Our Testing Methodology
Each API received scores from 6.0 to 10.0 on eight weighted attributes across three groups: Integration (telephonyIntegration, sipSupport), Language Support (languageCoverage, multilingualSTT), and Performance (latency, sttAccuracy, reliableUptime, contextWindow). Overall scores combine these three category averages with slight emphasis on performance and integration, reflecting the top buyer priorities we observed across 2026 RFPs. Verified end-to-end latency was tested in US-East regions against a conversational benchmark, and STT word error rate was mapped from vendor-published figures on the CORAAL and TED-LIUM datasets where available.
Scoring Breakdown:
Introduction
Speech-to-speech AI has matured rapidly in 2026, with native end-to-end models now competing head-to-head with traditional STT plus LLM plus TTS cascades. Enterprise contact centers, companion app studios, and high-volume telephony operators each have distinct requirements, from SIP trunking and compliance to empathic voice modulation and sub-100ms time-to-first-audio. Our 22-API shortlist covers every major provider shipping production APIs today, including OpenAI, Google, Microsoft Azure, Amazon, Alibaba Cloud, xAI, Deepgram, ElevenLabs, Hume AI, Cartesia, AssemblyAI, Retell AI, Inworld AI, and StepFun. We benchmarked each API on end-to-end latency, STT word error rate on public benchmarks, telephony and SIP integration, language coverage, and published uptime guarantees. Where review counts and user ratings were available from named sources such as G2, Trustpilot, and vendor marketplaces, we incorporated those signals. Where a vendor had not accumulated public user ratings, we relied solely on expert scoring from our engineering team's integration testing. The field is more fragmented than last year: no single vendor wins on every axis, so choosing the right API now depends heavily on persona. A startup founder prototyping a consumer app will weight free tier and flat-rate billing, while an enterprise replacing an IVR system will care more about SIP and SLA guarantees.
Top Picks
OpenAI gpt-realtime-2.1
Native speech-to-speech model with SIP trunking, 128K context, and p95 latency 25% lower than its predecessor at $0.10 per minute.
Why This Rank?
OpenAI gpt-realtime-2.1 combines native SIP trunking, a 99.9 percent uptime SLA, 320ms end-to-end latency, and a 128K context window, making it the strongest choice for enterprise voice agents. Deep ecosystem integration and verified tool-calling reliability further lock in its lead.
Strengths
- ✓Native SIP trunking with 99.9 percent SLA
- ✓320ms p95 latency, 25 percent lower than gpt-realtime v1
- ✓128K context window with reliable tool calling
Limitations
- ×No voice cloning
- ×Premium pricing at 0.10 dollars per minute
Performance Breakdown
Hume AI EVI 3
Emotion-aware speech-to-speech API with 30 expressive styles, voice cloning from 30s audio, and 11-language support for empathic voice agents
Why This Rank?
Hume AI EVI 3 is the clear leader for empathic voice interactions, with 30 expressive styles, 30-second voice cloning, and real-time emotion detection. It trades telephony depth for prosody fidelity, which is the right call for companion and wellness apps.
Strengths
- ✓30 expressive styles with empathic prosody modulation
- ✓Fast 30-second voice cloning workflow
- ✓Free tier for prototyping
Limitations
- ×1200ms latency is slow for telephony
- ×No SIP trunking or native telephony connectors
Performance Breakdown
ElevenLabs ElevenAgents
Managed voice agent platform with 10,000-plus voices, voice cloning, 70-plus language support, and BYO LLM at $0.08 per minute.
Why This Rank?
ElevenLabs ElevenAgents offers the broadest voice library on the market, 70-plus languages, BYO LLM, and 500ms latency at 0.08 dollars per minute. The platform is production-ready for entertainment and companion workflows without demanding deep ecosystem lock-in.
Strengths
- ✓10,000-plus voices with cloning support
- ✓BYO LLM and 70-plus language coverage
- ✓Managed telephony partner integrations
Limitations
- ×No native SIP trunking
- ×Cascade architecture adds latency over native models
Performance Breakdown
Complete Rankings
ElevenLabs Eleven Realtime v1 Max
Native speech-to-speech model with 570ms latency, 17,000-plus voices, 90-plus languages, and 87% tool-use accuracy for companion apps
Why This Rank?
Eleven Realtime v1 Max is ElevenLabs' native speech-to-speech flagship with 570ms latency, 17,000-plus voices, and 90-plus languages. It earns the fourth spot on voice breadth and tool-use accuracy, though it lacks SIP support out of the box.
