streaming-stt-deepgram

Original：🇺🇸 English

Translated

Real-time streaming speech-to-text via Deepgram WebSocket API — sub-300 ms latency, Nova-2 model, speaker diarization, auto-reconnect.

6installs

Sourceframersai/agentos-skills

Added on2026-05-30

NPX Install

npx skill4agent add framersai/agentos-skills streaming-stt-deepgram

SKILL.md Content

View Translation Comparison →

Deepgram Streaming STT

Use this skill when the user needs real-time speech-to-text transcription with the lowest possible latency. Deepgram's WebSocket API provides sub-300 ms interim transcripts using the Nova-2 model.

Prefer this provider over file-based Whisper when the agent needs live voice input during a conversation, or when speaker identification (diarization) is required without a separate processing step.

Setup

Set

DEEPGRAM_API_KEY

in the environment or agent secrets store before starting a voice session.

Configuration

json

{
  "voice": {
    "stt": "deepgram"
  }
}

To enable diarization and keyword boosting:

json

{
  "voice": {
    "stt": "deepgram",
    "providerOptions": {
      "model": "nova-2",
      "diarize": true,
      "keywords": ["AgentOS:2"],
      "endpointing": 300
    }
  }
}

Provider Rules

Use
```
nova-2
```
as the default model — highest accuracy on Deepgram's current tier.
Enable
```
diarize: true
```
when the conversation involves multiple speakers; word-level
```
speaker
```
labels are included in the transcript events.
Tune
```
endpointing
```
(ms of silence before finalization) to balance responsiveness vs. over-splitting. Default 300 ms is suitable for most conversations.
The provider auto-reconnects on WebSocket drops using exponential back-off (100 ms → 5 s cap).
Use
```
providerOptions.keywords
```
to boost domain-specific terms (e.g. product names, abbreviations).

Events

Event	Description
`transcript`	Every hypothesis (interim + final)
`interim_transcript`	Non-final hypothesis
`final_transcript`	Stable, final hypothesis
`speech_start`	First non-empty word in an utterance
`speech_end`	Deepgram `speech_final` flag raised
`error`	Unrecoverable provider error
`close`	Session fully terminated

Examples

"Start a live voice session using Deepgram for transcription."
"Enable speaker diarization for this multi-person meeting transcription."
"Use Deepgram with keyword boosting for AgentOS and Wunderland terms."

Constraints

Requires
```
DEEPGRAM_API_KEY
```
. Free tier available at console.deepgram.com.
Audio must be streamed as PCM/WebSocket-compatible frames.
Diarization adds slight latency; disable if single-speaker performance is the priority.