Sotto 1.5.1

Sotto 1.5.1

Changed

• The alerts that end a recording now say "conversation" instead of "lecture", because the recording can just as well be a meeting: "Complete this conversation?" and "End this conversation?" with matching text.
• The saved report's heading says "Conversation notes" (and the matching Korean heading) instead of "Lecture notes".

Sotto 1.5.0

Sotto 1.5.0

Added

• Four named modes: Dictation, Summary, Subtitle and Translation, in that order in the menu and in Settings, each with its own hotkey. Summary makes a summary with no live subtitles; Subtitle shows live subtitles and translation with a summary window.
• A new control row. Play/Pause, Stop and Picture sit on the left and Close on the right end. Stop ends the lecture and asks whether to save it; Close stops the session without saving. In Subtitle the row also has the translation and summary-window buttons. In Summary the row lives inside the summary window, and pausing stops the sound input.
• Both as a sound source. Choose System sound, Microphone or Both in the menu or in App settings. Both mixes your microphone and what your Mac plays into one transcript, and keeps your dictated words out of it.
• Settings in four tabs: Language, AI, Hotkeys and App settings.
• First and Second can be the same language. Dictation and Summary work as usual, the report shows one language with no First/Second toggle, and Subtitle and Translation show a short message instead of starting. A new install starts with English and English.
• The summary window makes room for a picture only where a picture was taken.

Changed

• In the summary window only the summary block that holds a picture is indented beside its thumbnail. Blocks without a picture use the full width, and deleting the picture gives the block its width back.
• The Summary window opens with the same size and place as the Subtitle summary window.
• The Subtitle control panel no longer has a Settings button; open Settings from the menu.
• A Translation request made right after launch no longer fails on the first try.
• A new session takes the language pair you have just chosen, so a same-language session after a language change is no longer refused.

Known limits

• With Both, when your microphone and the system sound overlap for a long time, only one voice at a time is recognized. Use headphones so the microphone does not hear the system sound again.
• Clicking the summary window makes Sotto the active app.

Sotto 1.4.14

• A sentence in your own ~/.codex/AGENTS.md that only names a model no longer raises the note that an AGENTS.md pins a model. For that file, only a model or effort setting such as model = "..." or model: ... counts. Sotto's own session AGENTS.md is still checked for every mention.
• The menu text for a capture that hears nothing now reads "Nothing is playing, or System Audio Recording is off for Sotto".

Known limits

• Sotto cannot tell a missing System Audio Recording permission from silence, so the menu text names both.

Sotto 1.4.13

• Pressing Play no longer waits for the dictation model to finish its first setup. After an update, the first Play could wait for minutes because the dictation (Whisper) model was compiling and the speech recognition model that capture needs had to wait behind it. The speech recognition model is now prepared first, and the dictation model starts after it. With the compile caches cleared, Play now starts in 7 to 13 seconds; the same test on 1.4.12 took 128 to 208 seconds.
• The dictation model still takes about two minutes to set up the first time, and it now loads right after the speech recognition model.

Known limits

• On a fresh install, the speech recognition model is downloaded at the first Play, so that first Play can still wait while the dictation model finishes its first setup.

Sotto 1.4.12

• Microphone capture handles a device change more reliably: Stop now waits for a restart that is already running instead of racing it, and a device change that arrives while recording is still starting is no longer skipped. Checked with AirPods: two short dictations were captured and pasted.
• The command-line helper's settings error now says Sotto instead of the internal name tapin.

Known limits

• On some long Korean and English recordings the error rate can be up to about 0.8 percentage points higher than with the single pass.
• Long speech that mixes Korean and English is still processed in one pass, so it is not faster.

Sotto 1.4.11

Sotto 1.4.11

• The setup and uninstall screens now say Sotto instead of the internal name tapin: the command-link and skill step, the Install skills confirmation, the Settings help text and the Uninstall list. Real folder and file names that you may need to find, such as Documents/Tapin, are unchanged.

Known limits

• On some long Korean and English recordings the error rate can be up to about 0.8 percentage points higher than with the single pass.
• Long speech that mixes Korean and English is still processed in one pass, so it is not faster.

Sotto 1.4.10

Sotto 1.4.10

• Whisper dictation now pastes sooner after you release the key. On Korean speech of 40 to 153 seconds it took 1.1 to 1.8 seconds, against 4.6 to 8.9 seconds before.
• Bluetooth microphones such as AirPods record again when macOS switches the headset input.

Known limits

• On some long Korean and English recordings the error rate can be up to about 0.8 percentage points higher than with the single pass.
• Long speech that mixes Korean and English is still processed in one pass, so it is not faster.

Sotto 1.4.9

Sotto 1.4.9

Fixed

• Translation now waits through a Codex rate limit instead of restarting. When Codex answers with HTTP 429 and says it is retrying, the in-app translator keeps the same Codex process and waits for the answer. It used to give up after 15 seconds and start a new process again and again, which met the same limit each time. Lines that waited are translated once the limit clears.
• The viewer says why a line is waiting. While Codex is retrying, the viewer shows "Translation is waiting: Codex connection lost. Codex is retrying the request (waiting since HH:MM)", and the message clears after the first translation arrives.
• A rate limit during the translator's warm-up no longer ends the session in agent mode. The session stays in the app translator and every line is translated.
• The translator's Codex process no longer gets a setting its Codex does not know. It was started with features.memory=false; Codex 0.163.0 names that feature memories. The process now starts with features.hooks=false and features.memories=false, and a Codex without those features gets no such arguments.

Known limits

• A translator that never answers is still replaced up to three times and then handed over to the translator agents, as before.

Sotto 1.4.8

Sotto 1.4.8

Fixed

• A long capture could crash the app. After roughly an hour and a half of continuous speech the app could run out of stack and quit. The cause was found and removed: the callbacks that carry live words, lines and speaker marks no longer pile up call frames each time they are read.
• tapin start no longer gives up after 10 seconds on a slow launch. It now waits up to two minutes for the app to open and for capture to begin, so a slow start is not reported as "Sotto app is not responding".
• Codex model lists show only models Codex really runs. Sotto asks Codex which of the listed models work and hides the ones it rejects. A stored model that stopped working is replaced by the translation model for the Summary, and the stored choice is kept.
• The Summary never runs on gpt-6-astra, even when that is the chosen translation model.
• A model that Codex refuses while an agent is running is no longer used for the Summary and Screen roles. The Report keeps it, because the report runs through codex exec.
• The report call always names its model, so it can no longer fall back to a model Codex picks on its own.

Added

• Summary effort in Settings, Lecture mode (Low, Medium or High). The choice is saved, survives a restart and is used for the summary.
• The viewer shows the model and effort that ran, for example Summary gpt-6-luna medium.

Known limits

• When Codex cannot be reached at launch, the model lists stay as they were; nothing is changed or hidden.
• Right after launch the model caption can show the wrong model for up to 5 seconds.

Sotto 1.4.7

Sotto 1.4.7

Fixed

• The enter word at the end of a sentence is pressed more reliably. Dictation now keeps listening for 0.3 seconds after you release the key, so the last word is no longer cut off at the moment of release.
• The first word of a sentence is cut less often. Dictation turns the microphone on before its other start-up work, and the recording indicator appears once sound is arriving.

Changed

• Text appears about 0.3 seconds later after you release the key, because of the extra listening time.
• With the Nemotron model, a short stretch of quiet is fed to the recogniser before your speech. Whisper is not changed.

Known limits

• A word spoken right at the key press or in the last moment before release can still be missed.