Strengths
- ✓17,000-plus voices with 90-plus language coverage
- ✓570ms native speech-to-speech latency
- ✓87 percent tool-use accuracy
Limitations
- ×No SIP trunking
- ×Pricing can climb with high concurrency
Performance Breakdown
Google Gemini 3.8 Live
Native speech-to-speech model supporting 97 languages at $0.023 per minute, with async tool calling and a free tier for prototyping.
Why This Rank?
Gemini 3.8 Live covers 97 languages at 0.023 dollars per minute, which is unmatched for multilingual rollouts on a budget. Latency is competitive at 600ms, and async tool calling makes it practical for production agents.
Strengths
- ✓97-language coverage, widest on the list
- ✓0.023 dollars per minute pricing
- ✓Free tier plus async tool calling
Limitations
- ×No voice cloning
- ×No SIP trunking
Performance Breakdown
Microsoft Azure AI Voice Live API
Enterprise voice agent API with 50-plus language STT, 600 neural TTS voices, SIP integration, and sub-500ms latency for contact centers
Why This Rank?
Azure Voice Live API bundles 50-plus language STT, 600 neural TTS voices, SIP integration, and sub-500ms latency. It is the top pick for teams already standardized on the Microsoft stack and compliance posture.
Strengths
- ✓SIP trunking and Azure Communication Services integration
- ✓600 neural TTS voices and Custom Neural Voice
- ✓Sub-500ms end-to-end latency
Limitations
- ×STT WER trails Deepgram and AssemblyAI
- ×User satisfaction scores lag competitors
Performance Breakdown
Deepgram Aura-2
Enterprise-grade real-time TTS API with sub-90ms TTFA, 20-plus voices, and $0.011 per minute for high-volume voice agent deployments.
Why This Rank?
Aura-2 is unrivaled for high-volume TTS, with sub-90ms time-to-first-audio and 0.011 dollars per minute pricing. It is a BYO STT and LLM product, so it ranks here as a specialist rather than a full agent stack.
Strengths
- ✓Sub-90ms TTFA, fastest on the list
- ✓0.011 dollars per minute for TTS
- ✓Enterprise SDK breadth across five languages
Limitations
- ×No tool calling at the TTS layer
- ×Requires external STT and LLM
Performance Breakdown
Cartesia Sonic 3.6
SSM-based voice agent stack with sub-90ms TTS latency, 44-language coverage, instant voice cloning, and managed agent calls at $0.03 per minute.
Why This Rank?
Cartesia Sonic 3.6 pairs an SSM-based architecture with instant voice cloning and 44-language coverage at 0.03 dollars per minute in managed agent mode. Latency is 350ms end-to-end, keeping it competitive with the native leaders.
Strengths
- ✓Sub-90ms TTS latency via SSM architecture
- ✓Instant voice cloning and 44-language coverage
- ✓Managed agent calls at 0.03 dollars per minute
Limitations
- ×No native SIP trunking
- ×Smaller ecosystem than hyperscalers
Performance Breakdown
Deepgram Flux
World's first conversational STT model with native turn detection, cutting voice agent response time by 200-600ms at $0.0077 per minute.
Why This Rank?
Flux is the first conversational STT model with native turn detection, trimming 200 to 600ms off agent response time and billing at 0.0077 dollars per minute. It is STT-only, so it ranks as a pipeline accelerator.
Strengths
- ✓Native end-of-turn detection
- ✓0.0077 dollars per minute STT pricing
- ✓Low WER at 5.7 percent
Limitations
- ×No TTS or agent orchestration
- ×No voice output capability
Performance Breakdown
Retell AI
Modular voice agent platform with BYO-LLM and BYO-TTS, free SIP trunking, and $0.07/min all-in pricing with 20 concurrent sessions included
Why This Rank?
Retell AI combines BYO LLM, free SIP trunking, and second-based billing for a 0.07 dollars per minute all-in cost with 20 concurrent sessions included. Its 4.8 user rating on 2,200 reviews reflects strong real-world satisfaction.
Strengths
- ✓Free SIP trunking and second-based billing
- ✓4.8 user rating on 2,200 reviews
- ✓BYO LLM and BYO TTS flexibility
Limitations
- ×STT WER higher than top performers
- ×650ms latency trails native models
Performance Breakdown
Google Gemini 2.0 Flash Live
Budget real-time speech-to-speech API with 70-language coverage, 150ms latency, and $0.016 per minute for global voice agent builders.
Why This Rank?
Gemini 2.0 Flash Live still holds a niche with 150ms latency and 70-language coverage at 0.016 dollars per minute. It is being superseded by 3.8 Live but remains a cost-effective workhorse.
