Skip to content

Releases

v0.4.0

Choose a tag to compare

@github-actions github-actions released this 15 Nov 23:44
80b8b59

Download

  • Windows 10/11: tinglongwu-setup-0.4.0.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.4.0/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.4.0/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪
Tiếng Việt (vi) ⚪ ⚪ ⚪ ⚪ ⚪
ไทย (th) ⚪ ⚪ ⚪ ⚪ ⚪
Bahasa Indonesia (id) ⚪ ⚪ ⚪ ⚪ ⚪
Español (es) ⚪ ⚪ ⚪ ⚪ ⚪
Deutsch (de) ⚪ ⚪ ⚪ ⚪ ⚪
Français (fr) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • Listen to sound in the room, not only what the computer plays: pick a microphone (Advanced > Microphone (room sound)) to get live subtitles for a lecture or an in-person meeting.
  • The settings window keeps its taller size in the one-column layout, so no setting is cut off when you make it narrow.
  • In the Windows menu, pressing Enter at the microphone question now just keeps the default instead of printing a usage line.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 119 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 119 MB
  • SHA256: a27b0b07753ab712cb25490b0883c69a4640f01f65ea171e0e558edf4347b14e (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: hardware checks deferred. The maintainer's self-hosted hw-tests runner did not start within 30 minutes, so this release shipped on ci, e2e and windows-build; hw-tests runs later on this commit and a failure opens a P0 issue.

Optional packs (downloaded on demand)

pack size sha256
ja 1133.1 MB dc5f44fcf1fa
ko 1133.1 MB dc5f44fcf1fa
ru 1196.3 MB 8b7bfd559dbc
speak 120.6 MB ef9f72fc5e31

Changes since v0.3.8

  • fix #1104: release gate skips never-started action_required runs (#1105)
  • fix #894: listen to room sound through a chosen microphone (#1106)
  • fix #571: nightly CPU eval to eval-results branch + chrF drop alert (#1108)
  • fix #1107: pin the Vulkan SDK installer SHA-256 so releases build the Vulkan pack (#1109)
  • fix #571: eval-nightly keeps history when the eval-results fetch fails (#1110)
  • fix #894: verify_offline.py --mic (offline check for live microphone input) (#1111)
  • fix #894: tlwu-menu listen-only asks for microphone (room sound) (#1112)
  • fix #1113: menuopts mic-args with no answer prints nothing (#1114)
  • fix #1054: last OPUS drift — e2e.yml comment + ARCHITECTURE default engine (#1115)
  • fix #1116: release-assets job installs python3-tk and decodes install.ps1 bytes (#1117)
  • fix #1118: hide subtitle debug stats in release (#1119)
  • release: bump version to 0.4.0 (#1124)
  • fix #1121: mobile subtitle screen redesign (bubbles + round mic button) (#1127)
  • fix #1126: version 0.2.0 + show in app (#1128)

Thanks

Thank you to everyone outside the core team who reported problems, tested and sent fixes:

v0.3.8

Choose a tag to compare

@github-actions github-actions released this 15 Nov 16:33
27e3106

Download

  • Windows 10/11: tinglongwu-setup-0.3.8.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.8/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.8/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪
Tiếng Việt (vi) ⚪ ⚪ ⚪ ⚪ ⚪
ไทย (th) ⚪ ⚪ ⚪ ⚪ ⚪
Bahasa Indonesia (id) ⚪ ⚪ ⚪ ⚪ ⚪
Español (es) ⚪ ⚪ ⚪ ⚪ ⚪
Deutsch (de) ⚪ ⚪ ⚪ ⚪ ⚪
Français (fr) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • Subtitles can label who is speaking (a small add-on pack downloads only when you turn it on), and a new 情境 row lets you pick meeting or video.
  • The mic now follows your Windows default device, and speak mode holds back more background noise.
  • The setup wizard and app speak 繁體中文, 简体中文, 日本語 and Русский, and the settings window can be resized and wraps to one column when narrow.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 119 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 119 MB
  • SHA256: 01f9cc6675cac53da8f38f60cad9b64750a7ebca28dc713c52ccf92bf5952325 (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: validated by hw-tests on the maintainer's self-hosted NVIDIA runner for commit 556de166c010. The zip below is the slim core: it has no CUDA libraries inside, was built on a GPU-less CI runner and tested there on the CPU path only.

Optional packs (downloaded on demand)

pack size sha256
ja 1133.1 MB dc5f44fcf1fa
ko 1133.1 MB dc5f44fcf1fa
ru 1196.3 MB 8b7bfd559dbc
speak 120.6 MB ef9f72fc5e31

Changes since v0.3.7

  • fix #1054: align README/docs with code (#1055)
  • fix #1002: wizard in 繁體中文 / 简体中文 / Русский too, one language per page (#1056)
  • fix #1057: Japanese UI table, so ui_lang=ja opens a Japanese app (#1062)
  • feat #1064: mobile grey source-text draft (顯示原文草稿) (#1066)
  • feat #1065: mobile device probe + 進階功能 recommendations (#1067)
  • fix #1068: installer language test expects ja→ja after #1057 (#1069)
  • fix #1063: Japanese text menus (ja.sh / ja.psd1) (#1071)
  • fix #1002: Japanese wizard opens a Japanese app (install.ps1 ui_lang ja) (#1072)
  • fix #1054: PRIVACY.md lists the GPU pack download (#1073)
  • fix #934: install-clean tag dispatches no longer cancel each other (#1074)
  • fix #1054: docs: drop OPUS fallback wording, note leftover OPUS code (#1075)
  • fix #1076: 看影片翻譯 on iOS (ReplayKit broadcast extension + PiP subtitles) (#1077)
  • fix #1079: one subtitle size slider drives overlay and in-app screen (#1081)
  • fix #1078: refit floating subtitle overlay on rotation (#1082)
  • fix #1076: TestFlight upload keeps working before the App Group exists (#1085)
  • fix #1054: align zh README badges with en (partial) (#1086)
  • fix #1070: 0.3.8 listen side — speaker labels (on-demand pack), 情境 row, mic follows default, strict speak gate (#1087)
  • fix #957: hotfix summary reads the gate result after the last '=' (#1090)
  • fix #1054: support.json next_release v0.3.2 -> v0.3.8 (#1091)
  • fix #1092: pending line and draft follow subtitle size slider (#1093)
  • fix #1089: speak picks VB-CABLE's CABLE Input over Hi-Fi Cable Input (#1094)
  • fix(mobile): overlay_geometry int types so analyze passes (#1095)
  • fix #939: settings grid wraps to one column below the two-column width (mockup A) (#1096)
  • release: bump version to 0.3.8 (#1099)
  • fix #611: topic notes window in the GUI (listen leg feeds TopicPanel) (#1101)

Thanks

Thank you to everyone outside the core team who reported problems, tested and sent fixes:

v0.3.7

Choose a tag to compare

@github-actions github-actions released this 15 Nov 11:25
a03fd43

Download

  • Windows 10/11: tinglongwu-setup-0.3.7.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.7/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.7/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪
Tiếng Việt (vi) ⚪ ⚪ ⚪ ⚪ ⚪
ไทย (th) ⚪ ⚪ ⚪ ⚪ ⚪
Bahasa Indonesia (id) ⚪ ⚪ ⚪ ⚪ ⚪
Español (es) ⚪ ⚪ ⚪ ⚪ ⚪
Deutsch (de) ⚪ ⚪ ⚪ ⚪ ⚪
Français (fr) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • New chat-style main window: the other side on the left, you on the right, with the original words in small text under each line; one big Start button.
  • A mic on/off button (like mute in Meet), and one click shows the subtitles floating over a video.
  • Settings now use plain words, with the technical options folded away under Advanced. Speak on Windows is still experimental.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 119 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 119 MB
  • SHA256: d250a113642d3d39f54eee0b9431938b37f7f879bf4f997950c96f15fc43b697 (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: validated by hw-tests on the maintainer's self-hosted NVIDIA runner for commit 66eef851c7ab. The zip below is the slim core: it has no CUDA libraries inside, was built on a GPU-less CI runner and tested there on the CPU path only.

Optional packs (downloaded on demand)

pack size sha256
ja 1133.1 MB dc5f44fcf1fa
ko 1133.1 MB dc5f44fcf1fa
ru 1196.3 MB 8b7bfd559dbc
speak 120.6 MB ef9f72fc5e31

Changes since v0.3.6

  • fix #946: updater applies only the app layer when the runtime key is unchanged (#952)
  • fix #959: rules §14 "Hotfix lane" paragraph + hotfix.yml as release-critical (#963)
  • fix #966: GPU pack message instead of 'update your driver' when cuBLAS/cuDNN are missing (#972)
  • fix #957: hotfix.yml dry run builds base tag + fix once and tests that build (#975)
  • fix #942: route Hindi (hi) through Hy-MT2 and gate hi->zh in eval-langs (#984)
  • fix #976 item 6: mobile-only PRs skip the desktop e2e / audio-routing / acceptance-windows steps (#986)
  • fix #981: session_route derives src/dst and capture from (mode, My language, Other side language) (#990)
  • fix #968: speak noise-gate eval (keyboard / mouse / fan + 10 zh clips) gated in eval-langs zh->en (#991)
  • fix #680: wrong-translation prefill adds the usage answer behind a ticked checkbox (#992)
  • fix #993: freeze-guard triggers on pull_request so it actually runs (#994)
  • fix #976: rules section 17 two lanes + release snapshot branch (items 2, 3, 5) (#995)
  • fix #967: tlwu-setup and the app resolve the same models dir (#996)
  • fix #967: setup summary says the voice is not installed (#997)
  • fix #966: first-run setup installs the GPU pack on NVIDIA (tlwu-setup --gpu-pack) (#999)
  • fix #1002: setup wizard language page; install.ps1 -Lang; ui_lang from the wizard (#1003)
  • fix #967: setup voice note also checks the Kokoro venv (#1004)
  • fix #1001: one Start runs listen + speak together (no mode choice) (#1008)
  • fix #974: speak smoke on a clean Windows VM (install-clean.yml speak-smoke) (#1009)
  • fix #1007: route --talk in run_setup_worker to the setup CLI (#1010)
  • fix #974: speak smoke review fixes (ClipMic.close, 30 s window, 2-channel capture) (#1011)
  • fix #1012: cut the release snapshot at the first commit with whats_new notes (#1013)
  • fix #1014: self-test and Hy-MT2 load error report "CPU lacks AVX2" (#1015)
  • mobile: 看影片翻譯 (playback capture + floating subtitles) (#1017)
  • fix #1016: require install-clean (speak-smoke) for the release snapshot (#1018)
  • fix #1007: voice install venv uses a system Python, not tlwu.exe (#1019)
  • fix #1020: clearer GUI home (freeze subtitles, localized status, 聽攏有, auto-restart on language change) (#1022)
  • fix #1023: keep Room's WorkDatabase_Impl constructor in R8 release builds (#1024)
  • feat #1025: iOS TestFlight upload from GitHub Actions (#1026)
  • fix #982: GUI shows only My language + Other side language (#1027)
  • fix #1029 part 1: copy models out of install-time asset packs on first launch (#1030)
  • fix #781: CPU-mode A/B script (cpu_mode_ab.py) (#1033)
  • feat #1031 part 2: first-run screen while Play downloads the model packs (#1035)
  • fix #1034: testflight pre-checkout steps use github.workspace (#1036)
  • fix(ci): mobile-ios secret check runs from workspace (#1039) (#1040)
  • fix #947: windows-build upgrade job covers app-layer-only and full update (#1041)
  • feat #1028: chat-style GUI, mic toggle, plain-language Settings (#1042)
  • ci(mobile): build iOS on macos-26 / Xcode 26 (#1043) (#1044)
  • fix #1007: in-app voice install runs tlwu.exe --run-setup in the core layout (#1045)
  • release: bump version to 0.3.7 (#1047)
  • fix #1048: mobile subtitle quality (#1049)
  • fix #958: hotfix.yml publish=true releases exactly the tested files (#1050)
  • fix #611: topic panel model + CUDA/bake-off default (part 4, partial) (#1051)
  • fix #1052: speak half-duplex (no 3x sends), send straight away by default, bubbles show original (#1053)
  • fix #773: ttfs_ci summary reports install size on disk (#1059)
  • fix #1060: ttfs green again — Linux install gets llama-cpp-python, Windows runs the wizard's tlwu.exe, macOS runner (#1061)

