Skip to content

STT Settings ​

Settings → STT controls speech input and transcription. The master Enable Speech-to-Text switch must be on before provider controls appear.

Choosing a provider ​

  • Desktop Local: runs Whisper on your machine via Electron. No network required, works offline. Requires a one-time setup.
  • Android Local: on-device Whisper, SenseVoice, or Dolphin models, Android only. Download model packs once and use them offline.
  • OpenAI / OpenAI-compatible: cloud transcription. No local setup, but requires an API key and an internet connection.
  • Chrome AI Multimodal: in-browser Gemini Nano on supported Chrome builds. No external server or API key.

After picking a provider, use Test Recording (3s) in Common controls to verify it's working before you start a session.

Common controls ​

ControlBehavior
Enable Speech-to-TextMaster switch. Provider sections are hidden until this is on.
ProviderActive transcription backend. Filtered by runtime.
Test Recording (3s)Records a short clip and transcribes it with the current provider.

Android Local ​

Android Local runs speech recognition on your device. No network connection is required after downloading a model.

To set it up:

  1. Download one or more models from the model cards. The available models include Whisper Tiny and Base in English-only or multilingual versions, SenseVoice for Chinese, English, Japanese, Korean, and Cantonese, and Dolphin Base or Small for 40+ Eastern languages.
  2. Choose the model and language. English-only Whisper models ignore the language setting. SenseVoice supports only its listed languages.
  3. Choose a Compute Provider. Auto benchmarks the available backends. CPU is the most compatible. NNAPI and XNNPACK require a custom build. QNN Snapdragon NPU is experimental and requires a compatible device.
  4. Tap Test Recording (3s) to verify transcription.

Use the delete control on a model card to remove it and free up storage.

ControlBehavior
Model cardsDownload and remove Whisper, SenseVoice, Dolphin, and supported SenseVoice NPU packs.
ModelSelect Auto, an installed Whisper variant, SenseVoice, or Dolphin.
LanguageSelect Auto, English, Spanish, Japanese, Chinese, German, or French where the model supports it.
Compute ProviderAuto, CPU, NNAPI, XNNPACK, or experimental QNN on supported builds and devices.

On supported Snapdragon devices, the experimental SenseVoice NPU pack runs on the NPU. Install the SenseVoice model first. The NPU option handles clips up to 30 seconds and splits longer audio automatically.

Desktop Local ​

Desktop Local runs speech recognition through Electron. In Settings → STT or the setup wizard, choose between faster-whisper and whisper.cpp. Both engines can be installed side by side. whisper.cpp is selected by default, and the STT Engine selector chooses which one handles transcription.

On AMD systems using ROCm, Desktop Local STT runs on the CPU. Faster Whisper's GPU path requires NVIDIA CUDA.

faster-whisper ​

The Python engine. It requires a one-time setup step. Select it in STT Engine to install and configure it.

To set up Desktop Local:

  1. Click Initialize (or Install) in the setup controls. Watch the Setup Log for progress.
  2. If anything goes wrong, click Verify / repair to re-run setup on the existing install.
  3. Once the status shows success, select a Whisper Model:
    • tiny: fastest, lowest memory
    • base: better accuracy, still fast (recommended for everyday use)
    • tiny.en / base.en: English-only variants, slightly faster if you only speak English
  4. Set Language. Leave it on auto unless you're getting wrong language detection, in which case fixing it to your language improves accuracy.
  5. Use Test Recording (3s) to verify it's transcribing correctly.

When Desktop Local is active, a microphone selector appears in the Actions area. If your recordings are picking up the wrong source, select your headset or preferred mic there.

Setup controls ​

ControlBehavior
Setup statusShows logs and success/failure banners.
Initialize / installStarts setup when the runtime isn't ready.
Verify / repairRe-runs setup on an existing install.
CancelStops an in-flight setup action.
Setup LogScrollable output during installation.

