Voice AI: infrastructure, platforms, and vertical applications
September 13 at 20:47 · $0.116 total
Voice AI Market Map
(1) Thesis
Voice AI is being restructured from a fragmented stack of ASR/TTS point solutions into vertically integrated, latency-optimized conversational systems — value is migrating away from commodity speech models toward orchestration layers and vertical applications that own the customer workflow and outcome (e.g., a booked appointment, a resolved claim).
---
(2) Segment Map
A. Core Speech Models & Infrastructure (ASR/TTS/Speech-to-Speech)
Companies: Deepgram, AssemblyAI, ElevenLabs, Cartesia, OpenAI (Realtime API), Speechmatics
Dynamics: Rapid commoditization at the model layer; pricing compression as OpenAI/Google bundle speech into frontier models. Survivors differentiate on latency (Cartesia's SSM-based approach), voice quality/cloning (ElevenLabs), or domain accuracy (Deepgram, AssemblyAI on telephony audio). ElevenLabs is the breakout, expanding upward into agents.
B. Voice Agent Orchestration Platforms (the "middleware" layer)
Companies: Vapi, Retell AI, Bland AI, LiveKit, Daily (Pipecat), Sindarin (less certain on Sindarin's current traction)
Dynamics: The hottest, most crowded layer — these stitch ASR + LLM + TTS + telephony into sub-second conversational loops. Low switching costs and thin moats today; LiveKit and Daily/Pipecat benefit from being infrastructure-neutral (LiveKit powers ChatGPT voice mode). Risk: squeezed from below (OpenAI Realtime API) and above (vertical apps building in-house).
C. Contact Center & Enterprise CX Automation
Companies: PolyAI, Parloa, Sierra, Cresta, Observe.AI, Replicant
Dynamics: Largest near-term revenue pool — displacing $300B+ in BPO/contact center labor. Sales cycles are enterprise-slow but contracts are large and sticky. Sierra (Bret Taylor) validates outcome-based pricing (per-resolution), which is the key business model innovation. Incumbents (NICE, Genesys, Five9) are acquiring or building fast.
D. Vertical Voice Applications
Companies: Abridge (healthcare scribing), Suki (clinical assistant), Slang.ai (restaurants), Numa (auto dealerships), Happy Robot (freight/logistics), Rilla (field sales coaching)
Dynamics: Deepest moats via workflow integration, proprietary domain data, and regulatory compliance (HIPAA). Healthcare is the standout — Abridge's EHR-embedded distribution (Epic partnership) is the template. Winners here are "voice-enabled software companies," not "voice companies."
E. Voice Interfaces & Consumer/Prosumer
Companies: ElevenLabs (again — dubbing/audiobooks), Wispr Flow (dictation), Otter.ai (meetings), Speak (language learning)
Dynamics: High adoption velocity, low pricing power outside of a few winners; mostly features awaiting absorption by OS/platform players (Apple, Google, Microsoft). Speak's language-learning wedge shows consumer verticals can sustain standalone businesses if voice is core to the value prop, not additive.
---
(3) Where Value Accrues: Vertical Applications (D), with C as the revenue giant
Contact center automation (C) captures the most dollars near-term — it's a direct labor substitution with quantifiable ROI. But vertical applications (D) win the most value per dollar of revenue: they compound proprietary conversational data, embed into systems of record (Epic, dealership DMS, TMS), and are insulated from model commoditization. Abridge reaching a reported ~$5B valuation on the back of Epic distribution is the proof point. The infrastructure layer (A) trends toward oligopoly with thin margins; orchestration (B) is strategically important but structurally squeezed unless a player (likely LiveKit or Vapi) becomes the "Twilio of voice agents."
---
(4) White-Space Opportunities
1. Voice agent trust, evaluation & compliance infrastructure. There is no dominant player for pre-deployment simulation testing, regression testing across model updates, real-time hallucination/compliance monitoring, and audit trails for regulated voice agents (TCPA, HIPAA, EU AI Act voice-disclosure rules). Early startups exist (Coval, Hamming AI — both early-stage, less certain of traction), but nothing at scale. Every enterprise deploying agents in segment C/D needs this; it's the "Datadog + Vanta for voice agents" gap.
2. Agent-to-agent voice interop / the "machine dial tone." As businesses deploy outbound voice agents (scheduling, procurement, logistics) that increasingly call other businesses' inbound agents, there's no protocol layer for agents to authenticate, negotiate, and transact machine-to-machine — skipping synthesized speech entirely when both ends are AI. Whoever builds the identity + handshake standard (think "SWIFT for voice agents") owns a toll-road position no current infra, orchestration, or app player addresses.
---
Confidence notes: Company lists in segments A–D are high confidence. Specific valuations/metrics cited (Abridge ~$5B, Sierra's pricing model) reflect reporting as of early-mid 2025 and may have shifted. Sindarin, Coval, and Hamming AI are flagged as lower-confidence given early stage.
Thesis: Voice AI is being restructured from siloed STT/TTS + legacy IVR into LLM-native, low-latency end-to-end conversational agents, with infrastructure rapidly commoditizing while platforms and especially verticals capture durable workflow and data moats.
Segments
1. Foundational speech infrastructure (STT, TTS, real-time voice models) Deepgram, AssemblyAI, ElevenLabs, Cartesia, OpenAI (Whisper + Realtime API), Google Cloud Speech/Gemini Live. Dynamics: Fierce competition on latency, naturalness, and cost; open models and hyperscalers are compressing margins, pushing vendors toward specialized real-time multimodal stacks.
2. Voice agent orchestration / developer platforms Vapi, Retell AI, Bland AI, Synthflow, Voiceflow. Dynamics: Abstraction layers that let teams compose agents with tools, telephony, and evals; high feature velocity but also high churn as core capabilities (turn-taking, function calling, monitoring) become table stakes.
3. Vertical applications — contact center / customer experience PolyAI, Cognigy, Parloa, Talkdesk, Genesys Cloud. Dynamics: Replacing traditional IVR and human agents in high-volume CX; value accrues from deep CRM/telephony integrations, compliance, and measured containment/resolution rates.
4. Vertical applications — healthcare & professional services Infinitus, Abridge, Ambience Healthcare, Nabla. Dynamics: High-stakes, regulated workflows (prior auth, documentation, patient outreach) where accuracy, auditability, and domain data create stronger switching costs than generic platforms.
Which segment wins the most value and why Vertical applications (especially CX and healthcare) capture the most durable value. Infrastructure is becoming a cost of goods sold; platforms face commoditization and switching. Verticals own the workflow, proprietary interaction data, regulatory positioning, and ROI metrics that buyers actually pay for, producing higher retention and pricing power.
White-space opportunities (no current player fully covers)
- Real-time, multi-party, highly regulated conversations (e.g., legal depositions, multi-clinician consults, or insurance appeals) that require simultaneous speaker diarization, live compliance redaction, and auditable reasoning traces.
- Privacy-first, on-device or edge voice agents for consumer hardware and air-gapped enterprise environments (no cloud round-trips), combining local SLMs with hardware-optimized TTS/STT.
(Less sure on exact current coverage of the second white space by emerging hardware players; most named companies above remain cloud-centric.)