Smart Speaker Voice Interaction Platform: Concepts, Processes, and Testing Metrics

Introduction With the rapid development of AI, intelligent voice interaction has become a core technology for smart speakers. A typical smart speaker relies on a voice interaction platform that includes wake‑word detection, ASR, NLP, and skill execution.

Wake‑Word (Voice Trigger) Wake‑word activation can be triggered by a button or by speaking a predefined phrase (e.g., "Hey Siri", "OK Google"). Shorter phrases increase false‑wake rates, so industry practice favors 3‑4 syllable wake words. Common mitigation methods include cloud‑side double verification and time‑of‑day sensitivity adjustments.

ASR (Automatic Speech Recognition) ASR converts spoken audio into text through two stages: training and decoding. Training builds acoustic and language models (e.g., CTC, CRNN) using large labeled datasets. Decoding uses these models to generate text, often enhanced with hot‑word boosting. Supporting technologies such as beamforming, noise reduction, acoustic echo cancellation (AEC), and voice activity detection (VAD) improve robustness.

NLU (Natural Language Understanding) NLU parses user utterances into domain, intent, and slot structures (e.g., "set an alarm for 8 am tomorrow" → domain: alarm, intent: create, slot: 8 am). Accuracy and recall are the primary evaluation metrics, reflecting the proportion of correctly identified intents.

Skill (Application Logic) A skill processes the intent returned by NLU and generates a response. Testing principles for voice skills include diversifying replies, placing critical information later (recency effect), and keeping responses concise yet complete when needed.

TTS (Text‑to‑Speech) TTS synthesizes speech from text using either concatenative (high naturalness, high cost) or parametric (lower cost, improving quality) methods. Evaluation includes subjective tests (MOS, ABX) and objective metrics (RMSE, latency, memory/CPU usage, crash rate).

Testing Metrics Overview Key metrics across the pipeline are wake‑word rate, false‑wake rate, wake‑word length, response time, power consumption, word error rate (WER) for AEC, and both subjective and objective scores for TTS. These indicators help assess the overall user experience and system reliability.