amanu v0.4.26

Amanu 0.4.26

Transcription

• Transcription no longer gets stuck after acoustic echo cancellation on macOS Tahoe. Amanu now lets Core Audio choose an AAC bitrate supported by the resulting 16 kHz tracks.
• Live transcription hides an echoed microphone phrase even when the cleaner call audio was split into several blocks or arrived through the decoder with a different delay.
• ElevenLabs Scribe v2 is available as a transcription engine. Amanu processes the microphone and call audio separately, retaining distinct speakers on each side.
• AssemblyAI transcripts now group word-sized fragments into readable paragraphs while preserving speaker changes and pauses. Existing transcripts can be reformatted from saved data without sending the audio again.

Automatic recording

• Calls in DION now start automatic recording. Amanu recognises both the main DION app and its helper processes.

Recordings

• Open a transcript directly from the recordings window in your preferred Markdown app. If no app is associated with Markdown, Amanu opens it in TextEdit.
• Open the selected recording folder directly from the same window.

Compatibility

• Universal binary for Apple Silicon and Intel Macs running macOS 14.2 or later. Local transcription requires Apple Silicon; Intel requires a cloud transcription key and has not been tested on physical hardware.

amanu v0.4.25

Amanu 0.4.25

Local transcription models

• Local transcription models are once again shown as separate cards, making the choice between Parakeet, Whisper, and GigaAM clear at a glance.
• Every card shows the expected download size before installation and the actual disk usage afterwards.
• Download progress and failures now stay attached to the model they belong to.
• GigaAM is described as an alternative for Russian rather than the default recommendation.

Compatibility

• Universal binary for Apple Silicon and Intel Macs running macOS 14.2 or later. Local transcription requires Apple Silicon; Intel requires a cloud transcription key and has not been tested on physical hardware.

amanu v0.4.24

Amanu 0.4.24

Homebrew installation

• Amanu can be installed with brew install --cask gsamat/tap/amanu.
• A Homebrew-managed amanu command now takes precedence over Amanu's private ~/.local/bin/amanu link, avoiding a duplicate command after installation.

Compatibility

• Universal binary for Apple Silicon and Intel Macs running macOS 14.2 or later. Local transcription requires Apple Silicon; Intel requires a cloud transcription key and has not been tested on physical hardware.

amanu v0.4.23

Amanu 0.4.23

Local transcription

• Whisper large-v3-turbo and GigaAM v3 can now be downloaded and used directly in Amanu alongside Parakeet.
• Model downloads show progress and the exact disk space used by both Whisper and GigaAM. GigaAM is tuned for Russian speech.
• A failed recording can be transcribed again from the recordings list, with a choice of available transcription engines.

Import existing recordings

• Drop audio or video files onto the recordings window, or choose Import from the File menu or menu bar.
• Imports show queue and transcription progress, can be cancelled, and appear alongside recordings when complete.

Summaries

• Ollama and custom OpenAI-compatible endpoints can now generate summaries without sending recordings to Amanu's default provider.
• Advanced settings include an editable summary template, prefilled with Amanu's default instructions.

Live transcript

• A lightweight echo filter reduces duplicated You/Them lines when recording without headphones.

Compatibility

• Universal binary for Apple Silicon and Intel Macs running macOS 14.2 or later. Local transcription requires Apple Silicon; Intel requires a cloud transcription key and has not been tested on physical hardware.

amanu v0.4.22

Amanu 0.4.22

The full-size feather promised in the preview

• The macOS 26 Dock icon now uses the large hollow feather shown in the original preview.
• Icon Composer no longer shrinks or fills the feather: its transparent outline is scaled to compensate for the system safe area and sits slightly above center.
• Default, Dark and tinted appearances keep their system-native treatment, while recording status remains in the Dock badge and menu bar.

Compatibility

• Universal binary for Apple Silicon and Intel Macs running macOS 14.2 or later. Local transcription requires Apple Silicon; Intel requires a cloud transcription key and has not been tested on physical hardware.

amanu v0.4.21

Amanu 0.4.21

The right interface language

• Amanu now follows the Mac's primary interface language. A Mac set to Russian opens Amanu in Russian; every other language falls back to English instead of picking Russian from farther down the preferred-language list.
• An explicitly selected English or Russian interface still overrides the Mac and takes effect after Amanu restarts.

