Skip to content

Releases: mbn-code/whisperly-download

Danish model: Edda v0.2 (q5_0) for Whisperly

Choose a tag to compare

@mbn-code mbn-code released this 07 Oct 01:54
1cf558c

The Danish speech model Whisperly for Windows offers, from version 0.6.3, as Edda v0.2 (Danish, experimental) in Settings → Models. You do not need to download it from here; the app does that, and checks the file against the SHA-256 below before using it.

It is opt-in. For Danish, Whisperly recommends Large v3 Turbo: in testing, this model sometimes dropped a sentence from recordings holding several, which Turbo did not.

What it is

Edda v0.2, by the Danish Foundation Models project: a full fine-tune of OpenAI's whisper-large-v3-turbo on Danish speech, released under the Apache License 2.0. Original weights: danish-foundation-models/edda-v0.2, revision 98ec5eee028b111a19733890fe32ea65267b662b.

This file is a modified form of those weights: converted from the Hugging Face checkpoint to whisper.cpp's ggml format and quantised to q5_0. Nothing else was changed. The licence is attached as LICENSE-Apache-2.0.txt.

The file

  • ggml-edda-v0.2-q5_0.bin — 574,041,195 bytes
  • SHA-256 9bdfd79be6fad58b0a6f273526fe7e3fdc856bea7b31c19d032c8352d0bf0197

It is byte-identical to an independent q5_0 conversion of the same revision published at Angel-Freak/edda-v0.2-ggml.

Using it with whisper.cpp yourself

Edda was trained without timestamp tokens. Start whisper-server with -nt (--no-timestamps), or it produces garbage, and with -mc 0 for audio longer than 30 seconds, or it can loop.

On 80 single-sentence Danish FLEURS validation clips, through the whisper.cpp server Whisperly ships, it had a word error rate of 6.5%, against 14.7% for the stock ggml-large-v3-turbo.

Whisperly 0.6.3 for Windows

Pre-release

Choose a tag to compare

@mbn-code mbn-code released this 07 Oct 11:28
1cf558c

Danish, heard properly — and a word-splitting bug fixed for every language.

Danish

0.6.2 taught Whisperly to write Danish. This release is about hearing it. Measured on 80 clips of real Danish speech, Large v3 Turbo's word error rate went from 20.0% to 14.7% with the fixes below, and Turbo is now free. A free user could not install it before and was on Small, at 38.7%.

  • Large v3 Turbo is free, and it is what Whisperly recommends for Danish. Medium (English) and the history picker remain Pro.
  • An experimental Danish model. Edda v0.2, a Danish fine-tune of Large v3 Turbo by the Danish Foundation Models project, is in Settings → Models as an opt-in. On single sentences it makes 6.5% word errors against Turbo's 14.7%, in a 548 MB download. It is not the recommendation yet: in testing it sometimes dropped a sentence from longer recordings that Turbo kept. It only understands Danish, and the app says so if another language is set.
  • Danish Windows starts in Danish. A new install on Windows set to Danish starts on Danish and on Large v3 Turbo; the setup wizard shows the dictation language and can change it. An existing install keeps its settings — choose Danish and the model in Settings.
  • Capitals stay where Danish puts them: "den 3. maj", "ca. ti", "kl. halv tre" no longer get a capital after the full stop.
  • Lists stay lists: "punkt 1, 2, 3 og 4" is no longer glued into a decimal.
  • Quotes by voice: "citationstegn", "start citat", "slut citat".

Every language

  • Words are no longer cut in half. A word that fell on a boundary inside Whisper's output was pasted in two pieces — "Af rikas". This alone took Danish word errors on Turbo from 20.0% to 14.7%.
  • Auto-detect detects. It used to transcribe everything as English. It now detects the language, and auto-detected Danish gets the Danish punctuation words and number formatting.
  • Spoken punctuation no longer doubles Whisper's own ("fint..", "Hej,,"), and the word after a spoken full stop starts a sentence.
  • Removing filler words leaves no stray commas.
  • Dictation experience: microphone and engine readiness checked before recording, a warning when the microphone is silent, five-minute takes with a countdown, and a take that fails to transcribe is kept so Transcribe File can retry it.

Install

Download WhisperlySetup-v0.6.3.exe and run it — over 0.6.2 is fine, settings and models are kept. It is unsigned, so SmartScreen will show "Windows protected your PC": choose More info → Run anyway. Per-user install to %LOCALAPPDATA%\Programs\Whisperly, no administrator prompt. Whisperly-win-x64.zip is the same build unpacked.

Whisperly 1.0.1 for Mac

Choose a tag to compare

@mbn-code mbn-code released this 01 Oct 23:35

The first stable Mac release, now packaged as a disk image: open it and drag Whisperly into Applications.

In 1.0

  • A warm engine. whisper-server keeps the model in memory between dictations. If it ever fails, dictation falls back to whisper-cli.
  • Transcribe File… from the menu bar, or drop a recording on the icon: wav, mp3, m4a, aac, flac, aiff and caf.
  • History picker on ⌃⌥⌘⇧V. Type to filter past dictations; Enter pastes one at your cursor.
  • Recent transcripts in the menu bar, one click to copy.
  • Esc cancels while recording or transcribing.
  • Models are checked against their SHA-256 before use.
  • A command line inside the app: transcribe, history and diagnostic.

Whisperly 0.6.2 for Windows

Pre-release

Choose a tag to compare

@mbn-code mbn-code released this 01 Oct 23:34

Danish end to end, and the release that makes a GPU optional.

Danish

Whisperly could always transcribe Danish — it is one of the ninety-nine languages Whisper hears — but everything downstream of the decoder was written for English. This is the language handled properly: word boundaries that know "å" is a letter, Danish punctuation by voice, Danish hesitation sounds, Danish numbers, and names that spell themselves. The initial-prompt priming was measured on 24 clips of real speech, made no difference, and is therefore off.

Transcription

  • CPU machines actually work. The bundled whisper runtime moves to the v2 Vulkan build with dynamic backend loading and per-ISA CPU variants. A machine with no usable Vulkan loader now genuinely falls back to CPU instead of dying with STATUS_DLL_NOT_FOUND.
  • Names come out spelled the way you spell them, and model size is called out where it matters.

Prompt mode

  • Prompt mode can be shaped by a cloud model, on the login you already have — no extra key.
  • A failed cloud turn now names its own failure instead of failing silently.

Reliability

  • The hotkey can no longer die. A failed mic start no longer kills it for the session, and a hook dropped by Windows re-registers itself.
  • Pastes stopped lying. A paste that UIPI silently discarded no longer reports success.
  • Setup got honest about downloads that stall or half-extract.
  • The shell grew manners — crashes get a voice, second launches surface Settings, upgrades close cleanly.
  • Updates exist. A once-a-day check against the website manifest, pinned host, redirects refused, with an off-toggle in Settings → About.

Install

Download WhisperlySetup-v0.6.2.exe and run it. It is unsigned, so SmartScreen will show "Windows protected your PC" — choose More info → Run anyway. Per-user install to %LOCALAPPDATA%\Programs\Whisperly, no administrator prompt. Whisperly-win-x64.zip is the same build unpacked, if you would rather not install.