AI Agent Directory — Category
AI voice agents let software talk with people in real time — answering phones, qualifying leads, booking appointments, and narrating content. This guide covers the full stack: conversational phone-agent platforms, developer voice orchestration, and the speech-to-text and text-to-speech engines they run on. Every price below is checked against the vendor's own 2026 pricing pages.
"Voice AI" used to mean two things: reading text aloud (text-to-speech) and turning speech into text (transcription). In 2026 the centre of gravity has shifted to the voice agent — a system that listens, reasons, and speaks in a live conversation, most often over a phone line. The building blocks are the same three layers (speech-to-text, a language model, and text-to-speech), but the products have split into two camps: all-in-one agent platforms that give you a working phone agent from a prompt, and infrastructure APIs that developers assemble themselves for maximum control over latency and cost.
If you just want the short version: Vapi and Retell AI are the developer-favourite orchestration platforms; Bland AI offers the simplest all-inclusive per-minute phone agent; ElevenLabs still sets the bar for voice naturalness and now ships its own agents; Deepgram is the default speech-to-text engine for real-time pipelines; Cartesia leads on ultra-low-latency text-to-speech; Synthflow is the no-code choice for agencies and mid-market teams; and Hume AI is the pick when emotional expressiveness matters. Pricing is almost universally per minute of conversation rather than per seat, which changes how you budget — a busy support line is priced on call volume, not headcount.
Below you will find eight reviewed platforms, a feature and price comparison table, a framework for evaluating voice agents, individual write-ups with verified pricing, and guidance for four common buyer situations: developers, call centres, clinics, and consumer apps.
08 Agents Reviewed
Each platform below is assessed on latency, voice naturalness, speech recognition accuracy, telephony integration, language coverage, pricing model, and compliance posture. Scores shown are AI Agent Square editorial scores; platforms we have not yet formally scored are marked "Not yet scored".
Top Voice Quality
9.3/10
The most natural-sounding AI voices available, with fine emotional control and voice cloning from about a minute of audio. Its ElevenAgents product now wraps that TTS with an LLM, knowledge base, and telephony for full conversational agents in 29+ languages.
Best for Developers
Not yet scored
A developer-first platform for assembling voice agents from any LLM, STT, and TTS provider. You bring your own keys and models and pay a thin per-minute platform fee, keeping full control over latency, cost, and call flow.
Best Phone Agent Builder
Not yet scored
A managed platform for building inbound and outbound phone agents with a visual flow builder, batch calling, and analytics. Bundles voice infrastructure, TTS, and your chosen LLM into a single per-minute rate.
Simplest All-In Pricing
Not yet scored
A self-hosted-model phone agent platform that folds LLM, speech-to-text, and text-to-speech into one flat per-minute price with no separate token charges — attractive when you want predictable call economics at scale.
Best Speech-to-Text API
8.3/10
The developer-first speech platform. Nova-3 delivers low Word Error Rates with real-time streaming transcription, and the Aura TTS models and Voice Agent API round out a production-grade voice stack used by call centres and voice-agent builders.
Best No-Code Builder
8.3/10
A no-code voice-agent builder aimed at agencies and mid-market teams, with white-label options, CRM integrations, and managed telephony. Increasingly positioned around enterprise deployments with dedicated onboarding.
Lowest-Latency TTS
8.4/10
Built around the Sonic model for ultra-low-latency streaming speech, Cartesia is a favourite TTS engine for real-time voice agents where every millisecond of response time counts. Also offers Ink speech-to-text and voice cloning.
Best Emotional Expressiveness
8.1/10
Hume's Octave TTS and Empathic Voice Interface (EVI) generate speech with adjustable emotion and respond to the caller's tone, making it a strong fit for consumer companions, wellness, and expressive brand voices.
Match the Tool to the Job
Voice AI serves very different workflows. Use our comparison tools and pricing guide to match the right platform to your use case — inbound support, outbound calling, a developer pipeline, or content narration.
