streaming-stt-whisper

name	streaming-stt-whisper
description	Chunked sliding-window streaming speech-to-text via OpenAI Whisper HTTP API
category	voice

Whisper Chunked Streaming STT

Streaming speech-to-text using OpenAI's Whisper model via the /v1/audio/transcriptions HTTP API. Audio is accumulated in a sliding-window ring buffer and sent as 1-second WAV chunks with 200 ms overlap for continuity.

Setup

Set OPENAI_API_KEY in your environment or agent secrets store.

Features

• Works with any OpenAI-compatible Whisper endpoint (e.g. local Faster-Whisper, Groq, OpenRouter)
• Sliding-window ring buffer: 1 s chunks, 200 ms overlap to prevent word boundary clipping
• Inline WAV encoding — zero runtime dependencies (no ws, no native binaries)
• Previous chunk transcript forwarded as prompt for cross-chunk continuity
• Simple RMS energy detector emits speech_start / speech_end events
• On fetch failure, emits error and continues processing — no session crash

Configuration

In agent.config.json:

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

Provider-specific options via providerOptions:

{
  "voice": {
    "stt": "whisper",
    "providerOptions": {
      "model": "whisper-1",
      "language": "en",
      "baseUrl": "https://api.openai.com"
    }
  }
}

Events

Event	Payload	Description
interim_transcript	TranscriptEvent	Emitted after each chunk is transcribed
final_transcript	TranscriptEvent	Emitted after flush() completes
speech_start	—	RMS energy crossed threshold (0.01)
speech_end	—	RMS energy dropped below threshold
error	Error	Fetch failure (session continues)
close	—	Session fully terminated