You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Dub Studio 4.5.0 — Vulkan for every stage, CrisperWhisper, the update screen
Qwen3-TTS on CUDA 12.9, Vulkan, the processor, Linux and the Mac, as in Speech Studio: the engine (qwentts.cpp engine-0.2.0, engine-0.2.1 on the Mac) is one library, and the voicing device of Settings → Models adds the module it runs on — CUDA 13 or 12.9 for the card, Vulkan for a card of another make — while the processor needs none and the Mac runs it on Metal. A render fetches the module of the device chosen; an installed studio counts the former engine as missing and downloads the new one in its place.
Vulkan for every local stage, as in Speech Studio: voicing (Higgs and Supertonic), recognition (the audio.cpp server and transcribe.cpp), separation and diarization each offer Auto, GPU (CUDA), Vulkan and CPU in Settings → Models; «Auto» takes Vulkan on an AMD or Intel card, and the GPU tab opens on any card that runs CUDA 13 or 12.9. transcribe.cpp has its Vulkan build as a component of its own.
Google Gemini voicing plans a render as Speech Studio does: a new «Auto» mode sends up to 10 phrases as direct requests and more as one Batch at half the price; direct requests are retried on rate limits and server errors, a refusal for money or access stops the rest at once, and the reasons Google gives are told in plain words. Every phrase voiced is kept in the project and is not paid again; a phrase Google refused names why. Settings → Models (Google) shows what the cloud cost so far and takes a daily limit for agents and programs.
Noise out of a library voice, as in Speech Studio: DPDFNet (10.5 MB) cleans a voice's reference before cloning, with a limit of how far the noise goes down; the original is kept in voices/.originals and comes back with one call. Renaming or deleting a voice takes its original along.
Parakeet Ultra hears a long recording in pieces cut at its pauses, as in Speech Studio: Silero VAD v6.2 (1.2 MB, run by transcribe.cpp) finds where the speech is, so silence does not go to the recognizer and a piece never ends mid-word.
Gemma runs on the llama.cpp build of the card, as in Speech Studio: CUDA 13, CUDA 12 on older cards and drivers, Vulkan on a card of another make, the processor without one. On a Linux whose glibc is older than 2.38, which the CUDA builds of llama.cpp need, it takes the Vulkan build, and Settings → Models says why.
The voicing model stays loaded between renders, as in Speech Studio: voicing a phrase again or the next render takes the model already on the card instead of loading it anew. It leaves after the minutes of Settings → Models → «Unload the voicing model after idle» (15 by default, or never), before the studio's own Gemma starts, on a card under 12 GB before recognition and separation, and when its engine stalls.
One Windows installer: the release has the NSIS setup.exe and the portable build; the MSI is no longer made, as in the other studios. A studio installed from an MSI of an earlier version is updated by running the new setup.exe.
Supertonic 3 voices in the audio.cpp server, as in Speech Studio: the server reads the installed int8 model as it is, on the build of the voicing device (the card's CUDA build, Vulkan or the processor); ONNX Runtime no longer runs it. Voicing a dub's audio alone fetches the server build it needs, as a render does.
Diarization runs in the audio.cpp server, as in Speech Studio: Nemotron 3 Diarization reads the installed model as it is, on the build of the diarization device — the card's CUDA build (13 or 12.9), Vulkan or the processor — and leaves the card when the speakers are found; a job fetches that build, and diarization no longer needs ONNX Runtime for the card and cuDNN.
Parakeet TDT v3 and Nemotron 3.5 recognise in the audio.cpp server, as in Speech Studio: the server reads the same GGUF, on the build of the recognition device — the card's CUDA build, Vulkan on a card of another make, or the processor — and Higgs voices take the build of the voicing device the same way; Parakeet Ultra stays on transcribe.cpp. The server's other builds are components of their own, fetched when a stage is set to them.