Thanks

Thank you to everyone outside the core team who reported problems, tested and sent fixes:

v0.3.6

Choose a tag to compare

@github-actions github-actions released this 15 Nov 05:09
b406c0a

Download

  • Windows 10/11: tinglongwu-setup-0.3.6.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.6/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.6/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪
Tiếng Việt (vi) ⚪ ⚪ ⚪ ⚪ ⚪
ไทย (th) ⚪ ⚪ ⚪ ⚪ ⚪
Bahasa Indonesia (id) ⚪ ⚪ ⚪ ⚪ ⚪
Español (es) ⚪ ⚪ ⚪ ⚪ ⚪
Deutsch (de) ⚪ ⚪ ⚪ ⚪ ⚪
Français (fr) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • Speak mode no longer sends keyboard or background noise as made-up sentences: unclear lines are held back and shown greyed.
  • A "Wrong translation" button on each subtitle line and a "Copy AI prompt" entry in the Help menu make reporting a problem one click.
  • Updates are quieter (asked at most once a day, installed at the next launch), and names in the app show as languages, not codes.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 119 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 119 MB
  • SHA256: ce178c85fdacd103cdc91d5e1592295e670a4a8f646c63edd8205ada65112bd5 (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: validated by hw-tests on the maintainer's self-hosted NVIDIA runner for commit 11e5e0e5bc2f. The zip below is the slim core: it has no CUDA libraries inside, was built on a GPU-less CI runner and tested there on the CPU path only.

Optional packs (downloaded on demand)

pack size sha256
ja 1133.1 MB dc5f44fcf1fa
ko 1133.1 MB dc5f44fcf1fa
ru 1196.3 MB 8b7bfd559dbc
speak 120.6 MB ef9f72fc5e31

Changes since v0.3.5

  • fix #888: windows-build saves the Nuitka cache from dev pushes only; dev pushes never cancelled (#893)
  • fix #677: one-click 'wrong translation' report (GUI + overlay menu) (#895)
  • fix #780: GUI shows "CPU mode" status line with live per-line MT ms (#898)
  • fix #884: GUI shows language/choice names instead of raw codes (#900)
  • fix #172: install-test runs real tiny-model download + CPU self-test (#904)
  • fix #779: update prompt at most once a day; quiet download installs at next launch (#905)
  • fix #897: Add-ons menu and screen strings come from i18n (en/ru/zh-TW) (#906)
  • fix #886: Adjust panel fits a 1366x768 screen with Advanced open (#908)
  • fix #730: mobile subtitle language picker (zh-Hant / en) (#910)
  • fix #680: one-time local "what do you mainly use it for?" question (#913)
  • fix #26: build provenance for every release asset (attest-build-provenance) (#914)
  • fix #665: blocking ruff check in CI (F, E9, B, UP) + existing violations fixed (#917)
  • fix #723: rules §14 recovery path for tagged-but-unpublished release (#919)
  • fix #918: first-Start download bar follows the largest file, not max over all tqdm lines (#920)
  • fix #662: 14-day backfill of inferred dimensions + triage os/lang rule (#922)
  • fix #923: Windows core in two layers (cached runtime layer + app layer) (#924)
  • feat #927: iOS app (shared Flutter code, Swift native channel, TestFlight config) (#928)
  • fix #805: pin the official VB-CABLE package SHA-256 (#929)
  • fix #775: install-clean ttfs job (Linux + macOS) publishes ttfs_seconds (#930)
  • fix #765: reuse the previous release's GPU pack when its pinned wheels did not change (#931)
  • fix #642: per-language eval matrix from eval/languages.json (#932)
  • fix #677: per-line "Wrong translation" action in the main-window log (#933)
  • fix #934: install-clean ttfs measures Windows (released setup exe) (#935)
  • fix #936: stats page refreshes even when scheduled runs are skipped (#937)
  • fix #939: resizable settings window with menu bar, remembered size (#941)
  • fix #642: 20 pinned clips per pair, recorded chrF for every runnable pair, hi->zh blocked (#942) (#943)
  • fix #939: drop unused import (dev red after #665 + #939) (#949)
  • fix #945: release publishes tinglongwu-app-layer.zip + manifest (runtime_key) (#951)
  • fix #940: report-with-ai prompt doc (two-stage support desk) (#954)
  • fix #948: #924 review follow-ups D1 D2 D6 D7 (#955)
  • fix #944: ci test fails a PR with an unknown git identity (#960)
  • fix #956: scripts/hotfix_plan.py decides tag, base and fix for a hotfix (#961)
  • fix #770: DEFAULT_LISTEN_WHISPER stays small (tiny/base fail the eval gate) (#962)
  • fix #773: ttfs windows job records setup_time.json on a clean VM (#964)
  • release: bump version to 0.3.6 (#970)
  • fix #967: offer one-click Kokoro voice install when speak starts without it (#971)
  • fix #968: GUI speak runs the #249 speak input gate; rejected lines shown greyed (#977)
  • fix #976: rolling dev build (dev-latest pre-release from each dev push) (#978)
  • fix #940: "Copy AI prompt" in the Help menu + README links (#980)
  • fix #983: start-time settings changed during a run show "restart to apply" + one-click restart (#987)
  • release-fix #969: whats_new.md v0.3.6 section (unblocks auto-release) (#988)
  • fix #953: freeze-guard check fails PRs into dev while a release freeze is on (#989)

