STT Configuration¶
File: configs/stt.yaml
Command: kenzy-stt [config_path]
The STT service accepts POST requests with base64-encoded PCM audio and returns a transcript. It is built on faster-whisper, a CTranslate2-optimized implementation of OpenAI Whisper.
Pulled from the server
kenzy-stt pulls this config from the server at boot — it discovers the server via mDNS (or KENZY_SERVER_URL) and blocks until it answers, so start the server first. Edit it from the dashboard's Services tab (writes configs/services/stt.yaml on the server and restarts the service). Passing an explicit path (kenzy-stt configs/stt.yaml) loads locally instead — a dev/offline escape hatch. See central config for backend services.
Full reference¶
| Key | Default | Description |
|---|---|---|
host |
"127.0.0.1" |
Bind address |
port |
8767 |
HTTP port |
log_level |
"info" |
What the service prints to its console |
log_capture_level |
"debug" |
How deep the dashboard log viewer can see (trace/debug/…), independent of log_level |
Whisper model¶
| Key | Default | Description |
|---|---|---|
whisper.model |
"tiny" |
Model size: tiny, base, small, medium, large-v2, large-v3. Larger models are more accurate but slower and need more RAM. |
whisper.device |
"cpu" |
Inference device: cpu or cuda |
whisper.compute_type |
"int8" |
Quantisation: int8 (fastest on CPU), float16 (GPU), float32 (highest quality) |
whisper.language |
"en" |
Language code (e.g. "en", "fr"), or null for auto-detect |
Model size guide¶
| Model | Size | Relative speed | Notes |
|---|---|---|---|
tiny |
~75 MB | Fastest | Good for fast hardware or simple commands |
base |
~145 MB | Fast | Better accuracy, still CPU-friendly |
small |
~460 MB | Moderate | Good balance for a dedicated CPU server |
medium |
~1.5 GB | Slow on CPU | Recommended with a GPU |
large-v3 |
~3 GB | Slow | Best accuracy; GPU strongly recommended |
Run STT off the node
Don't run STT on a room-node board (Orange Pi Zero 3 / Raspberry Pi 3–5) — run it on a more powerful server and point stt.url in server.yaml at it. The tiny or base model on a modern x86 CPU gives acceptable latency.
Example¶
host: "127.0.0.1"
port: 8767
whisper:
model: "base"
device: "cpu"
compute_type: "int8"
language: "en"