v2.0.0

rapidaai/voice-aiv2.0.0Feb 24, 2026by iamprashant

AI Summary

Major release introducing telephony reliability improvements with a unified channel architecture across Twilio, Vonage, Exotel, Asterisk, and SIP. Includes native SIP integration, Asterisk/AudioSocket support, WebRTC channel for browser-based voice, and deterministic audio framing.

Key Highlights

Unified channel architecture across all telephony providers (Twilio, Vonage, Exotel, Asterisk, SIP)
Native SIP stack with RTP handling, SDP negotiation, and session management
WebRTC channel with Opus codec support for browser-based voice
Deterministic 20ms audio framing with zero per-frame heap allocations
LLM Text Aggregator reduces first-word latency with sentence-boundary aggregation

Breaking Changes

New `call_contexts` table required (migration 000005)
`recording_url` split into `assistant_recording_url` + `user_recording_url` (migration 000006)

New Features

SIP Integration with native stack
Asterisk/AudioSocket integration
WebRTC Channel for browsers
Per-speaker recordings (assistant and user)
Atomic interruption with ClearOutputBuffer
Channel UUIDs for end-to-end transfer/hangup operations

Full Release Notes

## What's Changed

 ### Telephony: Rebuilt from the Ground Up
  - Unified channel architecture shared across Twilio, Vonage, Exotel, Asterisk, and SIP
  - Interruptions, end-of-call signals, and transfer/hangup events handled consistently across all providers
  - New \`call_contexts\` table persists call state — async provider callbacks resolve correctly even after call ends
  - Channel UUIDs propagate end-to-end for reliable transfer and hangup operations

  ### New: SIP Integration
  Full native SIP stack with RTP handling, SDP negotiation, port allocator, and session management.

  ### New: Asterisk / AudioSocket
  Native integration with Asterisk via AudioSocket and WebSocket. Inbound and outbound call flows tested.

  ### New: WebRTC Channel
  Browser-based voice with Opus codec support and gRPC signalling, sharing the same hardened base as telephony.

  ### Audio Pipeline: Deterministic Framing
  - Exact 20 ms output frames with zero per-frame heap allocations
  - Atomic interruption — \`ClearOutputBuffer\` drains buffers and signals output writer instantly
  - Per-speaker recordings split into \`assistant_recording_url\` + \`user_recording_url\`

  ### LLM Text Aggregator
  Sentence-boundary aggregation between LLM stream and TTS — reduces first-word latency with configurable delimiters and clean context-switch flush.

  ### Test Coverage
  - 31 \`BaseStreamer\` unit tests
  - Full telephony provider test suite (Twilio, Vonage, Exotel)
  - Transformer tests for AssemblyAI, Azure, Cartesia, Deepgram, ElevenLabs, Google, Resemble, Sarvam
  - LLM text aggregator: 972 lines of unit tests + 381 lines of benchmarks

  ### Bug Fixes
  - Google TTS stale response fix for outputs > 5 sentences
  - AgentKit executor stability fixes
  - First-token response time now tracked in LLM telemetry
  - MCP tool support for agent tool invocations

  ## Breaking Changes / Migrations
  | Migration | Change |
  |---|---|
  | \`000005\` | New \`call_contexts\` table required |
  | \`000006\` | \`recording_url\` split into \`assistant_recording_url\` + \`user_recording_url\` |