sokuji

Organization accounts: choose who pays

• Organization members can choose whether a session uses the organization's balance or their own from the account panel. The title bar shows the selected payer and its balance; during a session, it stays with that session's payer. (#703)
• The account panel shows organization and personal balances, the amount available after running-session reservations, estimated remaining hours, and each side's spending over the last 30 days. (#703)
• Clear notices explain when an organization is out of money, paused, or no longer includes the member, with top-up guidance for organization owners. These notices are localized in all supported app languages. (#703)
• Accounts outside an organization keep the same session requests, balances, and notices as before. (#703)

Full Changelog: https://github.com/kizuna-ai-lab/sokuji/compare/v0.49.0...v0.50.0

• chore(release): v0.50.0
• Merge pull request #703 from kizuna-ai-lab/feat/org-payer
• docs(wallet): a docstring on every declaration in the usage helper
• docs(wallet): docstrings on the usage helper's public surface
• fix(account): keep the request refs current after commit, not during render
• fix(account): another identity's figures may not survive the switch
• fix(account): a superseded usage response may not paint or cache over a newer one
• fix(account): a stale-refresh, a per-render fetch and an invented zero
• fix(account): tell 'no usage' apart from 'not known yet'
• feat(account): choose the payer in the panel, and show both sides' usage
• Merge pull request #705 from kizuna-ai-lab/fix/soniox-reserve-rates
• fix(soniox): mirror the backend's new session budget rates and $1.00/h cap
• Merge pull request #704 from kizuna-ai-lab/chore/drop-form-factor-log
• chore(diagnostics): drop the output form-factor log line
• chore(diagnostics): drop the output form-factor log line
• refactor(account): compute what sessions hold in one helper
• fix(payer): keep a switch held in memory when another document clears or removes the key
• i18n(zh): say "reserved" (预留/預留) for what running sessions hold, as the web dashboard does
• style(account): style the held line under the organization's balance
• feat(account): show the organization's committed balance, not what is left after holds
• fix(payer): keep the switch working when the storage cannot hold it
• test: give the repository-wide import check a 30 s timeout
• fix(payer): hide the payer line for a running session started with no payer
• fix(payer): show the running session's payer, not the live switch
• fix(wallet): warn once on a malformed organization or personal wallet; share the payer field in voice library requests
• fix(session): keep every organization refusal as itself mid-session; move payerBalance out of the storage module
• fix(account): do not invent a $0 personal balance; tidy organization refusals
• fix(account): organization payer gate, notice params and payer display
• chore(i18n): keep the escaped invisible characters in screenContext
• feat(account): show organization and personal balances
• fix(session): organization refusals on voice preview and mid-session reissue
• feat(session): organization refusal notices
• fix(session): carry the payer to the voice preview and read it from the shape
• feat(session): choose who pays before Start
• fix(account): type personal expiringSoon as the backend sends it
• feat(account): read organization and personal balances
• Merge pull request #701 from kizuna-ai-lab/docs/readme-demo-video-v2
• docs(readme): use re-cut demo video (no clipped sentences)
• Add language options to README
• Update README.md
• Merge pull request #700 from kizuna-ai-lab/docs/readme-demo-video-url
• docs(readme): replace demo video with 0.49.0 recording
• Merge pull request #699 from kizuna-ai-lab/docs/readme-demo-video
• docs(readme): add looping demo GIF, remov

Automatic reconnect when the connection drops

• When the connection to Soniox or Qwen drops mid-session, for example when a VPN or proxy switches or Wi-Fi blinks, Sokuji now reconnects on its own. The status line shows "Reconnecting…", then "Connection lost and restored" appears in the conversation, and the conversation and the session timer carry on. (#693)
• Works with Kizuna AI's managed Soniox and Qwen, which fetch new credentials for the reconnect, and with your own Soniox key. (#693)
• A connection that stays open but stops answering is now noticed and reconnected, on Soniox and on Qwen. Before, the session could sit silent until you stopped it. (#693)
• Read-aloud comes back after a reconnect: its connection is renewed together with the transcription. (#693)
• If the network stays down for more than 30 seconds, the session ends with "Connection lost" and the conversation stays on screen. What was said while the connection was down is not translated afterwards. (#693)

