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:
- Pricing power: they price against labor replacement ($30–60/hr fully-loaded agents), not API tokens — 10–100x the margin capture of infra
- Moats compound: workflow integration (EHRs, CRMs), compliance (HIPAA, PCI), and proprietary interaction data create switching costs infrastructure can't build
- 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.