Thanks

Thank you to everyone outside the core team who reported problems, tested and sent fixes:

tinglongWu mobile (preview)

Pre-release

Choose a tag to compare

@github-actions github-actions released this 13 Nov 06:17
7e830ed

1. Install the app: tinglongwu-mobile-arm64.apk

2. Get the translation pack (1.4 GB, once): tlwu-mobile-models.zip

Open the app, tap Import models, pick the zip.

Installed from Google Play instead? Nothing to import: Play downloads the models right after install.

Details

Rolling preview of the Android app (arm64), rebuilt from dev on every change under mobile/.
This is a preview, not a versioned release; it is replaced each time.
Full steps: mobile/README.md

  • commit: 7e830ed
  • signed with: release key
  • APK SHA256: 6d95b5c48acbe46dae278f49455da95f50b4c7ada88986b21dd18976c5611cab
  • models zip SHA256: 208aa2985552dd3c9af37502f7cbbac9929d358bc7cbb7b947948b444d748448

dev build 62fdd96 (untested)

Pre-release

Choose a tag to compare

@github-actions github-actions released this 16 Nov 00:04
62fdd96

Development build, untested. Not a release. Built from dev at 62fdd96 (code version 0.4.0).

Last change: fix #1139: release plan ignores zero-job PR runs flipped to failure after merge (#1140)

Only the automatic build + start-up smoke ran on it; the release checks run when dev goes to main.
For normal use, take the latest release: https://github.com/sunososobro-hub/tinglongwu/releases/latest

v0.3.5

Choose a tag to compare

@github-actions github-actions released this 15 Nov 00:19
bdaf48d

Download

  • Windows 10/11: tinglongwu-setup-0.3.5.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.5/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.5/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪
Tiếng Việt (vi) ⚪ ⚪ ⚪ ⚪ ⚪
ไทย (th) ⚪ ⚪ ⚪ ⚪ ⚪
Bahasa Indonesia (id) ⚪ ⚪ ⚪ ⚪ ⚪
Español (es) ⚪ ⚪ ⚪ ⚪ ⚪
Deutsch (de) ⚪ ⚪ ⚪ ⚪ ⚪
Français (fr) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • The first screen now just asks Listen or Speak and the other side's language; everything else sits under Advanced and a File / Settings / Help menu.
  • Speak mode waits for your OK before sending by default, and speak setup can install the virtual audio cable in one click.
  • On computers without an NVIDIA card, the app picks lighter settings automatically and marks models that are too heavy; setup progress now shows real download sizes live.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 132 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 132 MB
  • SHA256: 106b70237c2990bcacea655b14ba6ae1a17d156d4da10d74b5792d9f7ce36163 (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: validated by hw-tests on the maintainer's self-hosted NVIDIA runner for commit b1196f98f34a. The zip below is the slim core: it has no CUDA libraries inside, was built on a GPU-less CI runner and tested there on the CPU path only.

Optional packs (downloaded on demand)

pack size sha256
ja 1133.1 MB dc5f44fcf1fa
ko 1133.1 MB dc5f44fcf1fa
ru 1196.3 MB 8b7bfd559dbc
speak 120.6 MB ef9f72fc5e31