Recognition of other languages is CrisperWhisper 2.0 turbo instead of Faster-Whisper: 99 languages with word timestamps, run by the audio.cpp engine on the graphics card or the processor. When the source language is left on Auto, the VoxLingua107 language id hears it first; a language CrisperWhisper does not know is named in the error instead of being read as English. Settings → Models offers one model, without a choice of size and compute; the old Faster-Whisper program, its CUDA 11 libraries and its models are removed by the update screen once their replacement is in place (a model another program shares stays with it). Its licence is nyra health Non-Commercial (LICENSE.md beside the weights).
Fresher tools: the transcribe.cpp engine 0.1.4 (transcribe 0.4.1) for recognition, llama.cpp b11900 for translation, FFmpeg of 30 September, yt-dlp 2026.10.21 with deno 2.9.8 for links; an installed studio counts the older ones as missing and downloads the new ones in their place.
Higgs Audio v3 voices run in the audio.cpp engine (a server process of the studio) instead of the old library: a stop ends the phrase being spoken at once, and the weights leave the card when voicing is done.
Added
The update screen: the first start after an update that changes the studio's engines and models shows what is updated (old → new, size), what is new and what is no longer needed (the space it frees), each with the reason. One button downloads it all and then removes the old files; nothing is removed before its replacement is in place, so a cut-off update leaves a working studio. «Later» puts it off; Settings → Models keeps it under «What changed in this release», and agents have it as models_update_plan and models_update_apply.
News and notices of the studios' service in full: a notice can now open as a window over the studio, which waits until nothing is being made (analysis, voicing, export); notices show in the sections they are meant for, carry their picture, buttons and the ad label where there is one, and a button can lead to the news, the settings, the models or «About».
The project's glossary helps CrisperWhisper spell its terms: the source terms of the glossary go to recognition as words to expect, so names and terms come out as the glossary writes them.
The emotion of the original, per project (#44): with the cloned voice, the editor's voice section chooses it for this project: in every line (each line is cloned from its own piece of the original, so shouting and whispering carry over, but the timbre can shift from line to line), one voice per speaker (every line from the speaker's one reference, a steady voice for a narrator or a lecture), or as Settings - Quality says. Agents set it with voice_emotion_set.
Fixed
macOS takes its own engines: on a Mac the studio downloaded the Linux builds of its engines and tools; it now takes the Mac's (the Metal build of the audio.cpp server, transcribe.cpp, BSRoformer and llama.cpp for Apple silicon, ONNX Runtime for macOS, ffmpeg and yt-dlp for macOS), opens files with open, and restoring a deleted project says to put it back from the Bin in Finder, which macOS leaves to it.
Voice separation on Linux and on older NVIDIA cards: the studio takes the studios' own builds of BSRoformer.cpp (CUDA 13, and CUDA 12.9 for cards and drivers CUDA 13 does not run) instead of the author's, which on Linux needed a newer system than Ubuntu 22.04 and its CUDA libraries from the build machine, and on Windows compiled itself for the card on the first run; separation on the card now opens on any card a CUDA build takes.
Speech recognition on older NVIDIA cards and drivers runs on the card: where CUDA 13 does not run (a driver before 580, a card before Turing), recognition took the processor; it now takes the transcribe.cpp engine built for CUDA 12.9 (and CrisperWhisper the CUDA 12.9 audio.cpp server) with the CUDA 12.9 runtime, and «GPU» is open for recognition there.
An engine taken from another program counts as installed: the install record of the older build it replaced stayed beside it, so the studio kept it uninstalled and asked for it again; that record now goes with the old file and the new one is recorded as what it is.
A program's own file is no longer lost when another program's copy cannot be taken: at start the studio takes what another program already has, and a file on another drive (where it cannot be linked) left the studio's own older file deleted, so yt-dlp or an engine went missing until it was downloaded again; now the own file is replaced only by a finished link or copy.
Linux: Higgs voices and CrisperWhisper run again: the studio installs the audio.cpp server on Linux too (CUDA 13, or CUDA 12.9 with its runtime for older cards) instead of the engine library it no longer runs, and makes the programs it unpacks executable; Windows takes the same server release.
