v0.0.2

twinscribe-win64.zip (1.56 GB, sha256 07536e5a669fcbb701eb92fbb0c1554b88b291f1c917e0773796ef8181701b0f): a folder with twinscribe-app.exe (the window) and twinscribe.exe (the command line), every library, the CUDA runtime packages and the media decoder. No installation and no administrator rights.

The models are not inside the zip. On first start the window offers to fetch the models of the Standard level (about 2.4 GB) from the sources the catalogue names into a models folder beside the executables, recording every file's digest; twinscribe.exe fetch-models does the same from the command line. After that nothing in the program reaches the network.

Running it

1. Unzip anywhere and double-click twinscribe\twinscribe-app.exe. The executables are not signed, so Windows may warn that they come from an unknown publisher.
2. Press Download in the dialog that opens, or Later and press Get models when ready.
3. Drop recordings or folders into the window, choose a quality level, press Transcribe. Outputs go beside each recording: text, Word, subtitles, a review set and a run record. Review opens the verification screen on the marked spans.

Windows 10 or 11 on x86-64. An NVIDIA device is used for the checking engine when present; everything runs on the processor without one, more slowly. Linux and macOS run from source, as the README describes.

What this is not yet

Early. The design is measured and the bench reproduces its accuracy tables on public corpora, but speed is a property of the machine and the speaker labelling can split a voice it cannot separate. A transcript is a draft until it has been verified against the recording; the tool makes that verification targeted, it does not remove it.

Components

faster-whisper and CTranslate2 (MIT), Whisper weights (MIT), sherpa-onnx (Apache-2.0), NVIDIA Parakeet TDT 0.6B v2 and TitaNet (CC BY 4.0), Silero VAD (MIT), pyannote segmentation-3.0 (MIT), the CED audio tagger (Apache-2.0) with the AudioSet class labels (CC BY 4.0), PySide6 (LGPL-3.0). Attributions are in NOTICE. Licence: Apache-2.0.

Since 0.0.1

• The review reads in order. Words a listener keeps for a gap go into the line where they
  were said, marked as the listener's within it (braces in the text, italics in Word), instead
  of a line of their own that broke the sentence around it; and the second engine's echoes of
  the words either side of a gap no longer raise a mark or fill the words box, so the hint
  names only what the transcript lacks. The review list halves on two-party telephone calls.
• The speaker clustering measured. The threshold that decides when two voices are two
  speakers had been the library's own default and was the worst value measured on every set;
  the default is now 0.9, the value with the lowest error on the bench's telephone calls, and
  the Speakers box offers Auto for a long recording or a meeting, which keeps each voice
  together at 1.2. After the clustering, labels holding under three seconds of speech are
  folded into the nearest voice.
• The mark redrawn. The two facing panels on a dark tile with a thin blue edge, and the
  banner and logo painted from the same drawing.

TwinScribe 0.0.1

twinscribe-win64.zip (1.56 GB, sha256 56b5ee2292526a84e36ca4f28b16ab626dd2f764707b462398c7de8a4c5bf6e8): a folder with twinscribe-app.exe (the window) and twinscribe.exe (the command line), every library, the CUDA runtime packages and the media decoder. No installation and no administrator rights.

The models are not inside the zip. On first start the window offers to fetch the models of the Standard level (about 2.4 GB) from the sources the catalogue names into a models folder beside the executables, recording every file's digest; twinscribe.exe fetch-models does the same from the command line. After that nothing in the program reaches the network.

Running it

1. Unzip anywhere and double-click twinscribe\twinscribe-app.exe. The executables are not signed, so Windows may warn that they come from an unknown publisher.
2. Press Download in the dialog that opens, or Later and press Get models when ready.
3. Drop recordings or folders into the window, choose a quality level, press Transcribe. Outputs go beside each recording: text, Word, subtitles, a review set and a run record. Review opens the verification screen on the marked spans.

Windows 10 or 11 on x86-64. An NVIDIA device is used for the checking engine when present; everything runs on the processor without one, more slowly. Linux and macOS run from source, as the README describes.

What this is not yet

Early. The design is measured and the bench reproduces its accuracy tables on public corpora, but speed is a property of the machine and the speaker labelling can split a voice it cannot separate. A transcript is a draft until it has been verified against the recording; the tool makes that verification targeted, it does not remove it.

Components

faster-whisper and CTranslate2 (MIT), Whisper weights (MIT), sherpa-onnx (Apache-2.0), NVIDIA Parakeet TDT 0.6B v2 and TitaNet (CC BY 4.0), Silero VAD (MIT), pyannote segmentation-3.0 (MIT), the CED audio tagger (Apache-2.0) with the AudioSet class labels (CC BY 4.0), PySide6 (LGPL-3.0). Attributions are in NOTICE. Licence: Apache-2.0.

Revision

The zip was replaced on 2026-09-09 with a build from the same version that carries two fixes found on the first downloads: the first-start fetch reads an archive once for all of its files and reports extraction as its own stage, instead of showing the finished file at 100 per cent while each member was extracted with a fresh pass over the archive; and the speaker stage runs in a child process, so the window keeps answering while speakers are labelled. Later replacements the same day added an output device box to the player bar and a fourth quality level, Laptop, for machines without a graphics device: the checker decodes only the speech found where the published engine fell silent, which on telephone calls cuts the checker's time by two fifths while the review list keeps its quality.

Replaced again on 2026-09-14 with the review pass of 2026-09-10 and the corrections of a review of the application in use. The verification screen takes the words in a box that is always in view and writes every decision into the transcript and its outputs at once, with a box to say who spoke; words kept for a gap inside a line are placed at the gap with that line's speaker. Two recordings of one name sent to one output folder no longer overwrite each other; Escape in the models dialog during a download no longer closes the program; the buttons take the keyboard and Ctrl+C copies from the transcript; there is one review screen, at the player bar's volume; a recording can be transcribed again from its header; a speaker split in two is merged by renaming one to the other's name; a folder of hundreds of recordings loads in a second; a folder that refuses writes is named before a batch starts rather than hanging it. The README's Recent changes lists the rest. The digest above is the current zip's.

Replaced again on 2026-09-15 with the speaker clustering measured. The threshold that decides when two voices are two speakers had stayed at the library's own default since the first delivery, and on every set in the lab it was the worst value measured: six labels on a two-speaker telephone call, over a hundred on a four-speaker meeting, a one-reader chapter split in two. The default is now 0.9, the value with the lowest error on the bench's telephone calls, below the error with the count fixed at two; a second value, 1.2, keeps each voice together on a long recording or a meeting with several voices, and is offered as a Recording choice in the window and as --threshold on the command line. After the clustering, labels holding under three seconds of speech are folded into the nearest voice, which removes the fragments left at every threshold without hiding anyone who said a few sentences. The digest above is this zip's.