Listen Stt

Offline voice dictation for OpenCode V2 on Windows. Press a key, talk, and the text lands in your prompt.

• 99 languages. whisper.cpp with large-v3-turbo — the model's full range, not a curated subset.
• The right language, first try. The spoken language is pinned rather than guessed: more accurate, and roughly twice as fast.
• GPU accelerated. CUDA on NVIDIA, Vulkan on AMD and Intel Arc. Detected automatically — you do not pick a backend.
• Fully offline. No cloud, no API key, no account. Audio never leaves your machine.

Not affiliated with the OpenCode team. Independent community plugin.

Voice output

This plugin only listens. To hear answers read aloud — Microsoft Edge neural voices, free, no API key, any language Edge supports — see opencode-speak-tts. The two are independent; either one works without the other.

Why this exists

Dictation plugins for OpenCode tend to arrive in one of two shapes. Either they target the older plugin API and do not load on OpenCode 2.x at all, or they ship a single specialised model covering roughly two dozen languages, which rules out everything outside that set. Neither is a mistake — they are limits of the approach, and neither is fixable from a settings file.

whisper.cpp covers 99 languages and runs on the GPU you already have. This plugin is built around that: nothing curated, and no setting left to chance.

The language problem

Speech models have a quality cliff. A model can be excellent at some languages and unusable at others — and when it meets one it does not know, it does not say so. It returns confident, plausible, wrong text.

There are two separate causes, and they need different answers:

• The model does not cover the language at all. No setting helps. The model is the limit, and the fix is a different model.
• The model covers it, but nothing told it which one to use. Whisper reads a Turkish recording as English when left to auto-detect. Naming the language fixes it, and roughly halves the transcription time as well, because the model stops testing hypotheses.

The second is the quieter failure. A model with no coverage announces itself; a mis-detected one looks like the model is simply bad at your language, which sends people looking for a bigger download instead of a setting.

Turkish is where this is easiest to show, because the failure has a distinctive shape. The same recording, three ways:

	What comes back
A model with no coverage of the language	Merhaba Bubir Dikte Denemissidr. Lutvan Bujum Ned Orias.
Whisper, language left to auto-detect	Read as English, returned as Latin-script noise
Whisper, language set to tr	Merhaba, bu bir dikte denemesidir. Lütfen bu cümleyi doğru yaz.

So this plugin always sends the language. The installer writes tr; set any other with /voice-language. That is the whole of it.

Requirements

• Windows 10/11
• OpenCode V2 (opencode --version → 2.x)
• ffmpeg on PATH — winget install Gyan.FFmpeg
• A ggml model (the installer fetches one, ~547 MB)

Install

git clone https://github.com/fahrim/opencode-listen-stt "$env:TEMP\opencode-listen-stt"
pwsh -NoProfile -File "$env:TEMP\opencode-listen-stt\install.ps1"

The installer does the whole job: it detects your GPU, installs the plugin, fetches the whisper.cpp runtime, downloads the model, and writes your settings file. Run it again at any time to repair an install.

On AMD or Intel, one extra step: upstream publishes no Vulkan build for Windows, so the installer tells you where to get one:

# from https://github.com/fahrim/whisper-vulkan-win
# extract into %LOCALAPPDATA%\opencode-listen-stt\runtime

Then re-run the installer.

Restart OpenCode and check that it loaded:

opencode plugin list    # opencode-listen-stt, showing a local path

Press your hotkey. If no microphone has been chosen yet it asks on the first press, so there is nothing to configure before you start — /voice-mic is only there for changing it later.

Already run whisper.cpp?

If you have a whisper.cpp runtime and a model already, skip the installer and take the plugin from OpenCode:

opencode plugin add github:fahrim/opencode-listen-stt

Point runtime and model at what you have in opencode-listen-stt.json. opencode plugin check and opencode plugin update keep it current.

Do not do both. The installer puts the plugin in ~/.config/opencode/plugins/, and a local folder and a package of the same name are loaded separately — every command would end up registered twice.

Usage

Command	What it does
/voice-record	Start or stop dictation (also your configured hotkey)
/voice-submit	Start, then dictate and add in one step
/voice-cancel	Discard the current recording
/voice-mic	Choose the microphone
/voice-language	Choose the dictation language
/voice-hotkey	Change the recording shortcut
/voice-status	Show backend, language and device

All commands are in the palette (ctrl+p) under Voice.

Transcribed text is inserted into the composer for you to review before sending.

Performance

Measured on an AMD Radeon RX 7700 XT, large-v3-turbo-q5_0, a 5-second Turkish clip:

Backend	First transcription	Warm
CPU (12 threads)	11.1s	~11s
Vulkan	2.1s	0.20s

Whisper is CPU-only upstream on Windows, so the Vulkan build comes from whisper-vulkan-win.

Configuration

~/.config/opencode/opencode-listen-stt.json:

Key	Default	Meaning
language	tr	Spoken language passed to whisper. The installer writes tr; with no config file at all it falls back to auto, which is slower and guesses.
backend	auto	auto, cuda, vulkan or cpu
model	installer download	Full path to a ggml model file. The installer writes it; delete the key and the plugin still finds the model in models/.
runtime	detected	whisper.cpp folder
ffmpeg	ffmpeg	ffmpeg executable, if it is not on PATH
port	8801	Local server port
threads	0	0 means auto
sampleRate	16000	Capture sample rate, in Hz
idleMinutes	10	Model unload delay after the last recording
maxSeconds	300	Hard cap per recording
hotkey	ctrl+y	Recording shortcut
mic	""	Microphone device name; empty means ask on first use

Unknown or wrongly-typed values are ignored rather than fatal, so a typo cannot stop the plugin loading.

Languages

whisper.cpp recognises 99. /voice-language lists the common ones under their English names; anything else takes its whisper code — tr, en, de, fr, es, ru, uk, pl, ar, hi and the rest — in the config file.

auto is available, and it is the slowest and least reliable of the choices, because it is a guess. The language problem above is what a wrong guess looks like.

Troubleshooting

• No devices listed — check ffmpeg -list_devices true -f dshow -i dummy, and Windows Settings → Privacy → Microphone.
• Transcription is in the wrong language — /voice-language, or set language in the config. If it is already pinned, the audio may simply not be in that language.
• Slow — /voice-status shows the backend. If it says CPU on an AMD or Intel GPU, the Vulkan runtime is missing or in the wrong folder.
• The result is plausible but wrong — that is a mis-detected language, not a bad recording. Pin it; see the language problem.
• Diagnostics — %LOCALAPPDATA%\opencode-listen-stt\listen-stt.log.

Development

npm install
npm run typecheck
npm test

License

MIT — see LICENSE.

This plugin drives whisper.cpp, which is MIT-licensed and is not redistributed here — it is downloaded at install time. ggml models carry their own licences; check the model card.