Top 10 TTS and STT APIs for AI Agents (2026)

The complete market map and weighted assessment of 48 text-to-speech and speech-to-text providers for AI agents: benchmark data, real pricing, latency measurements, and the 10 best ranked by what actually matters.

Voice AI funding surged 8x to $2.1 billion in 2025. ElevenLabs hit $330 million ARR and an $11 billion valuation - TechCrunch. Deepgram raised a $130 million Series C at $1.3 billion - Deepgram. Production voice agent deployments grew 340% year-over-year, and Gartner predicts conversational AI will reduce contact center labor costs by $80 billion in 2026 - Gartner. The voice AI agent market hit $22.5 billion and is accelerating at 34.8% CAGR toward $47.5 billion by 2034 - Market.us.

This is not a future trend. Voice agents are in production now, at scale. The question for builders is not whether to add voice to their agents. It is which TTS and STT provider delivers the best quality at the right price for their specific use case.

We mapped 47 providers across the TTS and STT landscape: 21 cloud TTS APIs, 12 cloud STT APIs, 5 voice agent platforms, and 10 open-source models. We scored them on 7 weighted criteria derived from first principles (what does an AI agent actually need from voice?), verified every price against official pricing pages, and cross-referenced quality claims against independent benchmarks from CodeSOTA, Artificial Analysis Speech Arena, and Inworld's 2026 evaluation. As we explored in our guide to the top 10 agent capabilities, TTS and STT are among the most important external capabilities for production agents.

Contents

1. The Master Ranking: Top 10 TTS and STT APIs, Weighted and Ranked
2. Assessment Criteria and Weight Rationale
3. The Top 10: Detailed Profiles
4. The Full Provider Directory (47 Services)
5. TTS Pricing Comparison
6. STT Pricing Comparison
7. Benchmark Data: Quality, Latency, and Accuracy
8. Open Source: The Quality Gap Has Closed
9. Voice Agent Platforms: The Full-Stack Option
10. How to Choose: Decision Framework

---

1. The Master Ranking: Top 10 TTS and STT APIs, Weighted and Ranked

How to read this table: Each cell is a raw score (0-10). The Final Score is the weighted average: (Quality x 0.25) + (Cost x 0.20) + (Latency x 0.20) + (Agent Ready x 0.15) + (Languages x 0.10) + (Cloning x 0.05) + (Scale x 0.05). The "What It Does" column captures each provider's identity and standout metric. Ordered by final score, best first.

Criteria definitions and weight rationale:

• Quality (25%): MOS scores for TTS, WER for STT, based on CodeSOTA and Artificial Analysis benchmarks. Highest weight because bad voice quality kills user experience instantly.
• Cost (20%): Normalized to $/1M chars for TTS, $/hr for STT. 10 = cheapest tier, 1 = most expensive. High weight because voice agents run at volume (thousands of minutes/month).
• Latency (20%): Time to first audio (TTS) or streaming factor (STT). 10 = sub-100ms, 7 = 100-300ms, 5 = 300-500ms, 3 = 500ms-1s, 1 = 1s+. Equal to cost because conversational agents need sub-300ms response.
• Agent Readiness (15%): Voice agent platform, WebSocket streaming, SDKs, telephony integration. 10 = full voice agent stack, 5 = REST API + streaming, 1 = REST only.
• Languages (10%): Number of supported languages. 10 = 100+, 7 = 40-99, 5 = 15-39, 3 = 5-14, 1 = English only.
• Voice Cloning (5%): Instant cloning, custom voice training. 10 = instant + professional, 5 = basic cloning, 0 = not available. Lower weight because not every agent needs custom voices.
• Scale/Reliability (5%): Enterprise SLA, concurrency limits, uptime guarantees. 10 = SOC2/HIPAA + 99.99%, 5 = standard cloud, 1 = no SLA.