Head-to-Head
Best-for, verified 2026 pricing, and the single most important limitation for each platform. Prices are usage-based unless noted; model, telephony, and add-on costs can stack on top of platform fees.
| Agent | Score | Type | Best For | Verified Price (2026) | Key Limitation |
|---|---|---|---|---|---|
| ElevenLabs | 9.3/10 | TTS + Agents | Voice quality & narration | TTS from $5/mo; Agents free 15 min then $0.08/min | Agent add-ons (LLM, telephony) billed separately |
| Vapi | — | Orchestration | Developers wiring their own stack | $0.05/min platform fee + model + telephony | Assembly required; you own model/latency tuning |
| Retell AI | — | Phone Agents | Inbound/outbound phone automation | $0.07–$0.31/min bundled; $10 free credit | Rate swings widely with premium LLM choice |
| Bland AI | — | Phone Agents | Predictable all-in per-minute cost | $0.14/min all-inclusive (Start); enterprise custom | Less model choice than open orchestrators |
| Deepgram | 8.3/10 | STT/TTS API | Real-time speech-to-text pipelines | Nova-3 STT ~$0.0048/min; Aura-2 TTS $0.030/1k chars; $200 credit | Building blocks, not a finished agent |
| Synthflow | 8.3/10 | No-Code Agents | Agencies & no-code teams | Self-serve from ~$29/mo; enterprise from $30,000/yr | Enterprise-oriented pricing at scale |
| Cartesia | 8.4/10 | TTS/STT API | Ultra-low-latency real-time voice | Free 20k credits; Pro $5/mo; Startup $49/mo | Voice library smaller than ElevenLabs |
| Hume AI | 8.1/10 | Emotive TTS + EVI | Expressive consumer & wellness apps | Free tier; Starter $3/mo; Pro $70/mo | Emotive focus over broad enterprise telephony |
Voice is unforgiving in a way that text chat is not. A chatbot can pause for two seconds and no one minds; a voice agent that pauses for two seconds sounds broken. Because of that, the evaluation criteria for voice AI are dominated by real-time performance and the quality of the audio on both sides of the conversation. Here are the seven factors that matter most when you compare platforms.
The single most important metric is how quickly the agent starts responding after the caller stops talking. This "turn latency" is the sum of three stages: speech-to-text finalising what was said, the language model generating a reply, and text-to-speech producing the first chunk of audio. For a conversation to feel natural you want the total under roughly 800 milliseconds, and the best real-time stacks push toward 500ms. Streaming matters enormously here — engines like Cartesia Sonic and Deepgram's real-time models emit audio and transcripts incrementally rather than waiting for a complete result, which is what makes sub-second turns possible. When you demo a platform, test it on a real phone call over a cellular connection, not just in a browser, because network conditions add latency that a local demo hides.
Naturalness covers prosody (rhythm and intonation), the absence of robotic artefacts, and the ability to convey emotion appropriate to context. ElevenLabs remains the reference point for lifelike output, and Hume AI's Octave is built specifically to modulate emotion and even respond to the caller's tone. For an agent, expressiveness has a functional purpose: an empathetic tone on a healthcare or collections call changes outcomes, not just perception. Judge this with your own ears on your own scripts — vendor demo reels are cherry-picked, so paste in awkward sentences, numbers, dates, and your brand names to hear how the voice handles them.
If the agent mishears the caller, nothing downstream can recover. Word Error Rate (WER) is the standard measure, but the number that matters is WER on your audio: your accents, your industry vocabulary, background noise, and phone-quality audio (8kHz telephony is harder than studio audio). Deepgram, along with providers such as AssemblyAI, publishes competitive WER, and features like custom vocabulary, keyword boosting, and speaker diarization materially improve real-world accuracy. Test with recordings of genuine calls, including the messy ones.
A voice agent is only useful if it connects to your phone system. Check whether the platform provides native phone numbers, supports SIP trunking to bring your own carrier, and integrates with your contact-centre software. Retell AI, Bland AI, Synthflow, and CloudTalk AI include managed telephony; Vapi and Deepgram expect you to connect a provider such as Twilio or Telnyx. Also confirm support for warm transfers to a human, DTMF (keypad) capture for menus and payments, and call recording with the controls your compliance team requires.
Language support differs between the listening and speaking halves of the stack, so verify both. ElevenLabs covers 29+ languages for synthesis, Deepgram's multilingual models handle dozens for transcription, and Cartesia and Hume both offer multilingual voices. If you operate internationally, test each target language at production latency — some languages are noticeably slower or less accurate than English, and code-switching (a caller mixing two languages in one sentence) trips up many systems.
