Voice Telephony

• Tool IDs: telnyxVoiceCall, telnyxVoiceProvider
• Secrets: telnyx.apiKey, telnyx.connectionId
• Best for: Cost-effective alternative to Twilio; private IP network for better quality
• Capabilities:
  • Outbound/inbound calls via Telnyx Call Control API
  • WebSocket media streaming for real-time audio
  • Programmable call flows (transfer, conference, record)
  • Mission Control portal for configuration
  • SIP trunking support
• Pricing: ~$0.007/min outbound US (roughly half of Twilio); phone numbers from $1/mo

Plivo

• Tool IDs: plivoVoiceCall, plivoVoiceProvider
• Secrets: plivo.authId, plivo.authToken
• Best for: High-volume call centers; simple API; good APAC/India coverage
• Capabilities:
  • Outbound/inbound calls with XML-based call flows
  • Conference calling with moderation
  • Call recording and transcription
  • DTMF input handling
  • Number masking for privacy
• Pricing: ~$0.010/min outbound US; competitive international rates

STT (Speech-to-Text) Streaming Providers

Deepgram Streaming STT

• Extension: streaming-stt-deepgram
• Secrets: deepgram.apiKey
• Best for: Fastest real-time transcription; best accuracy for conversational speech
• Features:
  • WebSocket streaming with <300ms latency
  • Multiple models: Nova-2 (general), Enhanced (noisy), Base (fastest)
  • Interim results for responsive UX
  • Punctuation, diarization, smart formatting
  • 30+ languages
• Recommendation: Default choice for production voice apps

Whisper Streaming STT

• Extension: streaming-stt-whisper
• Secrets: openai.apiKey (for API) or none (for local)
• Best for: Self-hosted/local deployment; highest accuracy for non-English languages
• Features:
  • OpenAI Whisper model (local or API)
  • Chunk-based streaming (not true real-time, ~1-2s chunks)
  • 97+ languages with strong multilingual performance
  • Local mode: no API costs, requires GPU for real-time
• Recommendation: Use when Deepgram is unavailable or for local/offline deployments

Google Cloud STT

• Extension: google-cloud-stt
• Secrets: google.serviceAccountJson
• Best for: Enterprise Google Cloud integration; medical/legal domain models
• Features:
  • Streaming recognition via gRPC
  • Multiple models: default, phone_call, video, medical_conversation
  • Speaker diarization (who said what)
  • Word-level confidence and timing
  • Automatic punctuation

Vosk (Offline)

• Extension: vosk
• Secrets: None
• Best for: Fully offline/airgapped deployments; edge devices
• Features:
  • Local models, no internet required
  • Lightweight enough for Raspberry Pi
  • 20+ language models available
  • Speaker identification
• Recommendation: Use for privacy-critical or offline scenarios

TTS (Text-to-Speech) Streaming Providers

ElevenLabs Streaming TTS

• Extension: streaming-tts-elevenlabs
• Secrets: elevenlabs.apiKey
• Best for: Most natural-sounding voices; voice cloning; emotional expression
• Features:
  • WebSocket streaming with ~200ms time-to-first-byte
  • 30+ pre-built voices, custom voice cloning
  • Adjustable stability, similarity, style
  • 29 languages with accent control
  • SSML support
• Recommendation: Default choice for the best voice quality

OpenAI Streaming TTS

• Extension: streaming-tts-openai
• Secrets: openai.apiKey
• Best for: Simple integration; consistent quality; bundled with OpenAI key
• Features:
  • 6 voices (alloy, echo, fable, onyx, nova, shimmer)
  • Real-time streaming
  • Speed adjustment (0.25x to 4.0x)
  • HD quality option
• Recommendation: Use when already using OpenAI for LLM; quality is good but fewer customization options

Amazon Polly

• Extension: amazon-polly
• Secrets: aws.accessKeyId, aws.secretAccessKey
• Best for: AWS ecosystem; SSML control; Neural and Standard voices
• Features:
  • Neural voices (natural) and Standard voices (cheaper)
  • Full SSML support (pauses, emphasis, phonemes)
  • 60+ voices across 30+ languages
  • Newscaster and Conversational styles
• Recommendation: Use for AWS-native deployments or when SSML control is critical

Google Cloud TTS

• Extension: google-cloud-tts
• Secrets: google.serviceAccountJson
• Best for: Google Cloud integration; WaveNet voices; Studio voices
• Features:
  • WaveNet voices (very natural), Standard, Neural2, and Studio
  • SSML support with audio effects
  • 50+ languages, 380+ voices
  • Audio profiles (telephony, headphone, smart speaker)

Piper (Offline)

• Extension: piper
• Secrets: None
• Best for: Offline/local TTS; edge deployment; no API costs
• Features:
  • ONNX-based, runs entirely local
  • 100+ voices across 30+ languages
  • Fast inference on CPU
  • Configurable quality levels
• Recommendation: Use for offline deployments or when API costs are a concern

Voice Pipeline Architecture

A complete voice pipeline connects these components:

Microphone → VAD → STT Provider → LLM → TTS Provider → Speaker
                                    ↑
                              Memory/Context

Pipeline Components

1. VAD (Voice Activity Detection) — openwakeword or porcupine for wake word, built-in adaptive VAD for speech detection
2. STT — converts speech to text in real-time
3. LLM — processes the transcribed text and generates a response
4. TTS — converts the LLM response back to speech
5. Audio Transport — WebRTC, WebSocket, or telephony media stream

Provider Selection Guide

IVR (Interactive Voice Response) Setup

1. Provision a phone number from Twilio, Telnyx, or Plivo
2. Configure inbound webhook URL pointing to your AgentOS endpoint
3. Wire the voice pipeline: STT → LLM → TTS
4. Define call flow states: greeting, menu, transfer, voicemail
5. Handle DTMF input for numeric menu selections
6. Set fallback to human operator for unhandled cases
7. Enable call recording for quality assurance (with consent disclosure)

Best Practices

• Latency budget — total round-trip (STT + LLM + TTS) should be under 2 seconds for natural conversation
• Interruption handling — enable barge-in so users can interrupt the TTS playback
• Fallback chain — if primary STT/TTS fails, fall back to a secondary provider
• Cost management — use Vosk/Piper for development/testing; paid providers for production
• Audio quality — use 16kHz 16-bit mono PCM for telephony; 44.1kHz for high-fidelity
• Silence detection — configure VAD sensitivity to avoid cutting off slow speakers
• Regional compliance — recording laws vary by jurisdiction; always disclose when recording