Gemini

Gemini is a speech-to-text plugin. It uses Gemini Transcribe with your Gemini API key. Live models stream partial and final transcripts; gemini-3.5-transcribe transcribes each turn once it ends and needs voice activity detection. For Google Cloud Speech-to-Text, see Google STT.

Setup

Set your Gemini API key in the worker environment. Generate a key from the Gemini site:

export GOOGLE_API_KEY=<key>

Usage

Import the plugin and pass it to the pipeline’s stt slot.

from zeroruntime.plugins import GeminiSTT

stt = GeminiSTT()

# Pipeline(stt=stt, ...)

GeminiSTT ships in the Python SDK only. The Node JS SDK does not include this plugin yet.

Parameters

Constructor parameters for GeminiSTT.

Parameter	Type	Default	Description
`api_key`	`str`	`None`	Gemini API key. Falls back to the `GOOGLE_API_KEY` environment variable when unset.
`model`	`str`	`"gemini-3.5-transcribe-live"`	Transcribe model, e.g. “gemini-3.5-transcribe-live” for streaming or “gemini-3.5-transcribe” for per-turn transcription.
`languages`	`Union[str, list[str], None]`	`"en-US"`	BCP-47 language code or list of codes, e.g. “en-US” or [“hi-IN”, “en-IN”]. Set to None to auto-detect the language.
`custom_vocabulary`	`list[str] \| None`	`None`	Up to 1,000 words or phrases, such as names and product terms, to recognize more reliably.
`interim_results`	`bool`	`True`	Emit partial transcripts while the user is still speaking. Live models only.
`mode`	`Literal['verbatim', 'smart'] \| None`	`None`	”verbatim” keeps the exact words; “smart” removes fillers and disfluencies and adds formatting.
`start_of_speech_sensitivity`	`Literal['high', 'low'] \| None`	`None`	”high” detects the start of speech sooner; “low” ignores more background noise. Live models only.
`end_of_speech_sensitivity`	`Literal['high', 'low'] \| None`	`None`	”high” ends a turn sooner after the speaker stops, for faster final transcripts; “low” waits longer. Live models only.
`prefix_padding_ms`	`int \| None`	`None`	Milliseconds of audio kept from just before detected speech, so first syllables are not cut off. Live models only.
`silence_duration_ms`	`int \| None`	`None`	Milliseconds of silence before a turn is finalized. Lower is faster but may split sentences at pauses. Live models only.
`input_sample_rate`	`int \| None`	`None`	Sample rate of incoming audio in Hz. Leave unset to detect it automatically.
`input_channels`	`int \| None`	`None`	Number of channels in the incoming audio. Leave unset to detect it automatically.

Import paths

SDK	Import	Constructor
Python	`from zeroruntime.plugins import GeminiSTT`	`GeminiSTT()`

Transcribed text passes to the LLM once turn detection decides the caller has finished speaking.

Inference

This provider also runs through the Inference Gateway on your Zero Runtime auth token, with no Gemini API key. Import GeminiSTT from zeroruntime.inference instead of zeroruntime.plugins.

Get started

Quickstarts

Build

Monitor

Deploy

Plugins

Setup

Usage

Parameters

Import paths

Inference

​Setup

​Usage

​Parameters

​Import paths

​Inference

Setup

Usage

Parameters

Import paths

Inference