[Bug] Any size local Whisper have error

[vthpeke-dotcom]

Version *

v0.2.9 (latest Release)

Platform *

Windows

Affected area *

Timeline editing

Describe the bug *

transcribe_track: transcription failed: 536525664 请到 设置 → 转写 → 本地模型 检查 whisper 模型是否已下载（Settings → Transcription → Local models）。). The successful completion of this operation is not recorded.

000180.log

0814.mp4: 536525664

Steps to reproduce *

When i try use

Expected behavior *

whisper work on Russian language

Actual behavior *

crushing

Media info

No response

Screenshots / logs / error messages

000180.log

[0xsline]

@vthpeke-dotcom, thanks for the report. The local Whisper audio-handling fix and Russian language option are now on main (not yet in a desktop release); a real Agent test with a Russian clip completed successfully on macOS. Please retry with Russian selected after the next desktop release; if it still fails on Windows, share the model size and the new stage-specific error text so we can pinpoint the failing step.

[vthpeke-dotcom]

@vthpeke-dotcom, thanks for the report. The local Whisper audio-handling fix and Russian language option are now on main (not yet in a desktop release); a real Agent test with a Russian clip completed successfully on macOS. Please retry with Russian selected after the next desktop release; if it still fails on Windows, share the model size and the new stage-specific error text so we can pinpoint the failing step.

Ok. I tried several models: Qwen 3.6 35B A3B, Meta Muse Glimmer 30B, Gemma 4 31B in version V0.2.11. And i tried all size of local whisper. I ran the models through LM Studio.

[0xsline]

Update — the real root cause is found and fixed, and it explains exactly what you saw.

The Windows build shipped whisper-cli.exe without the DLLs it needs to start: the packaging step moved the executable out of the archive's Release/ directory but left ggml.dll, whisper.dll and the CPU-variant DLLs behind. Windows resolves a process's DLLs from the executable's own directory first, so the process died instantly with exit code 127 — before printing anything. That bare number is precisely the error you reported, and it's why every model size failed: local transcription on Windows never actually ran, for any model, in any version so far. Your LLM choice (LM Studio etc.) was never the problem — that part is unrelated to local Whisper.

Verified both ways on a real windows-latest CI runner: with the old layout whisper-cli.exe --help exits 127; with the libraries placed beside the executable it runs; re-nesting them breaks it again.

The fix is merged and will be in the next desktop release — please retry then. Any size should work at that point.

[0xsline]

v0.2.12 is out now: https://github.com/0xsline/OpenChatCut/releases/tag/v0.2.12 — please retry local Whisper (any size) on it.

[0xsline]

补充：Windows 请直接安装 v0.2.13（修复了 v0.2.12 关闭窗口时的报错弹框）：https://github.com/0xsline/OpenChatCut/releases/tag/v0.2.13

[vthpeke-dotcom]

补充：Windows 请直接安装 v0.2.13（修复了 v0.2.12 关闭窗口时的报错弹框）：https://github.com/0xsline/OpenChatCut/releases/tag/v0.2.13

Запрос не выполнен из-за сбоя инструмента (transcribe_track: transcription failed: service-unavailable: 模型加载失败：Local ASR model load failed; remote fallback is disabled: Can't create a session. ERROR_CODE: 1, ERROR_MESSAGE: qdq_actions.cc:137 TransposeDQWeightsForMatMulNBits Missing required scale: model.decoder.embed_tokens.weight_merged_0_scale for node: model.decoder.embed_tokens.weight_transposed_DequantizeLinear Если это сбой среды выполнения, проверьте, что проект открыт, выполнен вход в Codex (Настройки → Codex) и служба приложения запущена.). Успешное выполнение этой операции не зафиксировано.