Asterisk AI Voice Agent

Original Article ↗ Hacker News Discussion ↗

Perceived Uses & Abuse Potential

  • Many commenters immediately associate an Asterisk AI voice agent with more spam, scams, and “horrible customer support lines.”
  • A counter-use is proposed: running it as a honeypot to waste spam callers’ time (e.g., “Lenny”-style scripts).
  • Some fear businesses will use such systems to justify cutting remaining human support staff.

Caller Experience & Interface Preferences

  • Strong dislike for current voice-driven IVRs: unreliable in noisy environments, socially awkward to speak commands in public, and inconsistent UX.
  • Several prefer simple DTMF menus; many systems still accept keypad input even when pushing voice.
  • One practitioner notes that when IVR trees get complex, callers just mash “0” and demand a human; AI intent capture can be “less bad” than long menus in that scenario.

Legitimate Use Cases vs. “No Value”

  • People in the industry report large fractions of calls are simple or already self-serviceable (password resets, FAQ-style questions, “did you power cycle it?”, checking amenities) where an AI agent could help.
  • Concrete positive examples: dealership service line where an AI instantly books oil changes instead of putting callers on hold; potential hands-free tools for field staff or drivers.
  • Skeptics respond that if something is simple enough for AI, it should be a web/app flow instead; many only call when self-service has already failed or can’t be trusted.

Ethics, Deception & Expectations

  • Debate over background noise/“on brand” ambience: some see it as harmless branding; others call it deceptive, especially when it nudges callers to think it’s a human.
  • Strong sentiment that phone calls implicitly promise a human; using human-like voices without clear disclosure is framed as a “switcheroo” or scam.
  • Others push back: users mainly want to express needs in natural language; they don’t necessarily care about human vs machine if problems get solved.

Latency, Turn-Taking & Technical Challenges

  • Concern about the repo’s stated 2–3s latencies being “rage inducing.” Multiple commenters say SOTA can be sub-second, even ~300ms, with 2s+ causing hangups.
  • Practitioners report 500–1000ms as common and acceptable today, with major effort going into interruption handling and turn detection rather than raw speed.
  • Techniques discussed: streaming partial LLM output, chunking at punctuation, using fast TTS on short fragments, canceling/resyncing when the user interrupts, and blending with “thinking” sounds to mask latency.

Stacks, Deployment & VAD

  • Twilio integration: possible via SIP trunks or Twilio MediaStream WebSockets.
  • Pipecat mentioned as an open-source framework with many integrations (STT/LLM/TTS, turn detection model, state machine library); compared against proprietary players like Vapi/Retell/Sierra.
  • Deployment complexity is a pain point; some prefer Cloudflare Workers + Durable Objects with external STT (AssemblyAI/Deepgram with built-in VAD) and LLM/TTS for low-latency, low-cost scaling.
  • Discussion touches on where to keep conversation state (e.g., in Durable Objects) and compatibility with OpenAI-style realtime APIs.

Asterisk-Specific Notes & Nostalgia

  • Commenters are pleased to see Asterisk back on HN; default music-on-hold is widely recognized.
  • One person asks how to correlate CDRs with voicemail recordings for a unified dashboard; others suggest using channel vars, voicemail metadata files, or AGI/ARI logging.

Repository Style & Trust

  • Several note the GitHub repo looks heavily AI-generated: emoji-heavy headings, AI-like commit logs, Cursor traces.
  • This style triggers distrust for some: they assume docs may not be carefully reviewed, making them hesitant to rely on the project.

Overall Sentiment

  • Split roughly between:
    • Enthusiasm from builders and some users who see real operational value in handling simple, high-volume calls.
    • Deep skepticism from others who see little benefit to customers, expect more hostile support experiences, and are uncomfortable with human-mimicking automation on the phone.