Skip to content

Version 1.3.45 - #783

Merged
maboa merged 1 commit into
mainfrom
release/1.3.45
Nov 8, 2026
Merged

maboa merged 1 commit into
mainfrom
release/1.3.45

Conversation

@maboa

@maboa maboa commented Nov 8, 2026

Copy link
Copy Markdown
Member

Release 1.3.45. Fixes #773, fixes #775, fixes #776, fixes #777, fixes #780, fixes #781, fixes #782.

Changed

  • Caption editor rows. Each caption is one row: a play button, the In and Out times in one column with a cue button before each, the text box beneath them, and the reading rate with the row menu under it. The column is the transcript's width, and on a narrow screen the row folds. The menu offers Insert new caption, Merge with next and Delete, in the style of the Recents menu, and closes on a click outside or Escape. A new caption takes the gap it is inserted into, with a small margin on either side. Two captions that overlap in time are marked in red at the times that meet. The editor follows the media as it plays, a click on a time seeks to it, Enter moves the rest of a line down, and Alt+[ and Alt+] set In and Out from the playhead while Alt+Up and Alt+Down nudge a time by a tenth of a second. (Caption editor: one cue per row, and the editor follows the media #773)
  • The selected passage shows its length. A pill in the bottom-right corner of the transcript card gives the length of the current selection, as a clip would be exported. The read-only badge moves to the bottom-left. (Show the duration of a transcript selection #775)
  • One selected word is not a clip by default. Export media still offers the selection, but a single word no longer ticks the clip option on its own and shows no length. (A single selected word should not choose the Selection option in Export #776)
  • Hugging Face engine. The provider that served Parakeet withdrew it, and the route the engine used cannot be paid with pre-paid credits, so the engine was dead for everyone. It now posts to DeepInfra through the same router and the tab is called Hugging Face, with a Model select offering Nemotron 3.5 Transcribe 30B, Whisper large v3 turbo and Qwen3-ASR 1.7B. The Language list follows the model. The error card names a bad token, an account with no credit and a model the provider no longer serves. The remembered token and the dialog's fields are unchanged. (Parakeet (Hugging Face) fails with 410: Together no longer serves the model #777)
  • Estimated cost in the info panel. A Hugging Face run shows what it cost, worked out from the audio length the service reported and the model's published rate per minute, with the rate shown beside it. The two numbers are saved with the project, so a reopened project shows the same row. (Show the estimated cost of a cloud transcription in the info panel #780)
  • A host can name the URL for an imported media file. The Deepgram JSON, SRT, VTT and Whisper JSON imports ask an optional window.hyperaudioMediaUrl(file) before they put a chosen media file on the player as a blob URL. A host application that keeps the media itself and serves it over a scheme it has declared answers with its own URL, and the project is then a link to the host's copy rather than a second copy in the editor's storage. With no hook, or any answer off the declared schemes, nothing changes. Documented with the other host hooks in the format document. (Let a host name its own URL for a media file chosen in an import dialog #781)
  • Paragraphs on edge to edge timings. Nemotron, Whisper and Qwen end a word where the next begins, so a speaker's pause between sections is one to two seconds, never the four the paragraph rule looked for, and a talk came out as one paragraph. A paragraph now also ends at a sentence end followed by a pause of over a second, once it has twenty words.

Fixed

  • Whisper left a space before a hyphen. Whisper on the Hugging Face route returns a hyphenated word in pieces, so the transcript read "speech -to -text" and "built -in". The pieces are joined into one word with the timing of the whole.
  • The media link field showed only the tail of a pasted address. The field was capped at 320 pixels in a wider dialog, and a single-line field scrolls to the caret after a paste, so a long URL looked like its file name. The nine link fields in the transcribe and import dialogs take the dialog's width, and after a paste the address reads from the start. (The media link field shows only the tail of a pasted address #782)
  • Audio being transcribed wore the previous project's picture. After a video project, a recording that was plainly sound showed a film strip until the transcription finished. The wave shows from the moment the audio loads.
  • A new transcription inherited the previous project's summary and topics. Only Deepgram writes them, so every other engine's project started with whatever the last project had, and showed it as its own. They are cleared as the transcription starts.
  • Words with no duration opened the wrong paragraph. Qwen times one word in seven as a single instant, a sentence-final word at the moment the next word begins, and that word was placed at the head of the following paragraph. Such a word is now given a duration before the paragraphs are placed, which also lets the karaoke highlight and the captions time it.

@maboa
maboa merged commit 096ec73 into main Nov 8, 2026
1 check failed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment