Voice Support Bot
Real-time voice transcription and response
Combine a speech-to-text layer with an LLM to handle phone-based customer support calls end-to-end. The model transcribes speech, understands intent, queries your knowledge base, and generates natural-language responses fed back through text-to-speech.
GPT-6 Luna
for 1.5M tokens a month · 55% input / 45% output
Why this model
For this workload — 1.5M tokens a month, 55% of them input — it comes to about $0.42 a month, the lowest of the models suggested here.
Alternative models
Implementation tips
- 1
Use streaming responses to start text-to-speech synthesis while the LLM is still generating — this cuts perceived latency from 2–3 seconds to under 500ms for the first word.
- 2
Implement turn detection carefully: send partial transcripts to the LLM with a 'IS_COMPLETE' flag so the model can start processing before the caller finishes speaking on predictable phrases.
- 3
Keep the LLM's job narrow: handle intent detection and answer generation separately from speech processing, so you can swap in better ASR or TTS providers independently.