Skip to main content

Google Gemini Provider Guide

A single Google Gemini account can power all six Insights services. Gemini is the only provider with a native path to every service, which makes it the simplest single-vendor deployment.

Two Gemini APIs are used, and the distinction matters when you configure options:

  • Gemini Live (BidiGenerateContent) — a bidirectional WebSocket used by transcription (text responses) and ai_agent (audio responses).
  • Gemini REST models — Models.GenerateContent for translation and chat, the Interactions API for speech synthesis, the Files API + Batches.Create for meeting summarization.

Both Live and the Interactions API require api_version: "v1beta".

Supported Services​

ServiceGemini pathDefault model
transcriptionGemini Live BidiGenerateContent, ResponseModalities=[TEXT]gemini-3.5-transcribe-live
translationModels.GenerateContent with ResponseMIMEType: application/jsongemini-3.1-flash-lite
speech-synthesisGemini Interactions API returning audio/L16 PCM16gemini-3.8-flash-tts
ai_text_chatStreaming chat, up to 8 rounds of MCP tool callingrequired, for example gemini-3.1-flash
meeting_summarizingFiles API upload plus Batches.Create batch jobrequired, for example gemini-3.1-flash-lite
ai_agentGemini Live BidiGenerateContent, ResponseModalities=[AUDIO]gemini-3.8-live

Get an API key from Google AI Studio or Google Cloud, then define one account and reference it from every service.

Provider Account Configuration​

insights:
providers:
google:
- id: "default-gemini-creds"
credentials:
api_key: "YOUR_GEMINI_API_KEY"
options:
api_version: "v1beta"
KeyTypeRequiredDescription
idstringYesUnique account name. Services reference it, for example "default-gemini-creds".
credentials.api_keystringYesGemini API key.
options.api_versionstringNoAPI version. Set "v1beta", which is required for Gemini Live (transcription, ai_agent) and for the Interactions API (speech-synthesis).
options.endpointstringNoCustom gateway or proxy base URL. Leave empty to call Google directly.
options.auth_headerstringNoHeader used to carry the key when a gateway is in front of Gemini. "Authorization" sends Authorization: Bearer <api_key>. Any other name sends <header>: <api_key>.

Service Configuration​

transcription​

Streams the microphone tracks of allowed speakers to Gemini Live and receives text deltas. Live is opened with ResponseModalities=[TEXT], so the model never returns audio for this service.

transcription:
provider: "google"
id: "default-gemini-creds"
options:
model: "gemini-3.5-transcribe-live"
transcribe_mode: "VERBATIM"
api_version: "v1beta"
max_selected_trans_langs: 2
pricing:
default:
price_per_hour: 0.36
KeyTypeDefaultDescription
options.modelstring"gemini-3.5-transcribe-live"Live transcription model.
options.transcribe_modestring"VERBATIM""VERBATIM" transcribes word for word. "SMART" lets the model normalize and clean up speech.
options.api_versionstring"v1beta"Required for BidiGenerateContent.
options.max_selected_trans_langsinteger2Maximum simultaneous caption languages per room. Enforced server-side; a host cannot raise it from the API.

translation​

Translates a batch of strings into every configured target language in a single GenerateContent call. The request sets ResponseMIMEType: application/json, so the model returns one JSON object instead of prose, which keeps multi-language chat translation to one round trip.

translation:
provider: "google"
id: "default-gemini-creds"
options:
translate_model: "gemini-3.1-flash-lite"
max_selected_trans_langs: 5
pricing:
default:
price_per_million_characters: 10.00
KeyTypeDefaultDescription
options.translate_modelstring"gemini-3.1-flash-lite"Translation model. model is accepted as an alias.
options.max_selected_trans_langsinteger5Maximum target languages for chat and caption translation. Separate from the transcription limit.

speech-synthesis​

Calls the Gemini Interactions API and reads PCM16 audio (audio/L16) straight from the response, so no file is written to disk.

speech-synthesis:
provider: "google"
id: "default-gemini-creds"
options:
model: "gemini-3.8-flash-tts"
default_voice: "Kore"
tts_sample_rate: 16000
api_version: "v1beta"
pricing:
default:
price_per_million_characters: 16.00
KeyTypeDefaultDescription
options.modelstring"gemini-3.8-flash-tts"TTS model.
options.default_voicestring"Kore"Voice used when a language has no explicit voice mapping.
options.tts_sample_rateinteger16000PCM sample rate. Allowed values: 8000, 16000, 24000.
options.api_versionstring"v1beta"Required by the Interactions API.