Changes since v0.3.4

  • fix #113: real-Tk GUI tests run in CI under xvfb; display skips fail there, modal dialogs fail instead of hanging (#812)
  • fix #722: checked-in pack manifest passed to release notes by release.yml (#816)
  • fix #818: commit Piper ru dmitri pin (#819)
  • fix #762: release.yml windows job builds the slim core zip (--core) with the core budget gate (#820)
  • fix #753: setup progress shows the real download size (one named MB table) (#823)
  • fix #804: one-click VB-CABLE install for speak mode (redo of #811 with the review fixes) (#824)
  • fix #755: setup progress page updates live instead of freezing at "Starting install.ps1 ..." (#825)
  • fix #679: add_eval_case.py turns translation-error reports into eval cases (+ weekly loop docs) (#826)
  • fix #808: rules s1 retire home-PC worker lane (#828)
  • fix #763: build_windows.py --gpu-pack-only builds the GPU pack without Nuitka (#829)
  • fix #831: rules - fewer Opus reviews (no size gate, narrower sensitive paths, paperwork is not a FAIL) (#835)
  • fix #832: shared I/O-free send-mode state machine (SendDecider) (#836)
  • fix #764: release.yml builds and attaches the GPU pack (parallel job) (#837)
  • fix #738: File / Settings / Help menu bar for rarely used actions (#838)
  • fix #806: anonymous update-check counter (date/version/OS only) (#839)
  • fix #834: GUI send-mode setting (default confirm) + quality notice (#840)
  • fix #766: tlwu-setup --default-listen installs only the default en->zh-TW listen models (#841)
  • fix #767: build_installer.py --core-dir builds the prebuilt-core wizard (tlwu.exe) (#842)
  • fix #744: stats page shows quota minutes/day, PRs/day + per 100 quota min, release time (#844)
  • fix #774: scripts/ttfs.py times install -> setup -> first subtitle (#847)
  • fix #768: auto-update runs as tlwu.exe --run-selfupdate in a prebuilt-core install (#851)
  • fix #848: auto-release defers hw-tests when the home runner is offline (30 min) (#852)
  • fix #769: clean_machine_check.py setup-time measures wizard start -> ready (#853)
  • fix #818: list ru pack (Hy-MT2 + Piper dmitri) in packaging/packs.json (#854)
  • fix #737: first screen = Listen / Speak + other side's language, rest under Advanced (#856)
  • fix #777: stats page shows the last 10 time-to-first-subtitle values per OS (#857)
  • fix #776: release_check.py --ttfs prints time to first subtitle, warns above 180 s (#858)
  • fix #849: audio-routing job on windows-latest (VB-CABLE, tone RMS check, speak_selftest) (#859)
  • fix #833: GUI speak hold-then-send with countdown, Stop button and Space/Esc (#860)
  • fix #772: release.yml builds the setup wizard from the release's core zip (--core-dir) + size report (#861)
  • fix #800: release.yml builds a Vulkan llama.cpp pack zip (vulkan-pack job) (#862)
  • fix #771: core wizard runs tlwu.exe setup --default-listen (no pip, no language page) (#863)
  • fix #845: docs say auto-update checks GitHub once a day; TLWU_OFFLINE=1 for zero network (#864)
  • fix #614: add-ons manager logic (pack rows, one-time-internet note, GPU pre-tick, offline message) (#866)
  • fix #796: low-end CPU gate runs on dependency-only PRs, base run uses its own venv (#867)
  • fix #705: install the VC++ runtime before llama-cpp; stop before the model download if the engine cannot load (#868)
  • fix #822: CLI speak --send-mode defaults to confirm, notice for countdown/auto (#869)
  • fix #855: ttfs.py counts each downloaded file once (#870)
  • fix #801: tlwu-setup --vulkan-pack installs the release Vulkan llama.cpp pack (#871)
  • fix #865: GPU pack ships CUDA llama build so Hy-MT2 offloads with the core zip (#872)
  • fix #681: hw-tests nightly on dev (hw-nightly.yml gate + dispatch), commit SHA in summary (#874)
  • fix #780: CPU-mode auto settings, too-heavy model marks, CPU mode status line (#877)
  • ci: e2e skips heavy steps on docs-only PRs; stats-dashboard drops push trigger (#873) (#878)
  • fix #827: score eval/user_cases.jsonl text cases with an MT-only pass (--text-cases) (#879)
  • ci: eval.yml dispatch input whisper tiny/base/small (#876) (#880)
  • fix: GPU pack llama build only when its llama-cpp-python version matches the core (#881) (#882)
  • fix #614: Settings > Add-ons screen (installed/available packs, add, remove, local zip) (#883)
  • fix #885: read-aloud on the captured device turns itself off with one localized line; GUI never shows CLI flags (#889)
  • release: bump version to 0.3.5 (#891)
  • fix #739: missing speech model downloads on first Start with a progress bar (#892)
  • fix #896: missing output device no longer blocks listening (#902)
  • fix #901: installer .iss line 413 no longer starts with #13#10 (release-fix) (#903)
  • fix #662: infer mode/os/lang from changed files in the quota Sankey (#907)
  • fix #909: ignore a stale Vulkan pack (version.txt != app version) (#911)
  • fix #661: roadmap page generated from milestones + latest release (#912)
  • fix #915: Vulkan pack zip carries version.txt; tests use the app version (dev green again) (#916)
  • fix #662: stats Sankey spl...
Read more

v0.3.4

Choose a tag to compare

@github-actions github-actions released this 14 Nov 10:16
41d8376

Download

  • Windows 10/11: tinglongwu-setup-0.3.4.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.4/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.4/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪
Tiếng Việt (vi) ⚪ ⚪ ⚪ ⚪ ⚪
ไทย (th) ⚪ ⚪ ⚪ ⚪ ⚪
Bahasa Indonesia (id) ⚪ ⚪ ⚪ ⚪ ⚪
Español (es) ⚪ ⚪ ⚪ ⚪ ⚪
Deutsch (de) ⚪ ⚪ ⚪ ⚪ ⚪
Français (fr) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • Setup now pre-selects the English and Chinese language packs, and Korean is on the language page, so the first run works without picking packs by hand.
  • Language pack downloads resume after an interruption instead of starting over, and each file is checked before use.
  • Speak mode no longer offers a made-up subtitle-credit line when the room is silent.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 1597 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline; lib_cpu/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK (llama_cpp: CUDA build not loaded on the GPU-less runner, CPU fallback loaded)
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 1598 MB
  • SHA256: 2d622e106990ad4ec85ba0ac06d4b9780755e3adcc7dba78f701e807af8b176d (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: validated by hw-tests on the maintainer's self-hosted NVIDIA runner for commit cd56f49a8ccd. The zip below was built on a GPU-less CI runner and tested there on the CPU path only.

