stt

name	stt
description	Transcribe audio files to text. Use when you need to convert speech/audio to text.

Speech-to-Text (STT)

Transcribe audio files to text using the stt CLI.

Usage

# Transcribe an audio file
stt /path/to/audio.wav

# Transcribe with language hint
stt --language de /path/to/audio.ogg

# Transcribe from a URL (downloads first)
stt https://example.com/recording.mp3

Supported Formats

wav, mp3, ogg, m4a, aac, flac, webm, mp4 (audio track)

Direct API Usage

The STT endpoint is Whisper-compatible (OpenAI /v1/audio/transcriptions format):

curl -X POST "$STT_URL" \
  -F "file=@/path/to/audio.wav" \
  -F "response_format=json" \
  -F "language=de"

Response: {"text": "transcribed text here"}

The URL is resolved from: ATLAS_STT_URL env → STT_URL env → config.yml (stt.url) → default.

Limitations

• Long files (>2 min) are split into 120s chunks with 5s overlap — minor artifacts at chunk boundaries are possible
• Accuracy depends on audio quality; background noise reduces quality significantly
• The model runs on CPU (int8 quantized) — expect ~5-15s processing per minute of audio
• Single-speaker optimized; multi-speaker conversations may lose speaker attribution
• No diarization (speaker identification) — only raw text output

Signal Integration

Audio messages received via Signal are automatically transcribed before reaching the agent — no manual action needed.