Soniox: the app's own voice is no longer transcribed

• When Sokuji reads a translation aloud through this computer's speakers, it now tells Soniox what it is about to say, so the microphone picking up that playback no longer turns into a new line that gets translated again. In face-to-face mode this removes the loop of the app translating its own voice: in our tests, stray lines from the app's own speech went from 51 to 1. (#687)
• On by default; the switch is in Soniox's advanced settings. In Both mode with separate sessions it is not used. (#687)

Full Changelog: https://github.com/kizuna-ai-lab/sokuji/compare/v0.48.0...v0.49.0

• chore(release): v0.49.0
• Merge pull request #693 from kizuna-ai-lab/feat/session-reconnect
• docs(session): Qwen's reissue limit is four a minute (#692)
• docs(session): name the reissue residuals, and what Qwen's limit of 2 a minute now costs (#692)
• fix(session): a session-started retry asks again after its wait whether the run is still open (#692)
• fix(qwen): a token fetched ahead of a stepped-back clock leaves the cache (#692)
• docs(session): a Qwen reconnect's token is minted since its drop, and session-started is retried while open (#692)
• fix(qwen): session-started is tried again while its connection is open (#692)
• fix(qwen): a managed reconnect always dials a token minted since its drop (#692)
• docs(spec): Qwen's liveness probe, its push-to-talk commit wait and their measurement (#692)
• fix(qwen): a connection that stays open and answers nothing reconnects as a stall (#692)
• fix(status): a screen reader can still read the reconnect's attempt and seconds (#692)
• fix(soniox): a reconnect opens a fresh TTS socket, so speech comes back (#692)
• test: pin the reconnect's frame, the throttled detail and the reissue role (#692)
• docs(soniox): the stall threshold cites its measurement (#692)
• fix(status): the reconnect line's live region stays steady while its seconds tick (#692)
• fix(session): a success read past the deadline's time is taken, and the attempt contract says so (#692)
• fix(qwen): a close during an attempt's handshake follows the live close policy (#692)
• fix(soniox): shared Both drops gap audio before the mixer, so none reaches the new socket (#692)
• fix(soniox): session-started is tried again while its connection is open (#692)
• docs(spec): Qwen reissues take their place before the mint (#692)
• docs: no internal names or channels in the reconnect spec and plan (#692)
• fix(soniox): a 429 reconnects unless it says the quota is used up (#692)
• docs(session): the flap limits and the reissued key's window in the spec (#692)
• fix(locales): screenContext keeps its escaped soft hyphen and zero-width space (#692)
• fix(soniox): a reissued key's window is counted on the client's clock from the answer, and a connect still going when it closes is given up (#692)
• fix(session): the restored n

Palabra AI: the whole voice library, and your own voice

• Choose any of Palabra's built-in voices for the translation the other side hears and the one you hear, not only Default Low and Default High. Each voice has a preview button. (#684)
• Create your own voice from a recording or an audio file, then pick it like any other voice. You can also delete voices you made. (#684)
• Works with both kinds of Palabra credentials: Platform API key and App Client ID/Secret. (#684)
• The voice list is kept while the app is open, so going back to the settings no longer reloads it. (#684)

Browser extension: tab capture on older installs

• On some installs that once ran an older version, starting a session failed every time with "Extension has not been invoked for the current page". Updating fixes this. (#683)
• When Chrome refuses to capture a tab, the message now tells you to click the Sokuji icon on the meeting tab, which is what grants the capture. (#683)

Soniox: new transcription model

• Real-time transcription now uses Soniox's stt-rt-v6 model. Soniox is retiring the previous model on December 15. (#685)

Full Changelog: https://github.com/kizuna-ai-lab/sokuji/compare/v0.47.0...v0.48.0