Gemini uses one default_voice for every language rather than a per-language voice map. If you need distinct voices per target language, use the Microsoft Azure guide.

ai_text_chat​

Streaming chat for moderators (room copilot) and participants (learning assistant). The moderator path can call MCP room tools for up to 8 rounds before it must answer.

ai_text_chat:
provider: "google"
id: "default-gemini-creds"
options:
chat_model: "gemini-3.1-flash-Preview"
summarize_model: "gemini-3.1-flash-lite"
context_window: 5
pricing:
"gemini-3.1-flash-Preview":
input_price_per_million_tokens: 0.30
output_price_per_million_tokens: 2.50
"gemini-3.1-flash-lite":
input_price_per_million_tokens: 0.10
output_price_per_million_tokens: 0.40
KeyTypeRequiredDescription
options.chat_modelstringYesModel for the live conversation and tool calling.
options.summarize_modelstringNoModel used to condense long context.
options.context_windowintegerNo, default 5Number of recent messages kept verbatim per conversation.

meeting_summarizing​

Uploads the mixed room recording through the Gemini Files API, then runs a batch job with Batches.Create and stores the returned summary text. Uploads are removed after the job finishes.

meeting_summarizing:
provider: "google"
id: "default-gemini-creds"
options:
summarize_model: "gemini-3.1-flash-lite"
pricing:
default:
input_price_per_million_tokens: 0.30
output_price_per_million_tokens: 2.50
KeyTypeRequiredDescription
options.summarize_modelstringYesModel used for the batch summarization job.

Custom prompts for the summary are set per room, not here. See AI Meeting Summarization.

ai_agent​

Runs the voice and vision agent on Gemini Live with ResponseModalities=[AUDIO]: the model hears the room audio, sees whiteboard and share frames, calls agent tools, and speaks back.

ai_agent:
provider: "google"
id: "default-gemini-creds"
options:
realtime_model: "gemini-3.8-live"
summarize_model: "gemini-3.1-flash-lite"
api_version: "v1beta"
pricing:
realtime_model:
input_price_per_million_tokens: 0.75
cached_input_price_per_million_tokens: 0.075
cache_write_input_price_per_million_tokens: 0.9375
audio_input_price_per_million_tokens: 3.00
audio_cached_input_price_per_million_tokens: 0.30
audio_cache_write_input_price_per_million_tokens: 3.75
image_input_price_per_million_tokens: 1.00
image_cached_input_price_per_million_tokens: 0.10
image_cache_write_input_price_per_million_tokens: 1.25
output_price_per_million_tokens: 4.50
audio_output_price_per_million_tokens: 12.00
summarize_model:
input_price_per_million_tokens: 0.25
output_price_per_million_tokens: 1.50
KeyTypeDefaultDescription
options.realtime_modelstring"gemini-3.8-live"Live model used for the session.
options.summarize_modelstringnoneModel used for the end-of-session summary. If omitted, the server reuses the summary model from ai_text_chat or meeting_summarizing, and otherwise writes a deterministic local summary with no LLM call.
options.api_versionstring"v1beta"Required for BidiGenerateContent.

The agent voice is selected per room with ai_agent_features.voice, not in server configuration. Server-side options only choose models and the API version.

Pricing Configuration​

Prices only drive the estimated_cost field on usage artifacts; they never change what Google bills you. Rates are per million tokens for LLM services and per hour or per million characters for speech.

transcription​

pricing.default.price_per_hour prices SPEECH_TRANSCRIPTION_USAGE from transcribed duration.

translation​

pricing.default.price_per_million_characters prices translation output.

tip

CHAT_TRANSLATION_USAGE is metered by character count, not tokens, and is priced with pricing.default.price_per_million_characters. Set an explicit rate or translation cost estimates stay at zero.

speech-synthesis​

pricing.default.price_per_million_characters prices SYNTHESIZED_SPEECH_USAGE from the number of characters sent to TTS.

