Performance

Metric	Value
Latency	~150–300ms chunk → text
CPU	~20–40% (1 core)
RAM	~500MB (Int8)
Accuracy	~95% clean Spanish
Bandwidth	~256KB/s PCM; VAD reduces ~60% silence

Optimizations

• Quantization — Int8 is ~2% less accurate than Float32 but ~2× faster
• VAD — Saves bandwidth and server CPU by skipping silence
• Model choice — Canary180M (180M params) faster; ParakeetTDT (0.6B) more accurate
• GPU — transcribe-rs with OrtAccelerator::Auto / Cuda
• Chunk size — Smaller frameSamples (256) = lower latency but more overhead

Measuring

Server logs chunk timing; client shows end-to-end in UI. Test with pre-recorded WAV piped over WebSocket for reproducible benchmarks.