Changes since v0.3.3

  • fix #607: hw-tests scripts spawn children with CREATE_NO_WINDOW (#617)
  • feat(mobile): Google Play AAB + model asset packs + internal-track upload (#719) (#721)
  • fix #570: eval.yml PR gate (paths filter, cached CPU models, chrF/MT gate) (#725)
  • fix #671: install.sh asks which language to understand (--ask-lang) (#726)
  • fix #107: Hy-MT2 temperature sweep (0.7/0.2/0) + U+FFFD count; ship greedy (#727)
  • fix #689: README visual-first first screen (GIF + 3-node Mermaid + mode icons + badges), text folded (#728)
  • fix #688: route --packs getters through resume.fetch (Range resume + SHA256) (#731)
  • docs(privacy): Android app section for Google Play (#732) (#733)
  • fix #511: record draft first-line latency in --latency-jsonl (#741)
  • fix #742: rules - faster pipeline (area lock, PASS honoured, narrower gates, earlier unfreeze) (#743)
  • fix #646: stats Sankey platform column shows mobile / all (#748)
  • fix #710: GUI Speak speaks the translation into VB-CABLE "CABLE Input" (#749)
  • fix #589: auto-release merges + publishes the green release PR (kill switch) (#757)
  • fix #751: installer pre-checks English (and zh) on the language page (#761)
  • fix #785: rules picker overrides prompts, stale release PR never freezes, bump cadence (#786)
  • fix #750: low-end CPU eval gate (p50/p90 line latency, peak RSS), also on the release PR (#787)
  • fix #746: MT backend probe CUDA > Vulkan > CPU, TLWU_DEVICE=vulkan, Vulkan->CPU fallback (#788)
  • fix #754: Korean (ko) language pack on the setup page, pre-checked on a Korean OS, ko->zh eval slice tooling (#790)
  • fix #759: drop YoYo/優優獨播劇場 silence hallucination in speak mode (#791)
  • fix #792: rules §13 infra cap (<=10%) + 'which user does this help' test (#794)
  • feat #793: stats dashboard release downloads trend + per-release (#795)
  • fix #729: mobile README claims match the default pack (no Russian) (#798)
  • fix #789: pin the ko->zh-TW FLEURS slice and record its chrF baseline (#803)
  • fix #760: speak mode countdown auto-send; low-confidence lines still need confirm (#809)
  • fix #756: update controls in the Check setup row; uninstall hint for setup installs (#813)
  • release: bump version to 0.3.4 (#817)
  • fix #782: ASR engine selection cuda > vulkan > cpu with whisper.cpp Vulkan recognizer (mocked) (#821)

Thanks

Thank you to everyone outside the core team who reported problems, tested and sent fixes:

v0.3.3

Choose a tag to compare

@github-actions github-actions released this 14 Nov 05:25
059fbfc

Download

  • Windows 10/11: tinglongwu-setup-0.3.3.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.3/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.3/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪
Tiếng Việt (vi) ⚪ ⚪ ⚪ ⚪ ⚪
ไทย (th) ⚪ ⚪ ⚪ ⚪ ⚪
Bahasa Indonesia (id) ⚪ ⚪ ⚪ ⚪ ⚪
Español (es) ⚪ ⚪ ⚪ ⚪ ⚪
Deutsch (de) ⚪ ⚪ ⚪ ⚪ ⚪
Français (fr) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • Fixes and improvements since v0.3.1; the full list is below.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 1597 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline; lib_cpu/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK (llama_cpp: CUDA build not loaded on the GPU-less runner, CPU fallback loaded)
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 1598 MB
  • SHA256: 123570620c5e4af3797f60a5dd1c337956fa31dfcaf1f431b05f1d4cc81cf94e (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: validated by hw-tests on the maintainer's self-hosted NVIDIA runner for commit 8fbf873c846e. The zip below was built on a GPU-less CI runner and tested there on the CPU path only.

Changes since v0.3.2

  • fix #344: weekly stale-branch pruning workflow (dry_run on dispatch) (#583)
  • fix #341: packaged-app smoke on Windows in CI (exe --smoke + GUI launch) (#586)
  • fix #581: mobile zh speech bypasses Hy-MT2 (passthrough) (#590)
  • fix #596: language priority by market (docs part) (#597)
  • docs: README by user priority + VISION.md + demo GIF (#599)
  • fix #593: speak_selftest waits for a quiet CPU and re-runs a load-explained stream gap once (#603)
  • fix #600: installer pre-ticks Force CPU when no NVIDIA driver (#604)
  • Docs/fast lane (#605)
  • fix #610: topic summary CPU bake-off script (#616)
  • fix #602: tier-2 FLEURS eval pins (vi/th/id/es/de/fr -> zh) with chrF scores (#618)
  • fix #609: low-priority topic summarizer worker + Markdown topic notes next to --save (#621)
  • sync: main -> dev after docs fast lane #622 (#623)
  • fix #612: core zip flag (--core) + release_check size/CUDA budget (#624)
  • fix #626: stats page vision block from docs/VISION.md (#630)
  • fix #585: mode / lang dimension labels (rules §12, backfill, stats dashboard groups) (#631)
  • fix #632: fm subprocess pipe explicitly UTF-8 (macOS Chinese garbling) (#633)
  • fix #625: mobile queue merge skips cross-language pairs (#634)
  • fix #608: topic segmenter (pause / min / max / embedding-similarity drop) (#635)
  • fix #628: stats page speed tiles show medians only (#636)
  • fix #71: install-test on release, release-asset one-liners, timing, README badge (#637)
  • fix #638: enforce core zip size budget in windows-build (#640)
  • fix #613: GPU pack manifest, SHA256-verified install, CPU fallback when absent (#641)
  • fix #627: stats page quota-flow Sankey, type-mix bars, faucet level (#644)
  • fix #606: Hy-MT2-1.8B on full FLORES-200 devtest (cloud CPU workflow + runner) (#645)
  • fix #647: test mobile queue-trim via ContextBudget.trimPlan (#648)
  • fix #629: stats page dense side-by-side activity layout with captions (#649)
  • fix #650: speak latency test stopwatch marks text-ready like the product (#651)
  • fix #591: add CODE_OF_CONDUCT.md (Contributor Covenant 2.1) (#652)
  • fix #620: Indian English vs Hindi -> Chinese CPU eval (#653)
  • fix #639: dispatch install-test after release publish; skip mobile-* tags (#654)
  • fix #476: uninstall for Linux and macOS (#660)
  • release: bump version to 0.3.3 (#668)
  • fix #672: installer language page (pre-tick OS locale + GPU pack) (#675)
  • fix #656: FLORES-200 full devtest for ko + tier-2 languages (vi/th/id/es/de/fr) (#676)
  • fix #673: GUI offers a missing language pack at the moment it is needed (#682)
  • fix #21: listen-mode --end-silence / --pause-send knobs for the 0.5 s latency A/B (#685)
  • fix #678: link GitHub Discussions from README, issue chooser, suggestion form, CONTRIBUTING (#686)
  • fix #663: split reverse/proto.py into player / synth / translator / args / session / listen_side (#695)
  • fix #671: first setup step asks which language to understand (#696)
  • fix #588: harden tlwu_rules.md for auto-release (sections 14-15) (#697)
  • fix #690: README/CONTRIBUTING doc fixes (#699)
  • fix #701: installer reserves ~6 GB disk (ExtraDiskSpaceRequired) (#702)
  • fix #615: optional pack manifest, upgrades keep installed packs, release notes list pack sizes (#708)
  • fix #664: rotating diagnostic log, no silently swallowed exceptions (r2) (#709)
  • fix #712: rules - fix/N heads only, mergers flag others, leftover branch not a lock (#713)
  • fix #700: installer progress bar follows [n/N] from install.ps1 (#714)
  • fix #703: zh full READMEs put setup.exe first, manual commands under advanced (#717)
  • fix #716: release held text at pause-send after a forced cut / held fragment (#718)
  • fix #698: Windows auto-update (check, verified download, silent reinstall, opt-in) (#720)
  • fix #314: first-launch 3-step overlay, "?" off the control bar (r2 of #666) (#724)

