Verbalis 0.7.43

• Unclear speaker marks with the moment to listen to: two names on one voice were a global notice
  («… klingen für Verbalis gleich») without saying where – even for one stray mark out of six, e.g. a click while
  the other person was still talking. Now only unclear cases count (the second name has 2 marks or a quarter of
  them, or no speaker of its own). They are stored with the recording and listed in its speaker box: «Bei 1:16:03
  war «Lara» markiert, Verbalis hört dort «Maurice».» with ▶ to listen from 5 seconds before; × hides them.
• Speaker diarization several times faster: the segmentation looked at every moment ten times (10-second
  windows, 1 second apart); now half a window apart. An 81-minute call needed 34 minutes of diarization after the
  end (partly next to another Teams call); with the new step 3.6 minutes, the same speaker for 99.6 % of the talk
  time. More than 4 threads were slower.

Windows: download and run Verbalis-0.7.43-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.43-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.42

• Transcribe while recording (setting «Schon während der Aufnahme transkribieren», on by default): the
  growing recording is transcribed in the same ~2-minute pieces as afterwards; after the end only
  the last minutes and the speaker diarization are left. The recording view shows how far it is. On Windows
  Verbalis runs «below normal» meanwhile – Teams and other programs come first – and the recorder threads keep
  normal priority. If it can't keep up it falls behind; nothing is lost. The pieces are handed over to the normal
  transcription (also after a crash), which still starts by hand when «Sofort automatisch transkribieren» is off.
  «Anhalten» next to the status stops it for this recording; the finished pieces stay. First real test: a
  51-minute Teams call, transcript ready 2 minutes after the end (before: 20–35 minutes); about 0.4 s of computing
  per second of audio while Teams ran.
• «0 Personen» leaves a track out: e.g. the system audio recorded by mistake at a meeting on site. The
  transcript header says «Nicht transkribiert: Systemton». Before, only 1–30 were accepted.
• Participants decide the speaker diarization when no number is entered (also with automatic
  transcription): two or more names under «Am Mikrofon» = a meeting on site, told apart into that many people,
  without switching on «Mikrofon: verschiedene Personen unterscheiden»; one name = nobody to tell apart.
• Your own name at the microphone: «Wer spricht gerade?» shows you under «Am Mikrofon» with ✓ – no need to
  mark yourself. Typing your own name as a new name marks the microphone, no longer the system audio (where it
  named the other person's voice too and counted you as a second person in the call). The status line during the
  recording no longer adds «nach dem Ende fehlen nur noch wenige Minuten». The introduction mentions the marks.

Windows: download and run Verbalis-0.7.42-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.42-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.40

• Speaker box also without speaker diarization: when the voices of a track were not told apart («1 Person»,
  or switched off), the box only offered «Anzahl Personen» and «Gegenüber» could not be renamed. Now every track
  has a speaker with name field, list of participants and sample – also in older transcripts.
• One participant names the speaker: with exactly one name entered for a track under «Teilnehmende», its
  only speaker gets that name automatically (also when entered later). It can still be changed in the speaker box.
• More room for the transcript: the speaker box is grouped by track with «Personen [n] Neu zuordnen» next to the
  speakers, and closed to one line with the names once every speaker has a name (click to open). Its explanation
  is behind «?». Title, project and date share one line; the project is no longer shown twice.

Windows: download and run Verbalis-0.7.40-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.40-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.39

• Fix (Windows): «GitHub ist nicht erreichbar … CERTIFICATE_VERIFY_FAILED» on some PCs – Python only knew
  the root certificates already in the Windows store (Windows loads them on demand; company roots for TLS
  inspection were missing too). Windows now checks HTTPS certificates itself (truststore), like Edge does –
  for the update check and all downloads (models, dictionary).

Windows: download and run Verbalis-0.7.39-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.39-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.38

• Fix: «Anzahl Personen auf dieser Spur» was missing when no speakers had been told apart – e.g. the
  test recording of the introduction («1 Person» per track) or with «Sprechertrennung» off. Now the speaker
  box always offers the number of people and «Neu zuordnen».
• Fix (Mac, MLX): the speed test failed with «[load_npz] Input must be a zip file» – the converted model
  had a file name mlx-whisper 0.4.3 doesn't look for. An already converted model is renamed, not loaded again.
• Linux: the window starts with Qt right away when GTK's Python bindings are missing (no more error
  messages in the terminal); the diagnosis shows the graphics card.
• Speed test: uses the newest recording of at least one minute (a short, silent one gave no result).
• Measured: on a Mac M1 the graphics chip (MLX) is only 1.2× faster than the CPU, on an Nvidia T500 1.3× –
  MLX stays a test for now, CUDA stays on.

