Repository navigation
Local dictation returns empty transcript for non-English languages since 1.141.0 (foundry-local-sdk 2.0.1 drops language) #340656
Description
Does this issue occur when all extensions are disabled?: Not tested (the cause is in the bundled foundry-local-sdk in core, not in an extension)
- VS Code Version: 1.141.0 (user setup)
- OS Version: Windows_NT x64 10.0.26200
Steps to Reproduce:
- Set
"agents.voice.language": "ja"and use local dictation (nemotron-3.5-asr-streaming-0.6b-generic-cpu:3). - Start dictation in the chat input. It shows "Listening…".
- Speak Japanese, then stop dictation. No text is inserted.
- Speak English instead. Text is inserted normally.
Expected: Japanese speech is transcribed, as in 1.140.x.
Cause (investigated):
- 1.141.0 bumped
foundry-local-sdkfrom 1.2.3 to 2.0.1 (chore: bump foundry-local-sdk to 2.x #335329). - In 1.2.3,
LiveAudioTranscriptionSession.start()passedLanguagetoaudio_stream_start. - In 2.0.1,
settings.languageis stored but not forwarded to the native request, so the multilingual model does not receive the language. - In the same environment,
AudioClient.transcribe()withlanguage: "ja"on a WAV file returns Japanese text, so the model itself supports Japanese.
Upstream fix: microsoft/foundry-local#1064 and microsoft/foundry-local#1138, released in foundry-local-sdk 2.1.0. Please bump the bundled SDK to 2.1.0 or later.
Activity
opensas commentedon Oct 19, 2026
Same here on VS Code 1.142.0 / Ubuntu 22.04 with Spanish (
agents.voice.language: "es").English is transcribed fine; Spanish returns an empty transcript (
[chat-stt] skipped language model cleanup (reason=noTranscript)). I reproduced it outside VS Code with the bundled foundry-local-sdk, using a live transcription session at 16 kHz / mono / 16-bit: an English sample transcribes perfectly, while a clear Spanish recording returns nothing withsettings.languageunset,es,es-ESandes-US.Workaround: switching to the
maimodel works for Spanish:amIsmael commentedon Oct 26, 2026
Still reproducible on the current Insiders: Code - Insiders 1.144.0-insider (commit 131aec7, Oct 26, Linux x64),
"agents.voice.language": "es", local modelnemotron-3.5-asr-streaming-0.6b.Confirmed on disk that 1.144.0-insider ships the reverted SDK:
~/.config/Code - Insiders/chatDictationRuntime/2.0.1is still in use (2.1.0 was present on Oct 19, 2.0.1 since Oct 22).There has been no follow-up since the revert: no dictation entry in the 1.144 notes (updated Oct 21) and no new PR re-landing Foundry Local 2.1.x.
Status note: the 2.1.0 bump described here was already attempted in #341640, but reverted because 2.1.0 segfaults on Linux x64 with a cold model cache (#342641). The revert landed in #342681 (1.144) and #342703 (1.143). So the language is still not forwarded.
Could we get a follow-up that bumps to 2.1.0+ with the Linux x64 crash fixed?
mai(cloud) is not an option for me./cc Megan Rogge (@meganrogge)
meganrogge commentedon Oct 26, 2026
Hi, my focus has been elsewhere. I do plan to get to this ASAP