Compatibility

• Universal binary for Apple Silicon and Intel Macs running macOS 14.2 or later. Local transcription requires Apple Silicon; Intel requires a cloud transcription key and has not been tested on physical hardware.

amanu v0.4.20

Amanu 0.4.20

A readable Dock icon

• The Dock icon now keeps its recognizable slate background in every recording state instead of turning into a bare dark feather that disappears against a dark Dock.
• Recording status stays where macOS expects it: the system Dock badge shows elapsed time, while the menu bar feather remains red during recording and orange while paused.
• On macOS 26 and later, Amanu includes layered Default, Dark and tinted icon appearances. Sonoma and Sequoia keep the existing hand-tuned icon.

Sonoma compatibility

• Lowered the minimum system requirement from macOS 15 to macOS 14.2, including the main executable and the bundled echo-suppression library.
• Added a release check that inspects every embedded Mach-O slice and nested bundle, preventing dependencies from silently raising the minimum macOS version again.
• macOS 14.2 is the oldest system with the Core Audio process-tap API Amanu needs to record the other side of a meeting. Automatic per-process detection is limited before macOS 14.4; recordings can still be started manually.

Compatibility

• Universal binary for Apple Silicon and Intel. Local transcription requires Apple Silicon; Intel requires a cloud transcription key and has not been tested on physical hardware.
• The macOS 14.2 deployment target is verified by compilation and release-artifact inspection. Runtime testing on physical macOS 14 hardware is still pending.

amanu v0.4.19

Amanu 0.4.19

Automatic recording

• Added automatic call detection for Yandex Telemost, eXpress and Kontur Talk on macOS.
• Slack Huddles remain supported through the existing Slack application detection.
• Calls in browser versions of these services remain covered by the existing browser detection.

Compatibility

• Universal binary for Apple Silicon and Intel Macs running macOS 15 or later. Intel requires a cloud transcription key and has not been tested on physical hardware.
• Audio capture, transcription and echo-suppression behavior are unchanged in this release.

amanu v0.4.18

Amanu 0.4.18

Analytics delivery

• Fixed analytics events being rejected because their timestamps included fractional seconds. Existing queued events are converted automatically.
• Events are now removed from the local queue only after validating the server's acknowledgement. Failed entries remain queued, and successful entries in a partially accepted batch are not resent.
• Large backlogs are split into requests within the server's limit. The existing opt-out, seven-day retention and 500-event queue limit remain in effect.
• Events already discarded by earlier versions cannot be recovered. Statistics begin accumulating as updated installations send events.

Automatic recording

• Expanded default call detection to WhatsApp, Signal, Telegram Desktop, Viber, LINE, WeChat and current Webex.
• Added Opera, Vivaldi, Chromium, Yandex Browser, DuckDuckGo, Orion, Zen and Dia for calls in browser tabs. Existing custom application lists remain unchanged.

Compatibility

• Release packaging now discovers SwiftPM's output directory, including the new Xcode 27 build engine.
• Universal binary for Apple Silicon and Intel Macs running macOS 15 or later. Intel requires a cloud transcription key and has not been tested on physical hardware.
• Audio capture and echo-suppression behavior are unchanged in this release.

amanu v0.4.17

Amanu 0.4.17

This release removes acoustic echo after a meeting without changing the live
call or the saved recording.

What changed

• Amanu now cleans a temporary microphone copy with LocalVQE before
  transcription. The original microphone and system tracks remain untouched,
  so disputed passages can still be checked against the source recording.
  Processing never opens an output device, changes playback volume or reduces
  what you hear during the call.
• The model removes playback captured by the microphone without applying noise
  gating, denoising or automatic gain control. It continues cancellation for
  one second after playback stops to cover a delayed acoustic tail. During
  sustained reference silence, Amanu uses the original microphone samples
  exactly. Large or abrupt changes in playback delay can still leave some echo.
• A controlled prototype evaluation covered nine complete recordings. Matched
  duplicate words remaining after filtering fell from 340 to 16 with
  AssemblyAI and from 532 to 39 with Parakeet. Source hashes, durations and the
  system track remained unchanged in all nine runs. Some quiet words still
  changed or disappeared, so these results do not imply perfect transcription.
• Three complete recordings also passed through the actual application paths.
  Matched duplicate words fell from 154 to 17 with AssemblyAI and from 194 to
  33 with Parakeet. Original-file hashes and decoded track lengths matched.
  These results cover the production decoder and codecs and are separate
  from the nine-recording prototype figures above.
• Provider utterances entirely beyond the real audio duration are now
  discarded, while a genuine utterance crossing the boundary is clipped.
  Cleaned audio uses a processor-specific cache and is removed after
  transcription. Processing failures leave the source recording intact.

The bundled LocalVQE model and universal library are pinned, SHA-verified and
signed inside the application. The library supports Apple Silicon and Intel.
Local Parakeet transcription still requires Apple Silicon; Intel uses a cloud
provider. Physical Intel hardware and macOS 15 have not been tested.