Skip to content

Version 1.3.46 - #790

Merged
maboa merged 1 commit into
mainfrom
release/1.3.46
Nov 10, 2026
Merged

maboa merged 1 commit into
mainfrom
release/1.3.46

Conversation

@maboa

@maboa maboa commented Nov 10, 2026

Copy link
Copy Markdown
Member

Release 1.3.46. Fixes #760, fixes #785, fixes #786, fixes #788.

Changed

  • Parakeet TDT 0.6B Ultra beside v3. The Parakeet (Local) tab gets a Model select with NVIDIA's Parakeet TDT 0.6B v3 and Moondream's Parakeet TDT 0.6B Ultra, a further-trained v3 with the same architecture and tokenizer. The choice is remembered, each model's files are listed and removed on their own under Settings ▸ Downloaded models, and the project's provenance names the model that ran. On the intro clip Ultra's CPU output matched its GPU output word for word, where v3 on the CPU dropped two words. (Add Parakeet TDT 0.6B Ultra beside v3 as a local Parakeet model #785)
  • A DeepInfra key on the Hugging Face tab, with speaker labels. The key field takes a DeepInfra API key as well as a Hugging Face token. With a DeepInfra key the three models answer from DeepInfra directly, and a Label speakers toggle, shown with Nemotron, puts a speaker on every word and a paragraph at every change of speaker. The info panel names DeepInfra as the service and the error card names the DeepInfra key or account. (Add Nemotron 3.5 Transcribe as a cloud engine through Hugging Face, with speaker labels #760)
  • A Timing row when the times cannot be right. Nemotron on a long recording sometimes stamps a passage of speech into a few seconds and leaves the audio where it was spoken empty. The transcript keeps the words, the console says so, and Project info names the stretches, kept with the project so a reopened project shows the same row. (Nemotron's word timings collapse on long recordings, and the editor says nothing #786)
  • The cloud engines are sent the sound, not the picture. A video picked for Deepgram, AssemblyAI or the Hugging Face tab used to go up whole. It is now converted in the browser to 16 kHz mono Opus first, about 20 MB an hour, with progress on the card while it converts. An audio file goes up as it is. (Send the cloud engines the audio, not the video #788)

@maboa
maboa merged commit 568ed5b into main Nov 10, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

1 participant