Transcription controls ​

The Whisper Model setting controls the tradeoff between accuracy and speed:

  • tiny: fastest, smallest memory footprint. Good starting point.
  • base: noticeably more accurate, still reasonably fast. Recommended for everyday use.
  • tiny.en / base.en: English-only variants. Marginally faster than their multilingual counterparts if you only ever speak English.

Leave Language on auto unless you are getting wrong language detection. Setting a fixed language avoids detection overhead and can improve accuracy for non-English speech.

ControlDefaultBehavior
Whisper Modeltinytiny, base, tiny.en, or base.en. The .en variants are English-only but slightly faster.
LanguageautoAuto-detect or a fixed language.

whisper.cpp ​

Use STT Engine in Settings or setup to select whisper.cpp. Choose and install a runtime pack for your system, then download a GGML model. The current packs include Windows CPU, Windows NVIDIA CUDA, Windows Vulkan community builds, and Linux x64 CPU. No macOS runtime pack is listed. Windows Vulkan packs support AMD and Intel GPUs as well as other Vulkan-compatible GPUs.

The available models are tiny, tiny.en, base, and base.en. The multilingual tiny and base models are about 75 MB and 142 MB. The .en variants are English-only. Select a language in the shared Language control, then use Test Recording (3s) to check the active engine.

You can install more than one runtime pack. Use Use on an installed pack to select it. Downloaded models can be removed from their model cards.

Advanced Settings ​

ControlDefaultBehavior
Endpoint URLhttp://127.0.0.1:11438Override the local AI server endpoint.
Threads4CPU thread count, 1–16.

Microphone selection ​

When Desktop Local is active and microphones are available, an input device picker appears in the Actions area.

OpenAI ​

OpenAI transcription uses the Whisper API. No local setup needed but requires an API key and sends audio to OpenAI.

To set up OpenAI STT:

  1. Enter your API Key (starts with sk-).
  2. Select a Model from the dropdown. The list is fetched from OpenAI using your key.
  3. Click Test Recording (3s) to verify the connection.
  4. Optionally save as a named backend using Saved Backends.
ControlBehavior
Saved BackendsNamed presets storing API key and model.
API KeyYour OpenAI key.
ModelFetched from OpenAI based on the key.

OpenAI-compatible ​

Works with any transcription service that implements the OpenAI /audio/transcriptions endpoint.

To set up OpenAI-compatible STT:

  1. Enter the Endpoint URL for your compatible service.
  2. Add an API Key if the service requires one.
  3. Select a Model from the dropdown (fetched from the endpoint).
  4. Set Language if auto-detect isn't reliable for your service.
  5. Click Test Recording (3s) to verify.
  6. Optionally save as a named backend.
ControlBehavior
Saved BackendsStores endpoint, model, language, and optional key.
Endpoint URLBase URL for the compatible service.
API KeyOptional.
ModelFetched from the endpoint.
Languageauto or a fixed language.

Chrome AI Multimodal ​

Browser-native Gemini Nano transcription. Only available in supported Chrome builds.

To enable Chrome AI STT:

  1. Check the Status. If it says the model is downloadable, click Start Model Download.
  2. If the status says Chrome flags are required, click each flag's copy button, paste the URL into Chrome's address bar, enable the flag, then restart Chrome. After restarting, visit chrome://components and wait for the model component to update.
  3. Click Refresh Status to confirm the model is ready.
  4. Set your Output Language if needed.

INFO

Chrome 138 or newer is required. If your Chrome build doesn't support it, the tab shows a warning.

ControlBehavior
StatusAvailability and current download state.
Start Model DownloadAppears when Chrome reports the model is downloadable.
Refresh StatusRe-checks availability.
Output LanguageEnglish, Spanish, or Japanese.
Required Chrome FlagsCopy buttons for optimization-guide-on-device-model and prompt-api-for-gemini-nano-multimodal-input, plus restart instructions.

INFO

Chrome 138 or newer is required. If the runtime reports no Chrome AI support, the tab shows a warning.

| Delete | Removes the selected saved backend. |