Evaluation¶

Record provider/model versions, dataset version, environment, WER, RU/EN code-switch WER, terminology accuracy, STT latency, LLM TTFT, TTS TTFA, and end-to-end latency.

Deterministic tests validate contracts and lifecycle behavior without proving speech quality. Do not report STT, LLM, or TTS acceptance until configured real providers process real audio end to end. Keep private recordings out of the repository unless explicitly approved.