The ranking shows a clear top tier: ElevenLabs (7.9) leads on quality and agent platform completeness. Deepgram (7.7) and Cartesia (7.7) tie for second through different strengths: Deepgram on price-performance and full-stack voice agents, Cartesia on raw latency (40ms) and the arena-leading Sonic-3.6 model. AssemblyAI (7.2) is the STT specialist with the richest feature set, and the only vendor whose realtime model sits inside the human-parity band that independent testers define as 2-4% word error rate.

For a broader view of how voice fits into the agent capability stack, see our guide to the top 10 agent capabilities. For cost analysis across all agent infrastructure, see our AI agent cost analysis.

---

2. Assessment Criteria and Weight Rationale

The weights above are derived from first principles: what does an AI agent fundamentally need from voice capabilities?

An AI agent that speaks to humans needs three things above all else. First, the voice must sound natural enough that users do not disengage. A robotic voice triggers an immediate trust deficit that no amount of intelligence can overcome. This is why quality gets 25%: it is the gatekeeper for everything else. The MOS (Mean Opinion Score) benchmarks from CodeSOTA show the top providers now score 4.5-4.8 out of 5.0, within the range that is "generally indistinguishable from human speech in blind tests" - CodeSOTA. Below 4.0, users notice. Below 3.5, they leave.

Second, the agent must respond quickly. In a phone call, silence longer than 500 milliseconds feels like the agent is broken. In a customer support interaction, every additional second of latency increases abandonment. This is why latency gets 20%, equal to cost. The range across providers is enormous: from 25ms (Inworld Realtime TTS-2 Flash, P99 measured server-side) and 40ms (Cartesia Sonic Turbo) to over 800ms (hyperscaler defaults). For conversational agents, anything above 300ms creates perceivable delay.

Third, the agent must be affordable at scale. A voice agent handling 10,000 minutes per month (typical for a mid-sized customer support deployment) will cost between $100 and $1,500/month in TTS alone, depending on the provider. STT adds another $15 to $1,000/month for the same volume. This is why cost gets 20%: at production volume, the difference between providers is material.

The remaining criteria (agent readiness at 15%, languages at 10%, cloning and scale at 5% each) matter for specific use cases but are not universally load-bearing. A US-only English agent does not need 100+ languages. A content generation agent does not need voice cloning. The weights reflect the general case.

---

3. The Top 10: Detailed Profiles

3.1 ElevenLabs (Score: 7.9/10)

ElevenLabs is the market leader in TTS quality and the only provider that combines best-in-class TTS with competitive STT in a single platform. The company reached $330 million in ARR, raised a $500 million Series D at an $11 billion valuation in February 2026, and serves over 1 million developers - ElevenLabs.

The TTS quality is objectively the best available. ElevenLabs Turbo v2.5 scores MOS 4.8 on CodeSOTA benchmarks, the highest of any provider - CodeSOTA. Eleven v3 sits third on the Artificial Analysis Speech Arena's controlled-voice board at 1,066 Elo, behind Cartesia's two most recent Sonic releases. Flash models (lower quality but faster) score MOS 4.6 at roughly half the latency.

Pricing spans two tiers: Flash/Turbo and the newer v3 Conversational model at $50/1M characters, and Multilingual v2 and v3 at $100/1M characters. Subscription plans range from free to $990/month (Business). The STT product, Scribe, costs $0.22/hour for batch and $0.39/hour for real-time streaming - ElevenLabs Pricing.

The STT side moved further than the price list suggests. Scribe v2 Realtime transcribes with roughly 150ms of latency at 93.5% accuracy across 30 languages, the lowest word error rate of any low-latency ASR model on the FLEURS benchmark, and the batch Scribe v2 model was measured at 1.5% WER in Artificial Analysis's July 2026 study - ElevenLabs. On the TTS side, Eleven v3 Conversational delivers expressive realtime speech across 70+ languages at approximately 280ms latency, while Flash v2.5 trades expressiveness for roughly 75ms - ElevenLabs Models.

