On-deviceKokoroWhisper
Free cloud voicesMicrosoft EdgeGemini 3.8 Live
Paid realtimexAIgpt-realtime-2.1
gpt-realtime-2.1-miniGemini 3.8 LiveElevenLabs
We tested the whole voice stack
On-device, free cloud and paid realtime voices, each tested in live conversations.
Read the full notes
We tested on-device models (Kokoro for speech, Whisper for transcription), free cloud voices, and paid realtime engines from xAI, OpenAI, Google and ElevenLabs. Every engine went through live, hands-on conversations while we logged each failure: stalled responses, dropped replies, mid-sentence cut-offs. Lower-cost options had real limits. We chose the one that proved the most reliable and the most human, with the data to back it up. It powers our own receptionist today.