Almost every conversational voice platform bills per minute of conversation, not per seat. That is a fundamentally different budgeting exercise: your cost is driven by call volume and average handle time. Read the fine print carefully, because the headline rate is rarely the whole bill. Platform fees, the language model, telephony, and add-ons like PII redaction or knowledge-base retrieval each add cents per minute. Bland AI's flat $0.14/min is all-inclusive; Vapi's $0.05/min is only the platform layer with model and telephony passed through; Retell AI's $0.07–$0.31/min swings on which LLM you pick. Build a spreadsheet with your real minutes before comparing.
Voice calls routinely contain personal and sometimes protected health information. Confirm SOC 2 Type II, GDPR handling, and — for healthcare — HIPAA eligibility with a signed Business Associate Agreement. Deepgram, ElevenLabs, and Hume AI publish compliance posture for eligible tiers, and enterprise agent vendors will sign BAAs, but these protections almost never apply to free or entry self-serve plans. Also ask where audio is stored, how long recordings are retained, whether data is used to train models, and whether PII redaction runs before storage.
ElevenLabs earned its reputation on raw output quality: its voices carry emotional nuance and natural pacing that competitors are still chasing, and its cloning can reproduce a voice from about a minute of audio. In 2026 the company extended from pure text-to-speech into full conversational agents (ElevenAgents), pairing its TTS with an LLM, a knowledge base, workflow builder, and telephony. Verified pricing: the TTS product starts with a free tier and a Starter plan at $5/month, scaling through Creator (~$11/mo) to Pro at $99/month. ElevenAgents has its own free tier with 15 minutes of calls, then charges roughly $0.080 per minute for included call minutes (additional minutes about $0.003), with paid agent plans from $6/month. The catch for agents is that external LLM and telephony usage are billed on top, so model your total. Read our full ElevenLabs review.
Vapi is not a finished agent; it is the plumbing that lets developers assemble one from best-of-breed parts. You choose the LLM, the speech-to-text engine, and the text-to-speech voice — bringing your own API keys if you want — and Vapi handles the real-time orchestration, interruption handling, and call management. Verified pricing: the Build plan charges a $0.05 per minute platform fee, with model-provider and telephony costs passed through at cost (or free if you supply your own keys), 10 concurrent lines included, and additional lines at $10/line/month. The Scale plan offers committed-volume enterprise pricing. Vapi's strength — total control over latency and cost — is also its cost: you own the integration and tuning work. It is the right choice for teams with engineering capacity who want to squeeze latency and margins. Read our full Vapi review.
Retell AI sits between raw orchestration and no-code: it gives you a visual flow builder, batch outbound calling, post-call analytics, and managed infrastructure, while still letting you pick your LLM. Verified pricing is genuinely usage-based and component-driven: Retell quotes roughly $0.07 to $0.31 per minute depending on the model and features you enable, built from voice infrastructure (~$0.055/min), text-to-speech (~$0.015/min for standard voices), the LLM (from fractions of a cent to $0.30+/min for realtime models), and telephony (~$0.015/min). Add-ons such as knowledge base, PII removal, and AI quality assurance stack on top. New accounts get $10 in free credits. It is a strong default for teams building serious inbound or outbound phone automation without wanting to wire everything themselves. Read our full Retell AI review.
Bland AI takes the opposite approach to component pricing: it folds the language model, speech-to-text, and text-to-speech into a single flat rate with no separate token charges. Verified pricing: the Start plan is $0.14 per minute, all-inclusive, with an Enterprise tier contracted to your volume (custom pricing via sales). The appeal is predictability — you can forecast the cost of a campaign from expected minutes alone, without modelling LLM tokens. The trade-off is less flexibility in swapping models than an open orchestrator gives you, and you should benchmark its latency and voice quality against Retell and Vapi on your own scripts. Bland is a good fit for high-volume outbound where predictable unit economics matter more than fine-grained model control. Read our full Bland AI review.
Deepgram is the speech-to-text (and increasingly text-to-speech) layer under a large share of production voice agents. Its Nova-3 models achieve low Word Error Rates with fast streaming transcription, and the Aura TTS models plus a dedicated Voice Agent API round out the stack. Verified pricing: Nova-3 monolingual streaming is about $0.0048 per minute pay-as-you-go (multilingual ~$0.0058), the Aura-2 TTS model runs about $0.030 per 1,000 characters, and new accounts receive $200 in free credit. Add-ons like diarization and redaction add roughly $0.002/min each, and committed Growth-tier volume lowers the base rate. Deepgram is infrastructure, not a finished agent — but if you are building one, it is the most common recommendation for the listening half. Read our full Deepgram review.
