NowCaption 1.2.2

Live captions for any audio on your Linux desktop, recognised entirely on your own computer. This release brings NowCaption's new speech model, v4: slightly fewer mistakes, with less than half the CPU time.

Install

With automatic updates (recommended):

flatpak install --user --from https://techlm77.github.io/NowCaption/nowcaption.flatpakref

Installed that way already? flatpak update gets 1.2.2.

Or download a file below — x86_64 for most PCs, aarch64 for 64-bit ARM — then:

flatpak install --user NowCaption-1.2.2-x86_64.flatpak

Or the AppImage, for Ubuntu 22.04, Debian 12 and newer — no installation needed:

chmod +x NowCaption-1.2.2-x86_64.AppImage
./NowCaption-1.2.2-x86_64.AppImage

Gear Lever or AppImageUpdate can keep it up to date.

What's new in 1.2.2

• Speech model v4. NowCaption's own model, trained further on transcripts that keep the first and last words of each recording, and rebuilt so it doesn't repeat the same work at every step:
  • Fewer wrong words when captioning continuously: 13.3% → 13.0% across meetings, calls, podcasts, speeches and noisy recordings, and 13.2% → 12.9% with a fresh stream per clip. On 3,200 other clips never used to choose the model: 12.4% → 12.3%.
  • Accents: 13.8% → 13.5% over 2,992 clips in 20 regions. Hard cases (gaming, names, numbers, live streams): 9.1% → 8.8%.
  • Less than half the CPU time. The speech model now needs about 7% of one CPU core for non-stop speech (was about 18%), and NowCaption on its own 0.073 CPU seconds per second of audio (was 0.161).
  • Words show as soon as before: 0.204 s after they are spoken at the median (was 0.210 s).
  • A smaller download (111 MB, was 127 MB) and about 20 MB less memory.
  • Still no words over background noise, and fewer over instrumental music (27 words in 40 clips, was 35).
• Known limits:
  • Very occasionally v4 repeats a word several times ("sir sir sir sir sir sir sird" for an unusual name): in 4 of 9,762 test clips, where v3 did it in 1.
  • On LibriSpeech's harder audiobook speakers it is slightly worse in one of our tests (6.4% → 6.7% words wrong), though not in the others.
  • Some groups are slightly worse, though not beyond chance, such as Northern English accents (6.0% → 6.5%), and Common Voice's short sentences straight after a pause show 17 ms later.
  • The benchmark page lists every limitation with its numbers.
• Benchmarks you can check. The benchmark page has every result with its 95% range, the hardware and how it was measured, and nowcaption-1.2.2-benchmark-results.tar.xz below has the per-clip results behind the tables. The reference transcripts are left out, as some test sets' licences don't allow passing them on.

Everything else works as in 1.2.1, and your settings and history are kept. The .zip and .tar.gz below contain the source code. The speech model is downloaded separately from the models-v4 release; its training data and licences, including the transcripts from NVIDIA's Parakeet TDT 0.6B v2, are listed in MODEL-LICENSES.md. To go back to the v3 model, install 1.2.1's .flatpak or AppImage from the 1.2.1 release.

SHA-256 of the Flatpaks, AppImages and benchmark results:

• NowCaption-1.2.2-x86_64.flatpak: 26602d4c388d32d18749be5a0d63a763dcdcf4ab0e870c69cea3d7e25cf056cf
• NowCaption-1.2.2-aarch64.flatpak: e3295779547d3a505f97436df32c0613c8c70cc947a61755f655f1ac2ab90a2a
• NowCaption-1.2.2-x86_64.AppImage: 72c97f4bc59392bb2e39e8fc09d2be7c17b0b6ca87945022dc9cc83aae0fbce7
• NowCaption-1.2.2-aarch64.AppImage: 600d6dbb086ad27056cdb8eb457ec270118a7b1b34ceacc8f3fb5b9c0f9d5ecf
• nowcaption-1.2.2-benchmark-results.tar.xz: 491a7a0353fd173e75ef651abca00c1ed331b81a5ca5b335174f7018ad11efe1