streaming-stt-whisper

Original：🇺🇸 English

Translated

Chunked sliding-window streaming speech-to-text via OpenAI Whisper HTTP API — compatible with local Faster-Whisper, Groq, and OpenRouter endpoints.

2installs

Sourceframersai/agentos-skills

Added on2026-05-30

NPX Install

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

SKILL.md Content

View Translation Comparison →

Whisper Chunked Streaming STT

Use this skill when OpenAI Whisper is the preferred STT provider, especially when a single API key should cover both LLM and speech. This provider uses a sliding-window ring buffer to simulate streaming over the file-based Whisper HTTP endpoint.

Prefer this over the Deepgram adapter when real-time WebSocket streaming is not required, when the user wants to route through a local Whisper-compatible server (e.g. Faster-Whisper, Groq), or when OpenAI is the only configured provider.

Setup

Set

OPENAI_API_KEY

in the environment or agent secrets store. For local endpoints, override

baseUrl

in

providerOptions

.

Configuration

json

{
  "voice": {
    "stt": "whisper"
  }
}

For a local Faster-Whisper endpoint:

json

{
  "voice": {
    "stt": "whisper",
    "providerOptions": {
      "model": "whisper-1",
      "language": "en",
      "baseUrl": "http://localhost:8000"
    }
  }
}

Provider Rules

Audio is accumulated in a 1 s sliding window with 200 ms overlap to avoid word boundary clipping.
The previous chunk transcript is forwarded as
```
prompt
```
to the next request for cross-chunk continuity.
On fetch failure the provider emits
```
error
```
and continues — no session crash.
Use
```
language
```
to force a specific language code (BCP-47); omit for automatic detection.
Compatible with any OpenAI
```
/v1/audio/transcriptions
```
-compatible server.

Events

Event	Description
`interim_transcript`	Emitted after each chunk is transcribed
`final_transcript`	Emitted after flush() completes
`speech_start`	RMS energy crossed threshold
`speech_end`	RMS energy dropped below threshold
`error`	Fetch failure (session continues)
`close`	Session fully terminated

Examples

"Use Whisper for live speech transcription during our voice session."
"Transcribe my speech through a local Faster-Whisper server."
"Use OpenAI for both the LLM and the STT provider."

Constraints

Requires
```
OPENAI_API_KEY
```
or a compatible local endpoint via
```
providerOptions.baseUrl
```
.
Latency is higher than native WebSocket providers (Deepgram) due to HTTP chunking overhead.
Speaker diarization is not natively supported; use the
```
diarization
```
extension for post-processing.