The Conversational AI 2.0 platform is what pushes ElevenLabs to the top for agent builders. It provides a complete voice agent stack: STT, LLM orchestration, TTS, with WebSocket streaming, Twilio/Genesys/Vonage telephony integrations, and SOC 2/HIPAA/GDPR compliance. If you are building a voice agent and want one vendor for the full pipeline, ElevenLabs is the most complete option - ElevenLabs.

The main weakness is cost. At $50-100/1M characters, ElevenLabs is 3-7x more expensive than alternatives like OpenAI TTS-1 ($15/1M) or Inworld Realtime TTS-2 Flash ($15/1M). For high-volume, cost-sensitive deployments, the quality premium may not justify the price difference, especially as competitors narrow the quality gap.

3.2 Deepgram (Score: 7.7/10)

Deepgram is the most complete price-performance play in voice AI. Its Nova-3 STT model achieves 2.2% WER on LibriSpeech clean audio (lowest of any commercial API) with a 54.3% WER reduction over competitors in streaming scenarios - Deepgram. The company raised a $130 million Series C at a $1.3 billion valuation in January 2026 - TechCrunch.

STT pricing is aggressive: Nova-3 pre-recorded at $0.0043/minute ($0.26/hour) and streaming at $0.0077/minute ($0.46/hour), both dropping roughly 15% on Growth plans. The Aura-2 TTS costs $30/1M characters with 90-200ms time-to-first-byte, and the newer Flux TTS runs $45/1M - Deepgram Pricing.

The more consequential release is Flux, which Deepgram calls the first production Conversational Speech Recognition model: instead of transcribing and then guessing at turn boundaries with a silence timer, end-of-turn detection is built into the model itself. Deepgram measures that as 200 to 600ms cut from agent response latency versus a conventional pipeline and roughly 30% fewer false interruptions, at Nova-3 level word error rates - Deepgram. Flux Multilingual followed on April 29, 2026, extending model-based turn detection to 10 languages with end-of-turn decisions in under 400ms, and switching languages mid-conversation - Deepgram. The $200 free credit (enough for 46,000+ minutes of transcription) is the most generous trial in the STT market.

The Voice Agent API is Deepgram's strongest differentiator for agent builders. At $0.075/minute standard pay-as-you-go ($0.068 on Growth), it provides a bundled STT + TTS pipeline optimized for real-time conversation. Combined with Python, JavaScript, Go, and .NET SDKs plus Twilio integration, it is the fastest path to a production voice agent for developers who want to own the stack without using a full platform like Vapi or Retell.

The weakness is TTS quality. Aura-2 is competent but does not match ElevenLabs, Cartesia, or OpenAI on naturalness. For agents where voice quality is the primary differentiator (luxury brand, premium support), Deepgram's TTS may not be sufficient.

3.3 Cartesia (Score: 7.7/10)

Cartesia pairs top-tier latency with the best blind-preference scores in the market. Sonic Turbo achieves 40ms time-to-first-audio and Sonic-3 runs at approximately 90ms in production streaming - Inworld Benchmarks. The quality is strong at MOS 4.7 (Sonic 2), placing it in the top tier alongside ElevenLabs and OpenAI.

Sonic-3.6, released in August 2026, took the #1 position on both Artificial Analysis Speech Arena leaderboards at once, scoring 1,282 Elo on the Provider Voice board (a 43-point lead over the second-placed Speechify Simba 3.2) and 1,120 Elo on the Controlled Voice board, ahead of Sonic-3.5 at 1,095 and ElevenLabs Eleven v3 at 1,066 - Artificial Analysis. In blind head-to-head tests with native speakers, Cartesia reports Sonic-3.6 winning 92% of US English comparisons against Eleven v3, 91% in Spain Spanish and 86% in French, across 44 languages and 61 locales, still replying in under 90ms - Cartesia.

