Complete speech-to-text system with audio processing, multilingual support.

Request validation and parsing for incoming Twilio webhooks.

Implements the OpenAI Realtime API for voice conversation with audio streaming.

Orchestrates bidirectional audio streaming between telephony and AI providers.

Resamples 16-bit PCM audio to 8kHz telephony sample rate using linear interpolation.

LiveKit LLM plugin that delegates to ConversationEngine. This module provides VoicetestLLM, a LiveKit-compatible LLM that uses the same ConversationEngine as the test runner. This

Immutable transcript segment with start time, end time, and text content.

Indian-phone normalization. Shopify checkout returns phones in many shapes: +91 98765 43210 (formatted E.164) 919876543210 (no plus, with country code) 09876543210 (leading zero, n

Initiates outbound calls through a telephony provider, handling state management, configuring duration and transcript timers, and persisting call records.

Validates Twilio webhook signatures using HMAC-SHA1 with replay attack prevention through time-windowed request caching.

The canonical status vocabulary. call-api owns it; everything else is a dialect. The voice agent has its own words for how a call ended (`goal_achieved`, `no_audio`, `abandoned`, `

Do-Not-Disturb window enforcement for Indian telecom compliance. TRAI (Telecom Regulatory Authority of India) restricts commercial voice calls to 09:00–21:00 IST. Calling outside t

Pydantic model validating call initiation requests with a phone number field, extending a workflow initiate model for call orchestration.

zvonok voice agent — one goal-directed outbound PSTN call per job. Dispatched with metadata: {"number": "+34...", "goal": "...", "language": "en", "caller_id": "+44...", "max_durat

Ten protocol adapters (OpenAI Realtime, Deepgram Voice Agent, Twilio Media Streams, ElevenLabs, Gemini Live, Pipecat, Vapi, AssemblyAI, Cartesia, Native) handling wire-format negot

State machine managing lifecycle transitions for active phone calls.

Processor class for VibeVoice ASR models.

Controls voice session lifecycle, validates runtime values, grants, and handles turn state for voice interactions.

Abstract speech recognition interface providing flexible model integration and transcription methods.

Everything the agent says, and everything that checks it said it. Two kinds of text live here and they are not the same kind of thing: FIXED TEXT — the disclosure and the probe. Sp

Module exposing: TwilioService, TwilioPackage

Provides a provider-agnostic protocol and events for real-time speech-to-speech voice sessions, supporting multiple backends including Qwen, Gemini, and OpenAI Realtime.

TTS provider registry

Resolves and returns validated voice call configuration settings from environment or runtime sources.

Orchestrates VAD→STT→LLM→TTS pipeline with barge-in, silence detection, call recording, and ntfy push notifications. The core brain of a real-time voice AI.

Provides Twilio voice call integration handling outbound/inbound calls, webhook verification, TwiML responses, and call status normalization for telephony applications.

Records user audio via mic, encodes to webm/blob, handles permissions, uploads backend.

https://github.com/mohayonao/wav-decoder

Async text-to-speech backend wrapping the LuxTTS (ZipVoice) model for zero-shot voice cloning with 48kHz output. Reuse it as a pluggable TTS engine with device selection, cached vo

React list pane that loads effect presets through React Query and shows them as selectable cards. Includes creation and selection state via a Zustand store. Reuse it for preset bro

Unified TTS interface supporting ElevenLabs (cloud), OpenAI-compatible TTS servers (e.g. qwen3-tts-server with streaming PCM), and local TTS sidecars. Streaming vs buffered synthes

Voice participant builders for live calls. A VoiceParticipant produces a LiveKit AgentSession for one side of a call. Cascade (STT -> LLM -> TTS) is the only implementation today.

Kotlin call state graph managing call lifecycle transitions and signaling events.

Pure C G.711 mu-law/a-law encoder/decoder and L16 linear PCM utilities with canonical ITU algorithms. No external dependencies, fully unit-tested off-platform.

Collects and processes transient messages like screenshots, voice clips, and text for ingestion into agent memory pipelines.

Initiates Twilio voice calls, returns call/summary responses, and handles API errors via typed client.

Wire a room observer to transcribe one participant's published audio. Generalizes the per-worker self-observer: given the identity of the participant to watch and the role to attri

Centralized registry for discovering and registering voice platform integrations.

packages/apple-llm/ios wraps SpeechAnalyzer, AVSpeechSynthesizer, and NLContextualEmbedding behind AI SDK interfaces. Relevant to CobaltCall voice work on iOS with zero per-minute

Processes incoming call events, enforcing inbound policies (allowlist, open, disabled) and coordinating call lifecycle transitions including acceptance, rejection, and state persis

Custom exception for voice session broker operations, indicating failures during LiveKit dispatch creation or session orchestration.

Pure asyncio RTP protocol implementation with G.711 μ-law codec support and precise 20ms frame timing with drift correction. Complete RTP session management for bidirectional audio

Creates a voice call runtime for accepting inbound calls from telephony providers, configuring webhooks, tunnels, and lifecycle management.

Starts the HTTP server for the voice gateway, validates configuration, and manages lifecycle signals for graceful shutdown.

Tiptap extension providing AI-powered TTS suggestions and editing toolbar.

Service mode entrypoint for Pipecat voice pipeline. Runs an async HTTP server that accepts call requests, allowing the container to stay warm and handle multiple calls without cold

Main engine coordinator. ! ! Two pipelines: ! OUTGOING: Mic -> Deepgram(ru) -> Translate(ru->en) -> TTS(en) -> Speakers ! INCOMING: BlackHole 16ch -> Deepgram(en) -> Translate(en->

Composition root orchestrating perception, consciousness, skill surfaces, and output interfaces (TTS/STT/OBS) for an AI vtuber system.

Defines a Zod schema enumerating supported telecom providers (telnyx, twilio, plivo, mock) for a voice application.

Runs real-time voice sales conversations with configurable voice activity detection, multi-format audio support, and tool-enabled interactions.