Thanks

Thank you to everyone outside the core team who reported problems, tested and sent fixes:

v0.3.2

Choose a tag to compare

@github-actions github-actions released this 13 Nov 15:06
851ac9a

Download

  • Windows 10/11: tinglongwu-setup-0.3.2.exe. Run it, accept the license, then keep clicking Next; stay online while it downloads about 3 GB of packages plus 2-3 GB of models (10-40 minutes).
    SmartScreen says "Windows protected your PC"? Click More info, then Run anyway (the installer is not code-signed).
  • macOS: paste into Terminal: curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.2/mac_install.sh | bash
    It needs Homebrew and asks for your Mac password while installing the BlackHole audio driver (typing shows nothing). If a step fails, run the same command again.
  • Linux (Ubuntu 24.04 tested): curl -fsSL https://github.com/sunososobro-hub/tinglongwu/releases/download/v0.3.2/install.sh | bash
    Run it as a normal user, not root. If system packages are missing it asks Install them now with apt (uses sudo)? [y/N]; answer y (or set TLWU_YES=1).
OS Install Translate subtitles Speak (voice) GPU CPU-only
Windows 10/11 🟡 🟢 ⚪ 🟢 🟡
macOS 🟡 ⚪ ⚪ ⚪ ⚪
Linux (Ubuntu 24.04) 🟢 🟢 🟢 ⚪ 🟡
Language Hear it To 繁體中文 To 简体中文 To English Speak it (voice)
中文 (zh) 🟢 — — 🟢 🟢 Windows
English (en) 🟢 🟢 🟡 — ⚪
日本語 (ja) 🟢 🟢 🟡 ⚪ ⚪
Русский (ru) 🟡 🟡 🟡 ⚪ 🟢 Linux
한국어 (ko) ⚪ ⚪ ⚪ ⚪ ⚪

🟢 tested on a real device or in the release gate · 🟡 tested only in CI or the eval, not yet on a real device (Windows install: the setup wizard; the portable zip is real-device tested. CPU-only: latency not measured on a real laptop) · ⚪ not tested yet or not supported · — same language

Built with help from 4 people

sunososobro-hub
sunososobro-hub

💻 📖 🚧
cook1e0728
cook1e0728

💻
psgo555
psgo555

🐛
jchodev
jchodev

💻 🤔

AI helpers: 🤖 Claude (Anthropic) · 🤖 CodeRabbit (code review)
💻 code · 🐛 bug report · 📖 docs · 🤔 ideas · 🚧 maintenance · 🤖 AI tool (not counted as people)

What's new

  • Fixes and improvements since v0.3.1; the full list is below.

Windows package (tinglongwu-win64.zip)

check result
size < 2 GiB PASS: 1597 MB
check_dist_dlls.py (unpacked zip) PASS: every DLL dependent resolves
llama CPU instructions (AVX2 baseline) PASS: lib/ggml-cpu.dll: AVX2 baseline; lib_cpu/ggml-cpu.dll: AVX2 baseline
tlwu.exe --smoke PASS: OK (llama_cpp: CUDA build not loaded on the GPU-less runner, CPU fallback loaded)
tlwu.exe --selftest (CPU, whisper tiny + Hy-MT2) PASS: OK
  • tinglongwu-win64.zip: 1598 MB
  • SHA256: 003d311cf6e842c77720f6685a064c3b405a49e81e0b116d65d2e998725f2ef8 (check with Get-FileHash tinglongwu-win64.zip; also in SHA256SUMS.txt)
  • Built by the release workflow on GitHub Actions (windows-latest); unsigned, not reproducible.

GPU: validated by hw-tests on the maintainer's self-hosted NVIDIA runner for commit 9fccc0de439b. The zip below was built on a GPU-less CI runner and tested there on the CPU path only.