Pricing uses a credit system: Sonic bills one credit per character generated. Plans range from free (20,000 credits) to Scale ($299/month for 8 million credits, about 10,667 minutes of audio). That works out to roughly $0.028 per minute, or approximately $37/1M characters - Cartesia Pricing. Ink-2 STT is billed per second of audio, about 740 hours on the same Scale plan, so roughly $0.40/hour.

The Line voice agent platform positions Cartesia as a full-stack voice AI company. Instant voice cloning from 3 seconds of audio, 40+ language support, and WebSocket streaming make it purpose-built for the real-time conversational use case. If your agent's primary requirement is a voice response fast enough to feel like an interruption never happened (phone banking, emergency services, gaming), Cartesia is the safest choice, because it is the only vendor holding both the latency and the preference-quality position at the same time.

The limitation is ecosystem maturity. Cartesia is newer than ElevenLabs or Deepgram, with fewer framework integrations and a smaller developer community. The credit-based pricing can also be harder to predict than simple per-character or per-hour models.

3.4 AssemblyAI (Score: 7.2/10)

AssemblyAI is the STT specialist with the richest feature set in the market. Universal-2 achieves 2.4% WER on clean audio across 99 languages, and the current flagship, Universal-3.5 Pro, is the model to beat in independent testing - AssemblyAI Benchmarks.

The evidence is unusually specific. In Coval's July 2026 evaluation, Universal-3.5 Pro led all 31 models tested on overall word error rate at 3.4%, and topped the field on clipping and far-field audio, the two conditions that break most transcription in the real world. On Daily's Pipecat benchmark it sits on the Pareto frontier at 1.22% pooled semantic WER with a 282ms median time to final segment - AssemblyAI. As of the July 21, 2026 measurement window, Universal-3.5 Pro Realtime was the only model inside what AssemblyAI calls the human parity zone: the 2-4% WER a professional transcriber achieves under good conditions, delivered inside the roughly 200ms gap a person leaves before replying - AssemblyAI.

What sets AssemblyAI apart is the audio intelligence layer built on top of transcription. Diarization, sentiment analysis, entity detection, summarization, topic detection, PII redaction, content moderation, and a medical mode are all available as add-ons. For agents that need to understand conversations (not just transcribe them), this feature depth is unmatched.

Pricing is competitive: Universal-2 at $0.15/hour (batch), Universal-3.5 Pro at $0.21/hour (batch) and $0.45/hour (streaming), with Universal-Streaming at $0.15/hour in both English and multilingual. Add-ons range from $0.02/hour (async diarization) to $0.15/hour (Medical Mode or content moderation). The $50 free credit covers initial evaluation - AssemblyAI Pricing.

For agents that need pure STT with maximum intelligence, from meeting transcription bots to compliance monitoring systems, AssemblyAI's combination of accuracy, features, and price is the strongest in the market.

3.5 OpenAI (Score: 7.0/10)

OpenAI provides both TTS and STT with the simplest integration path for teams already using the OpenAI API. TTS-1 scores MOS 4.7 and TTS-HD reaches even higher quality. Whisper STT at $0.36/hour ($0.006/minute) and gpt-4o-mini-transcribe at $0.18/hour ($0.003/minute) make OpenAI the cheapest major STT provider for batch processing - OpenAI Pricing.

TTS pricing is competitive at $15/1M characters for TTS-1 (standard) and $30/1M for TTS-HD. Thirteen built-in voices with no voice cloning. gpt-4o-mini-tts combines TTS with LLM intelligence at $12/1M output tokens, letting the model control speech style and emphasis dynamically, and the gpt-audio family (including gpt-audio-1.5) handles speech in and speech out in one model at $32/1M audio input tokens.

The streaming line has since split out from Whisper. gpt-live-transcribe and gpt-realtime-whisper both run at $0.017/minute ($1.02/hour), roughly 6x the batch rate, which is the real cost to model when a voice agent transcribes continuously rather than in files - OpenAI Pricing.