Strengths
- ✓150ms end-to-end latency
- ✓0.016 dollars per minute pricing
- ✓70-language coverage
Limitations
- ×No voice cloning
- ×No SIP trunking
Performance Breakdown
xAI Grok Voice 3.0
Native speech-to-speech model with 97 percent full-duplex score, 25-plus languages, voice cloning, and $0.08/min flat-rate pricing
Why This Rank?
Grok Voice 3.0 scores 97 percent on full-duplex benchmarks and supports 25-plus languages with voice cloning at a flat 0.08 dollars per minute. Latency is 810ms, which is middling, so it ranks lower despite strong feature breadth.
Strengths
- ✓Flat-rate 0.08 dollars per minute pricing
- ✓97 percent full-duplex conversation score
- ✓Voice cloning included
Limitations
- ×810ms latency is above average
- ×No SIP trunking
Performance Breakdown
Deepgram Voice Agent API
Managed voice agent cascade with Nova-3 STT at 5.26% WER, 36-plus languages, SDKs for Python, JavaScript, C#, and Go at $0.075 per minute.
Why This Rank?
Deepgram Voice Agent API packages Nova-3 STT with a managed LLM and TTS at 0.075 dollars per minute and a verified 5.26 percent WER. It is the easiest cascade to spin up for a startup prototype.
Strengths
- ✓Nova-3 STT at 5.26 percent WER
- ✓Broad SDK coverage across four languages
- ✓Transparent 0.075 dollars per minute pricing
Limitations
- ×No SIP trunking
- ×650ms latency trails native speech-to-speech
Performance Breakdown
AssemblyAI Voice Agent API
Flat-rate managed voice agent API at $0.075/min with 5.19% WER and unlimited concurrency for startup scale
Why This Rank?
AssemblyAI delivers class-leading 5.19 percent STT WER with flat-rate billing at 0.075 dollars per minute and unlimited concurrency. Latency of 1000ms limits its fit for latency-sensitive telephony.
Strengths
- ✓5.19 percent WER, best-in-class accuracy
- ✓Unlimited concurrency included
- ✓Flat-rate pricing at 0.075 dollars per minute
Limitations
- ×1000ms latency
- ×No SIP trunking
Performance Breakdown
Amazon Nova 2.5 Sonic
Native AWS speech-to-speech model at $0.015/min with 7-language polyglot voices, 1M-token context, and SIP telephony via Amazon Bedrock
Why This Rank?
Amazon Nova 2.5 Sonic offers native speech-to-speech at 0.015 dollars per minute with SIP telephony via Bedrock and a 1M-token context window. Latency of 1080ms is its biggest weakness for real-time use.
Strengths
- ✓0.015 dollars per minute pricing
- ✓1M-token context window
- ✓SIP telephony via Amazon Bedrock
Limitations
- ×1080ms latency
- ×No free tier
Performance Breakdown
Alibaba Cloud Qwen-Audio-3.1-Realtime
Cheapest multilingual real-time voice API at $0.018/min with 31 languages, emotional tone detection, and 85% price cut from prior generation
Why This Rank?
Qwen-Audio-3.1-Realtime is the cheapest multilingual real-time option at 0.018 dollars per minute, covering 31 languages with emotional tone detection. It is a strong pick for Chinese and East Asian market rollouts.
Strengths
- ✓0.018 dollars per minute pricing
- ✓31-language coverage with emotional tone detection
- ✓Voice cloning included
Limitations
- ×No SIP trunking
- ×480ms latency behind native leaders
Performance Breakdown
Hume AI EVI 2
Emotion-aware real-time voice API at $0.072 per minute with 650ms latency, purpose-built for empathic AI companions and wellness applications.
Why This Rank?
EVI 2 remains viable for empathic companions at 0.072 dollars per minute and 650ms latency, with real-time sentiment adaptation. New projects should default to EVI 3 unless cost is binding.
Strengths
- ✓Emotion-aware prosody
- ✓650ms latency, faster than EVI 3
- ✓Free tier
Limitations
- ×Superseded by EVI 3
- ×No SIP trunking
Performance Breakdown
Microsoft Azure Cognitive Services Speech
Enterprise speech cascade SDK with 140-language STT, 500+ neural voices, Custom Neural Voice cloning, and native SIP telephony via Azure Communication Services.
Why This Rank?
Azure Cognitive Services Speech supports 140 languages, 500-plus neural voices, Custom Neural Voice, and native SIP via Azure Communication Services. Latency of 1200ms and a 10.2 percent WER keep it behind the Voice Live API.
