Voice AI: infrastructure, platforms, and vertical applications
September 13 at 21:06 · $0.126 total
Thesis: Voice AI is being restructured from a stack of separate ASR/NLU/TTS components priced per audio minute into an outcome-based, agentic application layer that collapses real-time speech models, telephony/orchestration, and vertical workflow integration into a single system of record.
Segments:
- Foundation models & speech infrastructure OpenAI (Realtime API/Whisper), ElevenLabs, Deepgram, AssemblyAI, Cartesia, Google DeepMind/Google Cloud (Gemini Live, Chirp). Dynamics: Rapid commoditization from hyperscaler bundling and open models; differentiation is shifting to latency, emotion/paralinguistics, language coverage, and on-device/edge deployment.
- Developer platforms & orchestration Vapi, Retell AI, Bland AI, LiveKit, Daily, Twilio (Voice Intelligence/ConversationRelay). Dynamics: Race to become the default voice-agent OS by abstracting telephony, routing, testing, and observability; squeezed between hyperscaler APIs and enterprises that want to own the stack.
- Vertical applications—healthcare Abridge, Ambience Healthcare, Suki, Nuance (Microsoft DAX), Nabla, Corti. Dynamics: Moving from transcription to ambient clinical intelligence and revenue-cycle workflows; EHR integration, HIPAA compliance, and clinician trust create high switching costs.
- Vertical applications—customer service/contact center Sierra, Decagon, PolyAI, Cresta, Observe.AI, ASAPP; incumbents NICE, Five9, Talkdesk. Dynamics: Largest near-term budget; AI-native startups and CCaaS incumbents are converging on outcome-based pricing and human-in-the-loop escalation.
- Vertical applications—financial services/sales/collections Skit.ai, Regal.ai, Nooks, Orum, Dialpad Ai, Aircall. Dynamics: High-ROI outbound, lead qualification, and collections; compliance (TCPA/FDCPA, consent, recording) and CRM/loan-system integration are the main moats. (Less sure about the current voice-AI depth of Nooks/Orum/Dialpad/Aircall; they are real companies in the segment but may be more dialer-assisted than fully agentic.)
Where value accrues: Vertical applications—especially healthcare and customer service—win the most value. They own the workflow, the system-of-record integration, and the outcome metric (documented encounter, resolved ticket, collected dollar). Infrastructure and platforms become usage-priced inputs with low switching costs; healthcare has the highest margin/regulatory moat, while contact center has the largest absolute spend.
White space:
- Voice AI for physical/field operations — noise-robust, edge/offline-capable voice agents for manufacturing, energy, construction, and logistics; integrated with wearables/AR, ERP, and work-order systems. Current players optimize for phone/office/clinical audio; no one owns the end-to-end industrial workflow.
- Voice AI for high-stakes negotiation and dispute resolution — real-time coaching or mediation agents for procurement, insurance claims, legal intake, and collections litigation; uses emotion/paralinguistic analysis, strategy optimization, and compliance guardrails. Existing players handle support/sales scripts, not adversarial multi-party conversations. (I’m not aware of a current end-to-end player; adjacent point solutions may exist.)
Voice AI is being restructured from point speech components (STT/TTS) into low-latency, tool-using conversational agents that sit inside workflows; infrastructure is commoditizing while orchestration platforms and outcome-owning verticals capture more of the stack.
1. Foundational voice models & APIs (STT, TTS, real-time streaming, cloning)
Deepgram, AssemblyAI, ElevenLabs, OpenAI (Whisper + TTS), Cartesia, Speechmatics (plus hyperscalers Google/Amazon/Microsoft). Dynamics: Fast commoditization on accuracy/latency/cost; winners differentiate on streaming reliability, noisy/accented speech, and multilingual coverage rather than raw model quality.
2. Voice-agent orchestration / developer platforms
Vapi, Retell AI, Bland.ai, Synthflow, LiveKit (real-time media layer). Dynamics: Abstraction layer for telephony, turn-taking, function calling, and evals; competing on developer experience, interruption handling, and time-to-production rather than owning the underlying models.
3. Contact-center / CX voice AI
PolyAI, Sierra, Decagon, Parloa, Observe.AI, Cresta (plus incumbents like Genesys/Talkdesk adding AI). Dynamics: High-volume, measurable ROI (containment, AHT, CSAT); shift from analytics/QA to fully autonomous or hybrid agents; enterprise sales cycles and integration depth create stickiness.
4. Vertical workflow applications (healthcare scribes and adjacent)
Abridge, Microsoft Nuance DAX, Ambience Healthcare, Nabla, DeepScribe, Suki. Dynamics: Ambient capture + structured output into EHR/workflow; regulatory and specialty-specific accuracy create moats; expansion from documentation into coding, orders, and patient-facing voice.
Value concentration: Vertical workflow applications (especially healthcare and high-stakes CX) win the most durable value. They own the outcome (time saved, revenue captured, compliance) rather than a component, command higher willingness-to-pay, and benefit from switching costs once embedded in systems of record. Pure infra faces pricing pressure; platforms capture some developer spend but risk being bypassed by large verticals or hyperscalers.
White-space (no dominant current player):
- Rugged, offline-first or high-noise field-service / industrial voice agents (construction, logistics, utilities) that combine STT with computer-use or IoT actions without assuming clean telephony.
- Privacy-preserving, on-device or air-gapped voice agents for highly regulated internal knowledge work (legal discovery, defense, certain financial ops) that never send audio to the cloud.
(Less sure on exact competitive intensity in some smaller CX names vs. the listed leaders; all companies named are real and active in the space.)