The Realtime API enables full voice agent workflows with integrated STT + LLM + TTS in a single WebSocket connection. For teams building on the OpenAI API, this is the zero-friction path to voice agents. The main limitation is the lack of voice cloning and the relatively small voice library (13 voices vs hundreds from ElevenLabs or Google).

3.6-3.10: Google Cloud, Inworld, Gladia, Azure Speech, Hume AI

The remaining five in the top 10 each serve a specific niche. Google Cloud (6.8) offers the most generous free tier (4M characters/month on standard voices, 1M on WaveNet) and broadest language coverage (125+ STT languages), with transcription at $0.016/minute and a dynamic batch tier 75% cheaper again for work that can wait up to 24 hours - Google Cloud. Inworld (6.7) publishes the lowest time-to-first-byte of any commercial TTS, 25ms at P99 on Realtime TTS-2 Flash, and clones a voice from 5 to 15 seconds of audio that then speaks across 200+ languages, from $15/1M characters - Inworld. Gladia (6.6) bundles every feature (diarization, NER, sentiment) into the base price with no add-on charges, at sub-300ms streaming latency on Solaria-3, $0.61/hour async on Starter falling to $0.20/hour on Growth - Gladia. Azure Speech (6.6) has the widest TTS language support (140+) with enterprise compliance, at $16/1M characters and $1/hour for transcription. Hume AI (5.8) is still the only emotion-aware TTS, but it is not the bargain it is often described as: pricing is plan-based, at $70/month for 1M characters on Pro and an effective $50/1M on the $500/month Business plan, with overages from $0.05 to $0.15 per 1,000 characters - Hume.

---

4. The Full Provider Directory (47 Services)

Beyond the top 10, we mapped 37 additional providers across four categories. For brevity, we list them with their key differentiator and pricing anchor. The list is shorter than it would have been a year ago: PlayHT, one of the best-known independent TTS APIs, wound down after Meta acquired its team in July 2025, shutting off the API and deleting hosted voice clones on December 31, 2025 - TechCrunch. Voice AI is consolidating as fast as it is growing, which is an argument for choosing a provider whose voices you can export or reproduce.

Cloud TTS APIs (11 more): Amazon Polly ($4.80-30/1M chars, cheapest standard voices), LMNT (150ms latency, no rate limits), Resemble AI (emotion control focus, $0.006/sec), WellSaid Labs ($0.0025/min enterprise), Murf AI ($10/1M chars Falcon model), Speechify ($10/1M chars PAYG, and its Simba 3.2 model is the closest challenger to Cartesia on the Artificial Analysis arena), Rime AI (sub-200ms, on-premise option), Smallest AI ($0.01/min, cheapest real-time), Neets AI ($1/1M chars, cheapest cloud TTS), Unreal Speech ($49/mo for 1M chars), MiniMax (ELO 1,156, $60-100/1M chars).

Cloud STT APIs (4 more): Soniox ($0.10/hr, cheapest commercial STT), Amazon Transcribe ($1.44/hr, deep AWS integration), Rev AI ($0.003/min API tier, hybrid AI+human option), Speechmatics ($0.21-0.45/hr, best diarization).

Voice Agent Platforms (5): Vapi ($0.07-0.33/min total, provider-agnostic orchestration), Bland AI ($0.09/min connected calls, phone specialist), Retell AI ($0.07-0.08/min, no platform fees), LiveKit ($0.01/min, open-source WebRTC), Vocode (free open-source core, modular).

Open Source (10): Whisper/whisper.cpp/faster-whisper (2.5% WER, runs on CPU), Kokoro 82M ($0.70/1M hosted, MOS 4.5, Apache 2.0), Sesame CSM (MOS 4.7, conversational focus), Orpheus 3B (MOS 4.6, emotion tags), Fish Speech S2 (MOS 4.4, 15K+ style tags), Qwen3-TTS (97ms latency, 10 languages), Dia 1.6B (multi-speaker dialogue), Bark (music + speech, MIT), Piper (30ms edge, Raspberry Pi), XTTS v2 (voice cloning from 6s, community-maintained).