ai_text_chat​

pricing.<chat_model> and pricing.<summarize_model> use input_price_per_million_tokens and output_price_per_million_tokens. Quote model names that contain dots or uppercase characters. A pricing.default entry applies to any model without its own block.

meeting_summarizing​

pricing.<summarize_model> (or pricing.default) uses input_price_per_million_tokens and output_price_per_million_tokens.

ai_agent​

Token pricing is multimodal. Put rates under pricing.realtime_model (or under the concrete model name) and under pricing.summarize_model:

KeyCovers
input_price_per_million_tokensText input
cached_input_price_per_million_tokensCached text input
cache_write_input_price_per_million_tokensText written to the prompt cache
audio_input_price_per_million_tokensParticipant speech sent to the model
audio_cached_input_price_per_million_tokensCached audio input
audio_cache_write_input_price_per_million_tokensAudio written to the prompt cache
image_input_price_per_million_tokensWhiteboard and share frames
image_cached_input_price_per_million_tokensCached image input
image_cache_write_input_price_per_million_tokensImages written to the prompt cache
output_price_per_million_tokensText output, including thinking tokens
audio_output_price_per_million_tokensSynthesized agent speech

Cache rates are optional. Missing cache rates are estimated at the uncached input rate. The usage fields these rates fill are listed in Summaries & Usage Artifacts.

Complete Configuration Example​

Every service on Gemini, one account:

insights:
enabled: true
providers:
google:
- id: "default-gemini-creds"
credentials:
api_key: "YOUR_GEMINI_API_KEY"
options:
api_version: "v1beta"

services:
transcription:
provider: "google"
id: "default-gemini-creds"
options:
model: "gemini-3.5-transcribe-live"
transcribe_mode: "VERBATIM"
api_version: "v1beta"
max_selected_trans_langs: 2
pricing:
default:
price_per_hour: 0.36

translation:
provider: "google"
id: "default-gemini-creds"
options:
translate_model: "gemini-3.1-flash-lite"
max_selected_trans_langs: 5
pricing:
default:
price_per_million_characters: 10.00

speech-synthesis:
provider: "google"
id: "default-gemini-creds"
options:
model: "gemini-3.8-flash-tts"
default_voice: "Kore"
tts_sample_rate: 16000
api_version: "v1beta"
pricing:
default:
price_per_million_characters: 16.00

ai_text_chat:
provider: "google"
id: "default-gemini-creds"
options:
chat_model: "gemini-3.1-flash-Preview"
summarize_model: "gemini-3.1-flash-lite"
context_window: 5
pricing:
default:
input_price_per_million_tokens: 0.30
output_price_per_million_tokens: 2.50

meeting_summarizing:
provider: "google"
id: "default-gemini-creds"
options:
summarize_model: "gemini-3.1-flash-lite"
pricing:
default:
input_price_per_million_tokens: 0.30
output_price_per_million_tokens: 2.50

ai_agent:
provider: "google"
id: "default-gemini-creds"
options:
realtime_model: "gemini-3.8-live"
summarize_model: "gemini-3.1-flash-lite"
api_version: "v1beta"
pricing:
realtime_model:
input_price_per_million_tokens: 0.75
cached_input_price_per_million_tokens: 0.075
cache_write_input_price_per_million_tokens: 0.9375
audio_input_price_per_million_tokens: 3.00
audio_cached_input_price_per_million_tokens: 0.30
audio_cache_write_input_price_per_million_tokens: 3.75
image_input_price_per_million_tokens: 1.00
image_cached_input_price_per_million_tokens: 0.10
image_cache_write_input_price_per_million_tokens: 1.25
output_price_per_million_tokens: 4.50
audio_output_price_per_million_tokens: 12.00
summarize_model:
input_price_per_million_tokens: 0.25
output_price_per_million_tokens: 1.50

Common Pitfalls​

  • Missing api_version: "v1beta". Gemini Live (transcription, ai_agent) and the Interactions API (speech-synthesis) fail on other versions. Set it on the account, on each of those services, or both.
  • Unquoted model names. Quote model names such as "gemini-3.1-flash-Preview", and always quote pricing keys that contain dots or uppercase characters.
  • Empty pricing block. Without rates the session still runs, but usage artifacts report zero cost.