Conversation Relay: Build a Text-In, Text-Out Voice Bot on an Indian Number (2026)

Conversation Relay lets you run a voice bot that only handles text. When someone calls your Indian number, Edesy converts their speech to text and sends it to your bot over a WebSocket; your bot replies with text and Edesy speaks it, handling barge-in and keypresses. The protocol is Twilio ConversationRelay-compatible, so existing bots move over unchanged.

Put your bot on a number: Open the Edesy voice portal. Buy a number, choose the text-in, text-out bot target and paste your wss:// URL.

Why a text-only bot

Running speech recognition and text-to-speech yourself means picking vendors, handling audio formats, tuning voice activity detection and managing interruptions. That is real work. If what you want to build is the conversation itself (an LLM with your prompts and tools), Conversation Relay lets you skip the audio layer.

	Conversation Relay	Media Streams
Your bot receives	Text transcripts	Raw L16 PCM audio
Your bot sends	Text (and play, sendDigits, end)	Audio
Who runs STT and TTS	Edesy	You
Barge-in	Edesy detects, tells your bot	You detect, send `killAudio`
Best for	LLM bots, fast builds, Twilio relay ports	Custom pipelines, Pipecat, specific voices

For the audio route, see WebSocket audio streaming.

How a call flows

A caller dials your Edesy number.
Edesy opens a WebSocket to your bot and sends setup with callSid, from, to, direction and your customParameters.
If you set a welcome_greeting, Edesy speaks it. Otherwise your bot speaks first.
When the caller finishes speaking, your bot gets prompt with voicePrompt (the transcript).
Your bot streams back text tokens, with last: true on the final one.
Edesy speaks the reply. If the caller talks over it, your bot gets interrupt.
The bot ends the call with end, optionally with handoffData.

Messages

You receive	When
`setup`	Once, before anything is spoken
`prompt`	The caller finished an utterance (`voicePrompt`, `lang`)
`interrupt`	The caller talked over the bot (`utteranceUntilInterrupt`)
`dtmf`	A keypress, when DTMF detection is on
`error`	A platform-side problem; informational

You send	Effect
`text`	Speak it; stream tokens, `last: true` on the final chunk
`play`	Play an audio file, then continue
`sendDigits`	Send DTMF out of the call (`w` is a half-second wait)
`language`	Switch TTS and/or transcription language
`end`	End the call; `handoffData` is recorded

Settings on the number

Setting	Default	What it does
`welcome_greeting`	none	Spoken on connect
Voice	platform default	TTS voice ID
Language	`en-IN`	TTS and transcription language
`interruptible`	on	Whether the caller can talk over the bot
`min_bargein_word_count`	1	Words needed to interrupt; 0 means any sound, so a cough can stop the bot
`dtmf_detection`	off	Keypresses interrupt the bot and arrive as `dtmf`
`custom_parameters`	none	Passed verbatim in `setup`

Put the tenant, campaign or store ID your bot needs in custom_parameters, so the bot doesn't have to look it up by phone number.

Three things that catch people out

Set last: true. It flushes the speech engine. Without it, the end of each reply stays buffered and the bot sounds like it swallowed its last sentence.

Handle interrupt. When the caller barges in, unplayed speech is discarded, so your history contains words the caller never heard. Trim the last bot turn to utteranceUntilInterrupt, or the conversation drifts.

Accept the WebSocket upgrade quickly. The connection attempt is time-bounded, and a fallback message plays if it expires. Authenticate during the upgrade, not after.

Set it up

Run your bot. Edesy's Node sample bot listens on ws://0.0.0.0:8787/relay; expose it as wss:// in production.
Preflight it:

curl -X POST https://voice-api.edesy.in/v1/relay-test \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  -d '{"relay_url":"wss://bot.example.com/relay","prompt":"what are your hours","seconds":5}'

In the portal, open your number, choose the text-in, text-out bot under "When a call comes in", paste the URL, set the greeting and voice, and save. The API equivalent is PATCH /v1/phone-numbers/:id/routing with inbound_target_type: "relay" and inbound_relay_url.
Call the number.

What you can see after a call

The call timeline (GET /v1/calls/:sid/events) shows call.relay.connected, call.relay.failed (with a reason), call.relay.handoff and call.relay.usage, which reports TTS characters, STT seconds, turns and interruptions.

Porting from Twilio

Point an Edesy number at the same WebSocket URL your Twilio ConversationRelay bot uses. The message types and fields are the same. What changes is everything around it: Indian numbers bought after KYC, a prepaid INR wallet with GST invoices, and a 60-second pulse with unanswered calls free. See the Twilio migration guide for the REST side.

Cost

Calls are billed at Rs 0.50 per minute on a 60-second pulse, unanswered calls free, plus number rental (Rs 499 per month for mobile or 92-series). Prices exclude 18% GST. Speech usage for each call is reported in call.relay.usage; see Voice API pricing for current rates.

More on the Conversation Relay page, the Voice API hub, the call routing options and the Voice API docs.

Frequently asked questions

What is Conversation Relay?

A way to put a text bot on a phone call. Edesy transcribes what the caller says and sends your bot text over a WebSocket; your bot sends text back and Edesy speaks it. Edesy also handles barge-in and keypresses, so your bot never touches audio.

Is Edesy Conversation Relay compatible with Twilio ConversationRelay?

Yes. It uses Twilio's ConversationRelay protocol (setup, prompt, interrupt, dtmf, text, play, sendDigits, language, end), so a bot written for Twilio works unchanged when you point an Edesy number at it.

Can the caller interrupt the bot?

Yes, by default. interruptible is on, and min_bargein_word_count (default 1) sets how many words the caller must say before the bot stops. Your bot receives an interrupt message with roughly what the caller heard.

Can the bot switch language mid-call?

Yes. Send a language message with ttsLanguage and/or transcriptionLanguage. The default language on a new relay number is en-IN.

How do I hand the call to a human?

Send an end message with handoffData. The call ends on the relay and the handoff data is recorded as a call.relay.handoff event that your system can act on.

How do I test my bot without making a call?

Use POST /v1/relay-test with relay_url, a sample prompt and seconds. It connects from Edesy's network, runs one turn, and reports DNS, TCP, TLS, upgrade and protocol stages, including whether your reply set last: true.