• chore(release): v0.48.0
• Merge pull request #685 from kizuna-ai-lab/chore/soniox-stt-v6
• Merge pull request #684 from kizuna-ai-lab/feat/palabra-voice-library
• test(palabra): a padded https sample url is used trimmed
• feat(soniox): transcribe with stt-rt-v6
• fix(palabra): preview the first sample at a usable https address
• fix(palabra): read a voice's nested fields defensively
• fix(palabra): a voice-list entry must be an object with a voice id
• fix(palabra): mask a credential of any length; a null list item is a typed error
• fix(palabra): address the review on #684
• fix(palabra): the remembered list stays right under overlap, failures and changes
• feat(palabra): remember the voice list for the app session
• fix(palabra): typed error for an empty answer; https only for samples and uploads
• fix(palabra): clone languages read as names; deleting a processing voice stays quiet; long recordings kept
• docs(palabra): verified with both credential types against the live API
• feat(palabra): voice library in Palabra's settings; CSP for voice uploads
• feat(palabra): create and delete voices, with processing state
• feat(palabra): voice picker with built-ins, own voices and previews
• feat(voice-library): optional sample language in the clone review step
• feat(palabra): voice source, clone languages, credential hook
• feat(palabra): voices REST client
• feat(palabra): voice ids are any string, stored with a name
• docs(palabra): voice library implementation plan
• docs(palabra): voice library design approved
• docs(palabra): filters are language and gender only (owner ruling)
• docs(palabra): voice library design (task #47)
• Merge pull request #683 from kizuna-ai-lab/fix/tab-capture-grant
• fix(extension): only the missing permission points at the icon
• fix(extension): tab capture survives a panel setting old builds left on
• Merge pull request #682 from kizuna-ai-lab/fix/tts-catalog-voice-type
• refactor(soniox): the TTS catalog declares its own VoiceOption
• Merge pull request #674 from kizuna-ai-lab/release/v0.47.0

Echo cancellation now removes Sokuji's own speech

Sokuji only echo cancellation removes Sokuji's translated speech again, so the translation no longer comes back through your microphone when you use speakers. On Linux, where this is the only mode, you get the benefit automatically. Translations also start playing sooner, and the delay no longer grows during a session. (#673)

Face-to-face: every translation is spoken

• If someone switches to the other person's language mid-conversation, the translation of those sentences is now read aloud, in the voice and on the output of its language. This applies to Soniox face-to-face and shared Both. (#671)
• These sentences have their own replay button and karaoke highlight. (#671)
• In face-to-face settings, the two speech rows are now named by language, so it is clear which device plays which translation. (#671)

Doubao AST 2.0

• A pause no longer triggers a false "Speech playback is degraded" notice. (#672)

Full Changelog: https://github.com/kizuna-ai-lab/sokuji/compare/v0.46.5...v0.47.0

• chore(release): v0.47.0
• Merge pull request #673 from kizuna-ai-lab/fix/output-context-sinks
• fix(audio): a rebuild a switch overtook reports no failure
• fix(audio): a switch that never answers cannot hold recovery off
• Merge pull request #672 from kizuna-ai-lab/fix/ast2-empty-clip
• fix(audio): a sink's rebuild cap counts a run, and a hung rebuild retries
• fix(audio): a sink's switch resolves only if the sink moved
• fix(volcengine_ast2): a page of an unknown Ogg stream version is the decoder's, not silence
• fix(volcengine_ast2): only whole header pages count as silence; anything past them goes to the decoder
• fix(audio): the outlets play through contexts the echo canceller hears
• test(audio): a sink that never ran sees a state change too
• feat(audio): a context sink recovers from a vanished device
• feat(audio): a context sink switches device on a new context
• feat(audio): an output sink on its own AudioContext
• docs(audio): plan the output context sinks
• docs(audio): design the output context sinks
• Merge pull request #671 from kizuna-ai-lab/feat/f2f-language-routing
• fix(volcengine_ast2): a clip of Opus header pages alone is silence, not a decode failure
• fix(soniox): a core's two voices number a row's audio entries together
• fix(playback): a replay's voice comes from an entry that will play, else the row's own
• fix(playback): forward the clip's voice from the app's playback port; replay from an entry that will play
• feat(contract): a row keeps the audio another leg's voice read for it — audio.voice
• chore(face-to-face): the tooltip names one choice, device and side together, as the list is
• chore(face-to-face): say the row's tooltip plainly — pick the device and the side the listener wears
• fix(face-to-face): an auto source keeps the generic row name; foreign chunks never enter the held text
• feat(face-to-face): name each speech row by where its language's translation plays
• feat(soniox): read a translation in the voice of its language, face-to-face and shared Both