Synthflow targets teams that want production voice agents without engineering: a no-code builder, white-label options for agencies, CRM and calendar integrations, and managed telephony. On pricing, Synthflow has historically offered self-serve plans starting around $29/month, but its own pricing page now emphasises enterprise deployment, stating that enterprise contracts start at $30,000 annually, custom-scoped to call volume, concurrency, integrations, and security needs. If you are an agency reselling voice agents or a mid-market team standardising on one no-code platform, Synthflow is a leading option; confirm the current self-serve tiers directly with the vendor, since the entry pricing has shifted toward larger commitments. Read our full Synthflow review.
Cartesia is built around the Sonic model, engineered for ultra-low-latency streaming speech — which is exactly what real-time voice agents need on the speaking side. It also offers Ink speech-to-text and instant voice cloning. Verified pricing is credit-based and generous at the low end: a Free plan with 20,000 credits, a Pro plan at $5/month (100,000 credits), a Startup plan at $49/month (1.25M credits), and Scale at $299/month, where TTS consumes one credit per character. Its main limitation versus ElevenLabs is a smaller voice library and slightly less emotional range, but for agent builders who prioritise latency and cost, Cartesia is frequently the TTS of choice. Read our full Cartesia review.
Hume AI differentiates on emotion. Its Octave TTS generates speech with adjustable emotional expression, and its Empathic Voice Interface (EVI) responds to the emotional tone of the caller, not just their words. Verified pricing: a Free tier (10,000 TTS characters and 5 minutes of EVI monthly), a Starter at $3/month, Creator at $14/month, and Pro at $70/month (1M characters, 1,200 EVI minutes), scaling to Business at $500/month and custom Enterprise with SOC 2 Type II, GDPR, and HIPAA. Octave overage runs roughly $0.05–$0.15 per 1,000 characters depending on tier. Hume is the standout for consumer companions, wellness, coaching, and any application where how something is said matters as much as what is said. Read our full Hume AI review.
Beyond the eight featured platforms, several adjacent tools show up regularly in voice-agent projects. PlayAI (Play.ht) offers a large TTS voice library with a free tier of 1,000 characters/month and paid plans in the low tens of dollars. Vocode is an open-source framework you can self-host for free when you want to own the entire stack. PolyAI targets large enterprises with fully managed, per-minute customer-service voice assistants. Ringly.io packages AI phone support for e-commerce brands with plans from $349/month (1,000 minutes, $0.29/min overage). CloudTalk AI layers voice agents onto a full cloud call-centre platform, with AI Receptionist tiers from about €99/month. And Murf AI remains a strong pick for timeline-synced voiceover production, with a free tier and paid plans from $29/month.
The best voice AI platform depends far less on a leaderboard and far more on who you are and what you are shipping. Here is how the picks change across four common buyer profiles.
If you have engineering capacity and care about latency and margins, assemble your own stack. Use Vapi as the orchestration layer, Deepgram for speech-to-text, and Cartesia or ElevenLabs for text-to-speech, connected to a telephony provider like Twilio. This gives you sub-second turns and per-component cost control. If you want to own everything and avoid platform fees entirely, Vocode is an open-source framework you can self-host. Pay-as-you-go pricing and generous free credits (Deepgram's $200, Cartesia's free tier) make prototyping cheap.
For inbound and outbound phone automation at volume, start with Retell AI or Bland AI. Retell gives you a flow builder, batch calling, and analytics with component-level pricing; Bland gives you flat, predictable $0.14/min economics. For no-code deployment and agency white-labelling, Synthflow is the leading option. If you are a large enterprise with a formal contact centre, evaluate PolyAI for managed customer-service assistants and CloudTalk AI if you want voice agents inside a complete call-centre platform. Prioritise warm human transfer, DTMF capture, and SIP integration in your evaluation.
Compliance is the gating requirement. Only consider platforms that will sign a Business Associate Agreement and document HIPAA and SOC 2 posture — Deepgram, ElevenLabs, and Hume AI publish compliance coverage for eligible enterprise tiers, and managed agent vendors will contract for it. Never process protected health information on a free or self-serve plan. Prioritise PII redaction before storage, configurable retention, and a clear statement that your audio is not used to train models. Latency and empathy still matter (patients notice a cold or laggy agent), but they come after the paperwork.