---

5. TTS Pricing Comparison

The 100x price gap between Neets AI ($1/1M chars) and ElevenLabs v3 ($100/1M chars) is the most extreme in any AI API category. The question is whether the quality difference justifies it. On MOS benchmarks, the gap is 4.8 vs approximately 4.0, a meaningful but not transformative difference for most agent use cases.

6. STT Pricing Comparison

The hyperscalers (Google, Azure, Amazon) are consistently the most expensive for STT, often 3-10x more than specialized providers. This is because their pricing was set for the pre-agent era when transcription was a low-volume enterprise feature. The agent-native providers (Soniox, AssemblyAI, Deepgram) have priced for volume from day one. Note that the chart is batch pricing: streaming costs meaningfully more at every vendor. Deepgram's Nova-3 goes from $0.26/hour pre-recorded to $0.46/hour streaming, and OpenAI's live transcription models run $0.017/minute ($1.02/hour) against $0.006/minute for the same audio submitted as a file.

7. Benchmark Data: Quality, Latency, and Accuracy

The independent benchmark landscape for voice AI has matured significantly in 2026. Three sources provide the most reliable quality data.

CodeSOTA maintains a continuously updated leaderboard of TTS MOS scores and STT WER measurements across standardized test sets. Their key finding: the top 6 TTS providers (ElevenLabs, Sesame CSM, OpenAI HD, Gemini, Cartesia, ElevenLabs Flash) all score within a 0.2 MOS range (4.6-4.8). At this level, quality differences are subtle and often preference-dependent rather than objectively measurable - CodeSOTA.

Artificial Analysis Speech Arena uses ELO ratings from blind human comparisons, and it changed hands in August 2026. Cartesia's Sonic-3.6 now leads the Provider Voice board at 1,282 Elo, 43 points clear of Speechify's Simba 3.2, and also leads the Controlled Voice board at 1,120, ahead of Sonic-3.5 (1,095) and ElevenLabs Eleven v3 (1,066) - Artificial Analysis. Inworld TTS-1.5-Max, which had held the top provider-voice slot at 1,236, was displaced, though Inworld's newer TTS-2 still leads the separate realtime arena. The Elo gaps are more meaningful than MOS gaps because they reflect head-to-head human preference rather than absolute quality scores - Inworld Benchmarks.

For STT, the critical insight from AssemblyAI's research is that "WER is fundamentally broken" as a metric - AssemblyAI. A 2% WER on clean LibriSpeech audio translates to 8-15% WER on real-world conversational audio with background noise, accents, and crosstalk. When evaluating STT providers, test on YOUR data, not on benchmarks.

8. Open Source: The Quality Gap Has Closed

The most significant development in voice AI in 2025-2026 is the collapse of the quality gap between open-source and commercial TTS models. In 2023, the best open-source TTS scored approximately 1.0 MOS below the best commercial API. By 2026, that gap has narrowed to 0.1 MOS: Sesame CSM (open source) at 4.7 vs ElevenLabs at 4.8 - CodeSOTA.

Kokoro 82M deserves particular attention. With only 82 million parameters (small enough to run on any hardware), it achieves MOS 4.5, ranks #1 on HuggingFace TTS Spaces Arena, and costs roughly $0.70/1M characters when hosted, 85x cheaper than ElevenLabs. The Apache 2.0 license means full commercial use with no restrictions - HuggingFace.

For STT, Whisper Large v3 Turbo (809M params) achieves 2.5% WER with 8x speed improvement over the original, and faster-whisper (CTranslate2 port) adds another 4x speed boost. Self-hosted Whisper on a modern GPU processes audio at 50-100x real-time, meaning an hour of audio transcribes in under a minute.