The log on disk survives a crash: logs/dub-studio.log gets every line as it is written, with panics and, on Windows, a crash of an engine library (its exception code and module); it goes into the bug report package, and past 8 MB the older part is kept as dub-studio.old.log.
An engine or program counts as installed only as the pinned build: a file of the right size but another build (an older engine, one another program left) is hashed once, and the studio downloads the pinned one in its place.
Downloads trust the system's certificates: behind a proxy or an antivirus that checks HTTPS with its own certificate installed in the system, models and engines download instead of failing on the certificate; the check of a large download now moves the bar instead of standing at 100 %.
Linux and macOS know the free space of the models folder: First run said the free space is unknown, and a download too big for the disk was not stopped before it began.
Linux: the system's ffmpeg and other programs of the computer start from the AppImage: they no longer load the image's own libraries (on systems newer than Ubuntu 22.04 ffmpeg stopped with «undefined symbol vaMapBuffer2»); yt-dlp, the file manager and the browser start the same way.
Settings → Models keep their place: the list no longer jumps back up after a choice, and its buttons are no longer rebuilt every few seconds.
The status line names the translator the studio really uses: Gemma's file, the OpenRouter or local server model, or the agent (MCP); it showed Gemma whatever was chosen, and it now follows a preset or a change in Settings at once instead of after a restart.
A hardware preset moves every stage and stays shown: «CPU + cloud» left recognition, separation and diarization on the card when they had a device of their own, and Settings → Models showed the recommended preset again once reopened; now a preset sets the device of each stage, and the list shows the preset the settings are on, or Custom.
A Higgs line longer than its slot is no longer cut as a runaway: the token cap of a line is the larger of its slot's and its own text's (the cap the studios share, by the line's length and script, so Chinese, Japanese and Korean get the time their syllables take), and a line that runs to the cap is retried as a runaway take instead of stopping the voicing as an engine error.
Speech recognition no longer closes the studio on processors without AVX-512 (#46): the transcribe.cpp engine of 4.4.0 was built for the processor it was built on and used AVX-512 instructions, so on most Intel and older AMD processors the first recognition stopped the program without a word, on the card and on the processor alike. The engine is now built for every x86-64 processor with AVX2 (Intel since 2013, AMD since 2017), as fast as before, and on a processor without AVX2 the studio says so instead of closing.
Whisper no longer writes over the silence at the end of a file (#45): words lying wholly in a silence of the recording (at least 0.3 s, 30 dB under its speech) are dropped, and a word's end stops where the speech does, so "subscribe to the channel" over a file's quiet tail is gone and the last line ends with its voice; the check of the voiced lines by ear reads them the same way.
Separation works in folders with non-Latin names: the BSRoformer engine reads its paths through the old narrow command line, so a model or project under a Cyrillic folder failed with "gguf_init_from_file: failed to open GGUF file"; it now gets the short 8.3 names of such paths, and on a disk without short names the error says to move the folder. The engine's own error line comes first in the message.
Diarization on the processor downloads ONNX Runtime before the job when it is missing, instead of stopping the analysis with "Failed to load ONNX Runtime dylib".
Stopping a job or a link download no longer hangs when Windows' process service (WMI) is stuck: the studio ends its processes itself instead of through taskkill.
A damaged or not fully downloaded file is refused when it is opened, with its name and what is wrong with it, instead of becoming a project on which every analysis and render stopped with "ffprobe exited with code 1 … Invalid data found when processing input". So is such a file downloaded by a link.
Separation of recordings longer than about 6 h 45 min: their 44.1 kHz audio is larger than a plain WAV can hold (4 GB) and is now written as RF64, and a WAV whose size is left unknown is read to its end, instead of stopping with "Ill-formed WAVE file: data chunk length is not a multiple of sample size". The separated vocals and music of a recording longer than about 3 h 20 min are written the same way; before, their size wrapped past 4 GB and the stems came out cut short.