STT (Speech-to-Text) is the agent's ears — it converts the caller's audio into text so the LLM can reason about it. Shopbell's managed STT default is Deepgram Nova-2 — sub-200ms latency, 90%+ accuracy on most US English accents, included in $99/mo. This guide covers the default and BYOK options.
Default — Deepgram Nova-2
Latency: under 200ms for first hypothesis. Accuracy: 90%+ on US English in normal noise conditions. Languages: English (US, UK, AU), Spanish (Spain, LATAM), French, German, Italian, Portuguese, plus 20+ more. Included in your $99/mo plan. Most Self-Serve DIY customers use Deepgram without configuration — it's already attached to your first workflow.
Pattern — Handling accents and noise
Deepgram handles accents well but call centers, speakerphones, and cars are harder. If accuracy is poor, ask the caller to repeat or move to a quieter spot. The agent's system prompt should say 'If you did not hear correctly, ask the caller to repeat — never guess.'
BYOK — Whisper / OpenAI Audio / AssemblyAI
If you run a high-volume call center and want a hosted Whisper deploy, or you have a paid AssemblyAI plan, configure it as a BYOK STT provider. Click Model Configurations → New → STT → paste provider API key. Most DIY customers do not need this — Deepgram is enterprise-grade.
Frequently Asked Questions
How accurate is Deepgram?
Tested at 91-94% Word Error Rate on clean US English, 86-89% on accented or noisy calls. WER below 10% feels indistinguishable from a human transcriber for short utterances.
Does STT latency affect call latency?
Yes — STT latency is part of the round-trip time the caller experiences. Deepgram's sub-200ms is best-in-class. Whisper on CPU is 500-1500ms — only use with a fast GPU-backed deployment for production.
What if the transcriber misunderstands?
The agent gets the wrong text and might respond weirdly. The agent's system prompt should include 'If the caller's response doesn't make sense in context, ask them to repeat.' The caller rarely notices — they just rephrase.