The practical implication: if you have GPU infrastructure and the engineering capacity to self-host, open-source models now deliver 90%+ of commercial quality at 1-5% of the cost. The commercial APIs' value proposition has shifted from quality superiority to operational convenience (no infrastructure to manage, SLAs, compliance certifications).

9. Voice Agent Platforms: The Full-Stack Option

For teams that want voice agents without assembling the STT + LLM + TTS stack themselves, five platforms offer turnkey solutions. The economics are different from buying components individually.

Vapi ($0.07-0.33/minute total) is the most flexible: bring your own STT, LLM, and TTS providers, and Vapi orchestrates the pipeline. Retell AI ($0.07-0.08/minute base, ~$0.13-0.31/minute with LLM and telephony) is the simplest, with no platform fees and all features included. Bland AI ($0.09/minute connected calls) specializes in phone call automation. LiveKit ($0.01/minute agent sessions) provides open-source WebRTC infrastructure for self-hosted voice agents. Vocode is fully open-source with the most modular architecture.

The build-vs-buy calculus: assembling Deepgram STT + Claude + Cartesia TTS yourself costs roughly $0.05-0.10/minute in API fees but requires significant engineering effort. A platform like Retell adds $0.07-0.08/minute in orchestration fees but eliminates months of infrastructure work. For most teams, the platform premium pays for itself in time-to-market.

As we covered in our guide to building a Claude chatbot, the voice layer adds significant complexity to the conversation engine. For teams building unified APIs that abstract this complexity, platforms like Suprsonic provide TTS and STT as part of a broader capability set, alongside search, scraping, enrichment, and 15+ other agent capabilities through a single API key.

10. How to Choose: Decision Framework

If quality is everything (premium brand, luxury, high-stakes): ElevenLabs. MOS 4.8, full agent platform, voice cloning. Accept the premium pricing.

If latency is critical (real-time phone calls, gaming, emergency): Cartesia or Inworld. Cartesia's 40ms TTFA and Sonic-3.6 at under 90ms in production are 5-10x faster than most competitors, and Inworld's Realtime TTS-2 Flash publishes 25ms time-to-first-byte at P99. Both numbers are vendor-measured under different methodologies, so benchmark them on your own audio before committing.

If cost matters most (high volume, background processing): Soniox ($0.10/hr) or AssemblyAI ($0.15/hr) for batch STT, Deepgram ($0.26/hr pre-recorded) if you also want the agent stack, and OpenAI TTS-1 ($15/1M chars) or Inworld Realtime TTS-2 Flash ($15/1M) for TTS. Or self-host Kokoro ($0.70/1M) and Whisper (free) if you have GPU infra.

If you need the full pipeline (STT + LLM + TTS, one vendor): ElevenLabs Conversational AI 2.0, Deepgram Voice Agent API (whose Flux model handles turn-taking inside the recognizer rather than with a silence timer), or a platform like Retell/Vapi.

If STT features matter (diarization, sentiment, medical): AssemblyAI. Richest feature set, medical mode, and the only realtime model independent testers placed inside the human parity band in 2026.

If language coverage is primary (global deployment): Azure (140+ TTS languages), Inworld (200+ via cross-lingual cloning), Google (125+ STT languages), AssemblyAI (99 STT languages).

If you want open source (full control, privacy, edge): TTS: Kokoro (best quality/size), Sesame CSM (best MOS), Orpheus (best emotion). STT: faster-whisper (best speed), Whisper Large v3 (best accuracy).

Yuma Heymans (@yumahey), who builds agent infrastructure at O-mega and has deployed voice agents across customer support and sales workflows, notes that the most common mistake teams make is optimizing for TTS quality when their bottleneck is actually STT accuracy or LLM latency. In a voice agent pipeline, the slowest and least accurate component determines the user experience, regardless of how good the other components are.

Every price and benchmark in this guide was checked against the provider's own pricing page or model card in September 2026. Voice AI is evolving rapidly, and promotional rates in particular expire. Verify current details on official pricing pages before purchasing.