tests only: none was accepted item by item in the running app, in a live meeting, on a phone call or with real cloud requests.

The binary was built from commit a6ecb51. The tagged commit differs from it only in source comments and documentation.