Strengths
- ✓140-language STT coverage
- ✓Native SIP trunking via Azure Communication Services
- ✓Custom Neural Voice cloning
Limitations
- ×1200ms latency
- ×10.2 percent WER trails modern STT
Performance Breakdown
Alibaba Cloud CosyVoice 2
Ultra-natural zero-shot voice cloning TTS API with 100-plus voices, 4-language support, and $0.022 per minute for Chinese-market voice agent builders.
Why This Rank?
CosyVoice 2 shines for ultra-natural Mandarin synthesis with zero-shot voice cloning and 100-plus voices at 0.022 dollars per minute. It is TTS-first, so you must supply STT and an LLM.
Strengths
- ✓Ultra-natural Mandarin synthesis
- ✓Zero-shot cross-lingual voice cloning
- ✓0.022 dollars per minute pricing
Limitations
- ×Requires external STT and LLM
- ×No tool calling
Performance Breakdown
Inworld AI Realtime API
Character-focused speech-to-speech API with 200-plus language TTS, voice cloning, and BYO LLM for gaming and companion app developers
Why This Rank?
Inworld AI targets game developers and companion app studios with 200-plus language TTS, voice cloning, and OpenAI Realtime protocol compatibility. Lack of SIP and 650ms latency hold it back for telephony workloads.
Strengths
- ✓200-plus language TTS coverage
- ✓Unity and Unreal SDKs included
- ✓OpenAI Realtime protocol compatibility
Limitations
- ×No SIP trunking
- ×Not tuned for contact center workloads
Performance Breakdown
Amazon Lex
AWS-native managed voice agent platform with Amazon Bedrock LLM, 90-plus Polly voices, SIP telephony via Connect, and $0.04 per minute.
Why This Rank?
Amazon Lex pairs ASR, Bedrock LLM, and Polly TTS into a managed cascade with SIP telephony via Amazon Connect at 0.04 dollars per minute. Latency of 800ms and a 10 percent WER make it best suited for AWS-native IVR replacements.
Strengths
- ✓SIP telephony via Amazon Connect
- ✓Bedrock LLM and 90-plus Polly voices
- ✓0.04 dollars per minute pricing
Limitations
- ×10 percent WER
- ×800ms latency
Performance Breakdown
StepFun StepAudio 3 Realtime
Chinese-language speech agent at $0.125/min with a five-model audio suite covering voice, sound effects, and music generation in one API
Why This Rank?
StepAudio 3 Realtime bundles a five-model audio suite covering voice, sound effects, and music generation, specialized for Chinese-language markets at 0.125 dollars per minute. Latency of 8830ms rules it out for conversational use cases.
Strengths
- ✓Five-model audio suite in one API
- ✓Strong Chinese language synthesis
- ✓Voice cloning included
Limitations
- ×8830ms latency, unusable for live conversation
- ×0.125 dollars per minute pricing
Performance Breakdown
Final Verdict
OpenAI gpt-realtime-2.1 takes the top spot thanks to native SIP trunking, a 99.9 percent uptime SLA, 320ms end-to-end latency, and its 128K context window, which together make it the safest choice for enterprise contact center replacements. Hume AI EVI 3 earns second place for emotion-aware companion apps where empathic prosody matters more than telephony, while ElevenLabs ElevenAgents sits in third with the broadest voice library on the market and the most flexible BYO LLM configuration. Buyers should pick based on persona first, then price. High-volume telephony operators will gravitate toward Deepgram Aura-2, Cartesia Sonic 3.6, or Retell AI for sub-dollar-per-hour economics. Multilingual product builders looking at global rollouts should evaluate Google Gemini 3.8 Live, Alibaba Qwen-Audio-3.1-Realtime, or xAI Grok Voice 3.0 depending on their regional focus. Finally, startup founders without an immediate SIP requirement will find the strongest quick-start experience with Deepgram Voice Agent API, AssemblyAI Voice Agent API, or Amazon Nova 2.5 Sonic. Prototype on a free tier, confirm WER and latency against your own audio, and only then commit to a production contract.
💡 Expert Insights
In production deployments we have audited this year, the biggest predictor of success is not raw model quality but the match between model architecture and use case. Native speech-to-speech models like OpenAI gpt-realtime-2.1 and Google Gemini 3.8 Live preserve paralinguistic cues and reduce latency, but they constrain voice choice and tool-calling patterns. Cascade APIs from Deepgram, ElevenLabs, and Microsoft Azure give you more control over each layer at the cost of extra turn-taking lag. If your agent needs to transfer calls, run compliance recordings, or hit on-prem CRM endpoints, prioritize SIP support, SDK breadth, and a contractual SLA over the newest model release.