If you are building it yourself.

The long-standing open telephony engine. Still the default answer for bridging SIP into something you control.

Telephony platform used under the hood by a surprising share of the vendors in this directory.

Media Streams is the usual reference implementation for getting live call audio into your own runtime.

The session protocol underneath all of this. Worth an hour when call setup starts failing in ways no dashboard explains.

How real-time audio reaches browsers and apps, and the transport most agent frameworks build on.

The reference open speech recognition model. Not streaming out of the box, which is the first thing you will discover when you try to use it on a live call.

Small, fast voice activity detection. Most self-built stacks end up using this or something very like it to decide when the caller has stopped speaking.

Toolkit for training and fine-tuning speech models, including multilingual and Indic work. The route if you need a model tuned to your own audio.

Offline recognition with small models across many languages. Useful when data cannot leave the building.

Comparative word error rates across open models. Read it knowing the evaluation audio is nothing like a phone call.

Streaming recognition API with telephony-oriented models. Their docs on endpointing and interim results are worth reading whichever provider you pick.

Recognition API with real-time streaming and speaker labelling.

Streaming, tool use and prompt caching — the three things that decide whether a model is usable inside a voice latency budget.

Includes realtime and speech-to-speech APIs, which collapse several stack layers into one call and change the cost model entirely.

High-throughput serving for open models. The practical answer if you are self-hosting and care about time-to-first-token.

Fast local neural speech synthesis, including many languages. Runs on modest hardware, which matters if you are synthesising at volume.

Training and inference toolkit for speech synthesis. Archived upstream but still the most complete open reference for how these systems fit together.

Streaming synthesis and voice cloning. Test any voice you like over an actual phone line before committing to it.

Low-latency synthesis, explicitly optimised for the real-time case rather than for offline narration.

Framework for real-time voice agents — transport, turn detection and the conversation loop. The clearest open reference for what orchestration actually involves.

Pipeline framework for voice and multimodal agents. Good for understanding how the pieces are wired before you commit to a vendor.

Open-source CPaaS you can self-host. Relevant when data residency or carrier control rules out a hosted platform.

India's telecom regulator. The source for commercial communication rules, DLT registration and do-not-disturb obligations on outbound calling.

India's data protection framework, which governs call recordings, transcripts, retention and deletion. Read it before designing your storage, not after.