Speech Recognition (STT)

Deepgram Nova-2 is our recommendation for live voice calls - it offers the lowest latency (<300ms) and excellent accuracy. Google Cloud is best for multilingual scenarios, especially Indian languages. Azure is great for enterprise customization.

Audio from the phone call is streamed to the STT provider in real-time. Transcription results stream back as the caller speaks, with partial results updating until the speaker finishes. This enables natural back-and-forth conversation.

Yes, we support Hindi, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Bengali, and more. Google Cloud has the broadest Indian language support, while Deepgram offers excellent Hindi recognition.

Modern STT providers are trained on diverse accents. Deepgram and Google handle Indian English accents very well. For specialized domains or heavy accents, Azure allows custom model training on your audio data.

All providers include noise suppression designed for phone-quality audio. They handle office noise, street noise, and multiple speakers. Deepgram's Nova-2 is particularly strong at isolating the primary speaker.

Yes, if you have a custom model on Azure or Google, you can configure our platform to use it. This is useful for domain-specific vocabulary (medical terms, product names, etc.) that general models might miss.

STT costs are based on audio duration processed. Deepgram and Google are competitively priced at ~$0.004-0.006 per 15 seconds. We pass through provider costs transparently as part of your usage billing.

Yes, you can configure backup STT providers. If Deepgram is slow or unavailable, automatically fall back to Google or Azure. This ensures calls continue without interruption.