When the voice is the product — a companion app, a game character, an interactive story, or brand narration — expressiveness wins. ElevenLabs for top-tier naturalness and multilingual reach, Hume AI when emotional responsiveness is central, and PlayAI or Murf AI for large voice libraries and content workflows. Consumer scale makes per-character and per-minute costs add up fast, so model your unit economics early and lean on free tiers to validate before committing.
Questions & Answers
Straight answers to the questions buyers ask most when choosing an AI voice agent in 2026.
An AI voice agent is software that holds a spoken conversation in real time. It combines speech-to-text (STT) to hear the caller, a large language model to decide what to say, and text-to-speech (TTS) to speak back, usually connected to a phone line or an app microphone. Platforms such as Vapi, Retell AI and Bland AI bundle all three layers so you can deploy an automated phone agent, while ElevenLabs, Deepgram and Cartesia provide the underlying TTS and STT that many agents run on.
Most conversational voice platforms bill per minute of conversation rather than per seat. All-in phone-agent platforms range from about $0.05/min in platform fees (Vapi) plus passed-through model and telephony costs, to $0.07–$0.31/min bundled on Retell AI, to $0.14/min all-inclusive on Bland AI. Pure TTS and STT APIs are usually billed by characters or audio minutes: Deepgram's Nova-3 speech-to-text is around $0.0048/min and its Aura-2 TTS about $0.030 per 1,000 characters. Always model your real monthly minutes, because telephony, premium LLMs and add-ons like PII redaction stack on top of the base rate.
For a natural phone conversation you generally want end-to-end response latency (the gap between the caller finishing speaking and the agent starting to reply) under about 800 milliseconds, and ideally near 500ms. That budget is split across speech-to-text, the language model, and text-to-speech, so low-latency streaming TTS engines like Cartesia Sonic and fast STT like Deepgram Nova-3 or Flux matter as much as the LLM you choose.
Developers who want full control typically choose Vapi or Retell AI for orchestration, paired with Deepgram for speech-to-text and Cartesia or ElevenLabs for text-to-speech. Vapi lets you bring your own model and STT/TTS keys and pay only a per-minute platform fee, while Vocode is an open-source option you can self-host for free when you want to own the whole stack.
For contact-centre automation, look at Retell AI and Bland AI for outbound and inbound phone agents, Synthflow for no-code deployment, PolyAI for large enterprise customer-service lines, and CloudTalk AI when you need voice agents layered onto a full cloud call-centre platform. The right pick depends on call volume, existing telephony, and how much of the stack you want managed.
Some are, but only under the right plan and with a signed Business Associate Agreement. Deepgram, ElevenLabs and Hume AI publish HIPAA and SOC 2 posture for eligible tiers, and enterprise voice-agent vendors will sign BAAs. Never process protected health information on a free or self-serve tier without confirming compliance coverage in writing first.
If you are building a custom application and care about latency and cost, assembling best-of-breed STT (Deepgram), TTS (Cartesia or ElevenLabs) and an LLM through an orchestrator like Vapi gives you the most control. If you want a working phone agent quickly, an all-in-one platform such as Retell AI, Bland AI or Synthflow bundles every layer so you configure a prompt and a phone number instead of wiring components together.
Yes. ElevenLabs supports 29+ languages for TTS, Deepgram's multilingual models cover dozens of languages for transcription, and Hume's Octave and Cartesia's Sonic both offer multilingual voices. For voice agents specifically, confirm that both the STT and TTS layers support your target language at low latency, since language coverage often differs between the listening and speaking sides of the stack.
Guides & Research
Deep-dive resources on evaluating and deploying voice AI for phone automation, developer pipelines, and content.
Everything procurement and engineering teams need to evaluate voice-agent platforms and APIs — from latency budgets and per-minute pricing to compliance requirements.
How voice cloning works, where it helps with brand consistency and accessibility, and the consent and compliance requirements to get right.
How AI phone agents fit into a modern support operation — routing, warm transfers, compliance, and choosing between managed and self-built stacks.
New independent, hands-on reviews added to this category. Scores are AI Agent Square editorial scores.
Stay Current
New model releases, latency benchmarks, and pricing changes for voice AI platforms — delivered monthly.
Related reading