Question 1

What do I need to build a voice assistant?

Accepted Answer

You need speech-to-text (STT) to understand what users say, a language model to generate responses, and text-to-speech (TTS) to speak the responses. Low-latency models like ElevenLabs Turbo v2.5 and Deepgram Nova 2 are ideal.

Question 2

What latency should I expect?

Accepted Answer

With real-time STT and streaming TTS, you can achieve response times under 1 second. Use models optimized for speed like ElevenLabs Turbo or OpenAI TTS-1 for the fastest responses. Deepgram offers sub-300ms transcription.

Question 3

How do I handle noisy environments?

Accepted Answer

Apply audio enhancement as a pre-processing step before STT. Use noise-robust models like Deepgram Nova 2. Implement push-to-talk as fallback. Consider wake word detection to filter unintended activation.