Windows: download and run Verbalis-0.7.38-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.38-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.37

• Downloads under fixed names: every release also has Verbalis-Windows.exe and Verbalis-Mac.dmg, so
  links like …/releases/latest/download/Verbalis-Windows.exe always lead to the newest version. frehner.tech
  showed 0.7.30 while 0.7.36 was out, because GitHub answers only 60 version queries per hour and network.
  The app's updates still use the files with the version in their name.

Windows: download and run Verbalis-0.7.37-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.37-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.36

• Experimental tests in Einstellungen → Diagnose, so new systems can be tested without the command
  line; the results go into «Diagnose kopieren» and «Problem melden» (numbers only, never text):
  • Graphics chip status and a button to load the add-on once: the Nvidia libraries (1.3 GB,
    Windows/Linux) or MLX for Macs with M chip (90 MB) – downloaded from PyPI and checked, so the
    installers stay small.
  • «Geschwindigkeit messen»: transcribes 2 minutes of the newest recording on the CPU and on the
    graphics chip and shows the time and how far the texts agree.
  • «Meldung testen» (also notes when the notification is clicked) and «Anruf-Erkennung prüfen»
    (which programs use the microphone right now).
• With the Nvidia libraries loaded, the app transcribes on the graphics card (fall-back to the CPU on any
  error). On the Mac, MLX is only measured for now – the app keeps transcribing on the CPU.

Windows: download and run Verbalis-0.7.36-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.36-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.35

• Fix: a recording no longer ends when an audio device disappears. A headset switched off or
  unplugged, or a Bluetooth headset leaving hands-free mode at the end of a Teams call, ended the
  recording with «Gerätefehler: Error 0x100000001». Now the track reconnects – to the same device once it
  is back, otherwise to the default device – and the other track records on; the gap stays silent. The app
  says that the device is gone and the recording continues. meta.json notes reconnects and devices used.
  The error message, if a device can't be reconnected, names the real cause.
• Nvidia graphics card (experimental, Windows and Linux): verbalis gpu shows whether Verbalis can use
  it, verbalis gpu --install downloads NVIDIA's cuBLAS and cuDNN once (about 1.3 GB, checked). Then the
  app transcribes on the graphics card, with an automatic fall-back to the CPU on any GPU error.
  With the installed app on Windows: %LOCALAPPDATA%\Programs\Verbalis\verbalis-cli.exe gpu --install.
• Mac with M chip (experimental, development version): verbalis compare … mlx:MODEL measures
  transcription on the graphics chip with MLX against today's CPU transcription.

Windows: download and run Verbalis-0.7.35-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.35-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.34

• Linux (experimental): Verbalis runs from the repository or with pipx install straight from GitHub
  (README «Linux (experimental)»); there is no installer and no automatic update yet. The app says so next
  to the audio devices. New on Linux: call detection (programs using the microphone, via pactl),
  notifications with a click that opens Verbalis (notify-send), and a second start is refused like on
  Windows. Optional Qt window: verbalis[qt]. Not tested on real hardware yet – please report problems
  with «Problem melden».
• Fix (macOS): Verbalis could crash on quitting while an audio device was still open (e.g. after waking
  the Mac). Everything was saved already; now the app ends cleanly.

Windows: download and run Verbalis-0.7.34-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.34-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.

Verbalis 0.7.33

• Speaker box: one list of participants. In the app the name field already lists the participants of
  the track, so «Teilnehmende ▾» from 0.7.32 was a second list – it is gone. A name Verbalis gave
  («Gegenüber 2») moves to the placeholder while the field has the focus, so the list shows every
  participant; clicking away without choosing keeps the old name.
• Fix: «Anzahl Personen auf dieser Spur ()» showed empty brackets – now «(Systemton)» or «(Mikrofon)».

Windows: download and run Verbalis-0.7.33-setup.exe. An update is installed the same way – settings, correction list, models and recordings are kept.

Mac (Apple Silicon): open Verbalis-0.7.33-macos-arm64.dmg and drag Verbalis onto «Programme» (Applications) – or just open Verbalis: it offers to move itself there (into the personal Applications folder without an administrator account). The ZIP is for the automatic updates. On the first start macOS blocks the app because it is not notarized yet: close the message, then System Settings → Privacy & Security → «Open Anyway» (once; updates keep it). From macOS 14.2 Verbalis records the others' audio itself (allow the «Audio Recording» permission); only older Macs need BlackHole, see the README.