Voice AI: infrastructure, platforms, and vertical applications

Voice AI Market Map

(1) Thesis

Voice AI is restructuring from a stack of discrete components (ASR → NLU → TTS) into vertically integrated, real-time conversational systems, collapsing infrastructure margins while value migrates to vertical applications that own workflows, data, and compliance moats.

---

(2) Segments

A. Core Model & Speech Infrastructure

Speech-to-text, text-to-speech, and speech-native foundation models

• ElevenLabs – TTS/voice cloning leader, now expanding into full conversational agents
• Deepgram – developer-focused ASR/STT APIs, low-latency streaming
• AssemblyAI – ASR plus speech understanding (summarization, sentiment)
• Cartesia – ultra-low-latency TTS built on state-space models
• OpenAI / Google – realtime speech-to-speech models (GPT-4o Realtime, Gemini Live) commoditizing the layer below

Dynamics: Rapid price compression as hyperscalers bundle speech into multimodal LLMs; independents survive on latency, voice quality, and developer experience — but each is racing up-stack.

B. Voice Agent Orchestration Platforms

Middleware for building, deploying, and monitoring voice agents

• Vapi – developer API for assembling voice agents from any model stack
• Retell AI – voice agent platform for phone-based automation
• LiveKit – open-source real-time infrastructure (WebRTC), powers ChatGPT voice mode
• Bland AI – enterprise phone agents at scale
• Sindarin (less certain on current traction) – conversational speech layer

Dynamics: The hottest, most crowded layer; thin differentiation and squeeze risk from both model providers above and vertical apps below. Winners will look like Twilio (distribution + reliability), losers like thin wrappers.

C. Vertical Applications — Healthcare

• Abridge – ambient clinical documentation, deep Epic integration
• Suki – voice assistant for physicians
• Nabla – ambient AI copilot for clinicians
• Hippocratic AI – patient-facing voice agents for care coordination
• Assort Health (earlier stage) – medical call center automation

Dynamics: Highest willingness-to-pay in voice AI; wins driven by EHR integration depth and clinical validation, not model quality. Abridge's Epic partnership shows distribution is the moat.

D. Vertical Applications — Customer Operations & Sales

• PolyAI – enterprise contact center voice assistants
• Sierra (Bret Taylor) – agentic customer experience, voice + chat
• Parloa – enterprise contact center automation (strong in Europe)
• Observe.AI – contact center intelligence and agent copilots
• Decagon (voice is newer for them) – AI customer support agents

Dynamics: Massive TAM (multi-hundred-billion BPO/contact-center spend) shifting from seat-based to outcome-based pricing; incumbents (NICE, Genesys, Five9) are acquiring or bundling aggressively.

E. Telephony & Real-Time Delivery Rails

• Twilio – programmable voice, now with AI-native offerings
• Telnyx – telephony infrastructure with AI voice APIs
• Daily – real-time video/audio infra; created the Pipecat open-source framework
• Agora – real-time engagement infrastructure

Dynamics: Stable, lower-margin utility layer enjoying a volume boom as every AI agent needs a phone number and low-latency transport; commodity risk offset by reliability requirements.

---

(3) Where Value Accrues: Vertical Applications (C & D)

Vertical apps win the most value because:

1. Pricing power: they price against labor replacement ($30–60/hr fully-loaded agents), not API tokens — 10–100x the margin capture of infra
2. Moats compound: workflow integration (EHRs, CRMs), compliance (HIPAA, PCI), and proprietary interaction data create switching costs infrastructure can't build
3. Infrastructure is deflating: model costs fall ~10x/year; the orchestration layer is being squeezed from both sides. Value flows to whoever owns the customer relationship and the outcome.

Healthcare specifically leads on ACVs today; customer ops leads on total value in 3–5 years due to sheer market size.

---

(4) White-Space Opportunities

1. Voice agent trust & assurance layer. No dominant player provides real-time hallucination detection, compliance monitoring, regression testing, and audit trails specifically for production voice agents — an "observability + certification" layer that enterprises will require before deploying agents in regulated calls. (Coval and Hamming AI are early attempts in eval, but nobody owns runtime assurance + compliance certification end to end.)

2. Voice-native interfaces for deskless/industrial workers. Field service, logistics, manufacturing, and construction workers (~80% of the global workforce) can't type on the job. Nobody has built the hands-free, noise-robust, workflow-integrated voice layer for frontline work — the "Siri for the shop floor" that connects to CMMS/WMS/ERP systems. Current players are all screen-first or consumer-first; the hardware+software+vertical-integration combination remains unowned.

---

Confidence notes: Company positioning is accurate as of my knowledge cutoff; funding stages and exact product scope for earlier-stage names (Sindarin, Assort Health) should be re-verified. The orchestration segment is evolving fast — competitive positions there may have shifted materially.