Speech-to-Text

Transcribe voice input (from the web chat mic button and Telegram voice messages) into text the agent can read. Optionally rewrite the raw transcript with an LLM to clean up filler words and punctuation.

URL: /settings?tab=stt

Enabled

Master toggle. When off, the mic button is hidden in the UI and Telegram voice messages are passed through to the agent as unknown attachments.

{ "stt": { "enabled": false } }

Provider

Which backend handles the actual transcription. The openai and ollama options reference an entry you already configured on the Providers page — the API key / base URL is read from there, not from Secrets.

Value	Notes
openai	Uses an OpenAI-compatible provider from providers.json (typically your OpenAI entry). Calls its /v1/audio/transcriptions endpoint with OpenAI's Whisper / GPT‑4o transcription models.
whisper-url	A self-hosted Whisper-compatible HTTP endpoint (e.g. faster-whisper-server, whisper.cpp server). Standalone — does not use a provider entry, just a raw URL. Local-first, zero API cost.
ollama	Uses an Ollama provider from providers.json. A Whisper-family model served by that Ollama instance handles the transcription.
deepgram	Hosted Deepgram pre-recorded audio API. Standalone — does not use a provider entry. Uses the API key configured directly in the Deepgram card below (stored in stt.deepgramApiKey).

The dropdown lists each matching provider from providers.json as its own entry, e.g. OpenAI Whisper (My OpenAI) or Ollama (Ollama Mac). If no matching provider is configured, the option appears disabled.

The selected provider's id is stored in stt.providerId; stt.provider stores only the backend type:

{ "stt": { "provider": "openai", "providerId": "openai-main" } }

OpenAI model

Shown when provider is openai. The model is sent to the selected provider's /v1/audio/transcriptions endpoint.

Value	Notes
whisper-1	Classic Whisper, cheap, decent.
gpt-4o-transcribe	Newer, higher quality.
gpt-4o-mini-transcribe	Cheaper gpt-4o variant.

{ "stt": { "openaiModel": "gpt-4o-mini-transcribe" } }

Whisper URL

Shown when provider is whisper-url. Full URL to the transcription endpoint, e.g.:

http://whisper:9000/v1/audio/transcriptions

Must be an OpenAI-compatible /audio/transcriptions route. Usable over HTTPS or plain HTTP on the internal Docker network.

{
  "stt": {
    "whisperUrl": "http://whisper:9000/v1/audio/transcriptions"
  }
}

Ollama model

Shown when provider is ollama. The model tag as pulled into the Ollama server referenced by the selected provider entry, e.g. dimavz/whisper-tiny.

{ "stt": { "ollamaModel": "dimavz/whisper-tiny" } }

Deepgram model

Shown when provider is deepgram. The Deepgram transcription model used for incoming audio. nova-3 is Deepgram's current production model and a sensible default.

The dropdown is pre-populated with a small list of common models. Click the refresh icon next to the dropdown to fetch the full, up-to-date model catalog from your Deepgram account — useful when Deepgram releases a new model. Refreshing requires the API key to be saved first.

{ "stt": { "deepgramModel": "nova-3" } }

Deepgram language

Shown when provider is deepgram. ISO‑639‑1 language code (en, de, es, ...) or multi for multilingual transcription. Leave empty to auto-detect.

{ "stt": { "deepgramLanguage": "de" } }

Deepgram API key

Shown when provider is deepgram. The API key used to authenticate against Deepgram. Stored encrypted at rest in stt.deepgramApiKey. Get a free key at console.deepgram.com.

The field is rendered as a password input. Once saved, it shows a masked preview (e.g. dg_••••••abcd) — leave the masked value untouched to keep the existing key.

{ "stt": { "deepgramApiKey": "dg_..." } }

Rewrite enabled

After transcription, run the raw text through an LLM to:

• Strip filler ("äh", "umm", "like").
• Fix punctuation and capitalization.
• Keep the original language.

Useful for messy dictations. Adds latency and token cost — leave off if your transcriber is already clean.

{ "stt": { "rewrite": { "enabled": false } } }

Rewrite provider

When rewrite is on, picks which provider+model does the cleanup. Any enabled chat provider from the Providers page can be selected. If left empty, rewriting is silently skipped and the raw transcript is returned.

The stored value is a composite providerId::modelId string:

{ "stt": { "rewrite": { "providerId": "openai-main::gpt-5.4-mini" } } }