Changes since v0.3.1

  • fix #335 (part 1): Hy-MT2 translates it/pt/vi/th/id directly (#470)
  • Upstream the Ollama and Apple FM engines from jchodev's fork (#356) (#477)
  • docs: add all-contributors list to README (#480)
  • docs: tell users what to expect during install (#481)
  • fix #352 (part 1): deterministic lane gate for outside PRs (#482)
  • docs: README install table points at the latest release, not main (#483)
  • fix #360: scripts/external_scan.py lists outside-contributor activity (#484)
  • fix #361: python -m tlwu_eval --gate compares a smoke report to eval/baseline.json (#485)
  • fix #486: eval gate fails when no pair was evaluated (#487)
  • fix #121: silence/noise hallucination test in eval smoke (#488)
  • fix #127: wizard, shortcuts and tlwu.exe share the committed app icon (#491)
  • docs: roadmap as version transit map (single source support grid) (#492)
  • fix #373: meeting --save records both speakers (▶ 我 / ◀ 對方, side me/them) (#493)
  • fix #384: speaker label on the subtitle line, speaker in --save, rename/merge logic, eval adapter (part 1) (#495)
  • feat: stats dashboard from day one + label dimensions (#496)
  • feat: 聽龍 mascot, app icons, CLI splash (#497)
  • fix #371: linux menu speak flow no longer asks confirm-vs-direct (#498)
  • feat: stats page leads with throughput and speed (#499)
  • fix #123: eval --history JSONL + --trend SVG charts (#500)
  • fix #122: eval smoke records peak VRAM per pair and flags over-budget red (#501)
  • fix #455: compact control bar + "Adjust" panel holding every setting (#503)
  • fix #494: confirm mode records side=me rows only when the line is shown (#504)
  • fix #389: merge short finished fragments before MT (--min-words, per preset) (#505)
  • feat #40: meeting menu optional local preview, route line, Enter hint (#506)
  • feat #55: issue forms for bug / translation error / suggestion (#507)
  • fix #294: easy start first screen - "Listen to others" / "I speak", Loading -> Ready (#508)
  • fix #319: --voice male|female for the Russian Vosk voice (#509)
  • fix #347: latency p50/p90 tables + tier chart generated from committed runs; FAQ (#510)
  • fix #434: speak mode --paced file input + per-line latency records with TTS first audio (#512)
  • fix #429: ja eval set gets a Tatoeba dialogue subset reported as ja->zh[dialogue] (#514)
  • fix #40: meeting mode: no preview prompt, voice route line, correct send hint (#515)
  • fix #435: nightly eval latency reads native --latency-jsonl records (#517)
  • fix #399: GUI engine dropdown behind a collapsed Advanced section (#518)
  • fix #472: smoke eval --langs it/pt/vi/th/id with COMET x100 floor table (#519)
  • fix #448: reap failed TTS worker, drop unplugged voice_device, GUI voice-off notice (#520)
  • fix #471: remove the OPUS pivot from the listen engine path (part 1: cli/engines/gui) (#522)
  • fix #350: record Qwen3-ASR vs Whisper + Hy-MT2 result (report, chart, FAQ) (#525)
  • ci(mobile): build the Android APK in CI and publish it to the mobile-latest pre-release (#526)
  • fix #390: per-language pre/post-MT hooks; RU person names to nominative of the real gender before MT (#527)
  • fix #524: drop Qwen wording from install.ps1, requirements-*.txt and packaging (#528)
  • fix #523: mobile translator swap Qwen3-4B -> Hy-MT2-1.8B GGUF (#530)
  • fix #516: remove Qwen from docs, notices and test wording (part 1) (#532)
  • fix #391: document history rewrite and fork reset steps (#534)
  • fix #478: keep newest 0.x release as Latest (owner decision a), document policy + guard test (#536)
  • feat(mobile): import models on the phone (system file picker) + models zip on mobile-latest (#537)
  • fix #427: Linux live speak mode exits cleanly on the first Ctrl+C (#538)
  • fix #531: convert mobile Hy-MT2 output to Traditional Chinese (Taiwan) like desktop (#539)
  • fix #535: Hy-MT2 glossary window (current + previous 2 lines), with/without measurement, Hy-MT2 docs (#540)
  • fix #533: keep mobile translator requests inside the 768-token context (#541)
  • fix #40: drop local preview, dedupe meeting menu (owner decision) (#543)
  • fix #345: add docs/support.md (release policy + platform support table) and one README link (#544)
  • ci: disable claude-review (workflow_dispatch only) (#546)
  • fix #239: opt-in local recording (--record): one WAV per line + segments.jsonl (#547)
  • maint #548: add .agent/tlwu_rules.md (#549)
  • ci(mobile): simplify mobile-latest release page (#550)
  • fix #552: rename the Android app display name to 聽龍 (#553)
  • feat(mobile): navy dragon-head launcher icon (#554)
  • docs: v0.3.2 headline card on roadmap + roadmap link on Pages home (#555)
  • fix #112: release workflow accepts pre-release tags cut from dev (#557)
  • fix #559: stats page: show p90 next to median for issue/PR speed (#560)
  • fix #414: one-click release (release PR, Windows build + checks, gated publish) (#561)
  • fix #566: resolve conflicts in place; overlap-aware picking (#567)
  • fix #562: stop installer/setup from requiring or self-testing OPUS packs (#568)
  • rules: restore ideation budget and record the work-mix ratio (#572)
  • fix #569: make_ci_smoke.py rebuilds the smoke set from pinned FLEURS clips (#573)
  • rules: ledger picker + throttle faucet (#575)
  • fix #563: mobile default pack = SenseVoice-small auto-detect (zh/en/ja/ko/yue) (#578)
  • fix #577: speak mode plays the queued TTS when the input ends on its own (#579)
  • fix #521: delete OpusStep and PivotTranslator from tinglongwu/translate.py (#580)
  • fix #582: probe/speed/selftest no longer name OPUS as an engine (#584)
  • release: bump version to 0.3.2 (#587)
  • fix: drop sentencepiece from the exe smoke list (blocks v0.3.2) (#594)

Thanks

Thank you to everyone outside the core team who reported problems, tested and sent fixes:

Read more