Speech-to-Text Packages

Built-in speech-to-text packages transcribe supported audio or video attachments before the model request. Each provider package installs an input slot at agent/input/speech_to_text, so install only one STT provider package in a core at a time.

Packages

Package	Provider	Capability
stt_openai	OpenAI	network.fetch
stt_groq	Groq	network.fetch
stt_deepgram	Deepgram	network.fetch
stt_assemblyai	AssemblyAI	network.fetch
stt_gemini	Gemini	network.fetch
stt_dashscope	Alibaba Cloud Model Studio	network.fetch
stt_baidu	Baidu Cloud Speech Recognition	network.fetch
stt_tencent	Tencent Cloud ASR	network.fetch

Install

Use the interactive manager:

uv run demiurge package

Or preview and install with subcommands:

uv run demiurge package install stt_openai --core assistant --preview
uv run demiurge package install stt_openai --core assistant

To switch providers, uninstall the current STT package first:

uv run demiurge package uninstall stt_openai --core assistant
uv run demiurge package install stt_gemini --core assistant

Shared Behavior

All STT packages install:

agent/input/speech_to_text/
agent/lib/stt_common/
agent/lib/<provider>/

The input slot appends to the serial input pipeline. When a turn has a supported audio or video attachment, the slot requires network.fetch, sends the attachment to the provider, and adds a Voice message transcript section to the current user prompt. If no supported attachment is present, it does nothing.

Supported media includes common audio formats such as AAC, FLAC, M4A, MP3, OGG, OPUS, WAV, and WebM, plus MP4/WebM video attachments when the provider accepts them.

Options

Package	Options
stt_openai	api_key, language, context_hint, timestamp_granularity, include_metadata
stt_groq	api_key, language, context_hint, timestamp_granularity, include_metadata
stt_deepgram	api_key, language, detect_language, diarization, include_metadata
stt_assemblyai	api_key, language, detect_language, speaker_labels, include_metadata
stt_gemini	api_key, language, include_timestamps, include_metadata
stt_dashscope	api_key, language, include_metadata, enable_itn
stt_baidu	api_key, secret_key, access_token, language, dev_pid, include_metadata
stt_tencent	secret_id, secret_key, region, engine_model_type, include_metadata

timestamp_granularity choices are none, segment, and word.

Optional secret options can be passed during installation or left empty so the installed provider config reads environment variables at runtime.

Credential Environment Variables

Package	Environment variables
stt_openai	DEMIURGE_OPENAI_API_KEY or OPENAI_API_KEY
stt_groq	DEMIURGE_GROQ_API_KEY or GROQ_API_KEY
stt_deepgram	DEMIURGE_DEEPGRAM_API_KEY or DEEPGRAM_API_KEY
stt_assemblyai	DEMIURGE_ASSEMBLYAI_API_KEY or ASSEMBLYAI_API_KEY
stt_gemini	DEMIURGE_GEMINI_API_KEY, GEMINI_API_KEY, or GOOGLE_API_KEY
stt_dashscope	DEMIURGE_DASHSCOPE_API_KEY or DASHSCOPE_API_KEY
stt_baidu	DEMIURGE_BAIDU_ACCESS_TOKEN, or DEMIURGE_BAIDU_API_KEY plus DEMIURGE_BAIDU_SECRET_KEY
stt_tencent	DEMIURGE_TENCENT_SECRET_ID plus DEMIURGE_TENCENT_SECRET_KEY

Verify

List installed packages:

uv run demiurge package list --core assistant

Send an audio attachment through the TUI or a configured channel. The model request should include a Voice message transcript section before or beside the original user text.

Uninstall

uv run demiurge package uninstall stt_openai --core assistant --preview
uv run demiurge package uninstall stt_openai --core assistant

Uninstall removes the package-owned input slot and provider lib. Shared agent/lib/stt_common/ is kept if another installed package still references it.