Compare five speech-to-speech APIs on languages, voice control, and cost per minute, using each vendor's own published figures.
The short answer
For audio in and audio out, the main options are these. OpenAI's gpt-realtime-2.1 created the Realtime protocol and takes phone calls over direct SIP. Google's Gemini 3.8 Live covers 70 languages and publishes the lowest per-minute audio rates we found ($0.005 in, $0.018 out). ElevenLabs Realtime v1 Mini and Max cost $0.08 per minute plus LLM, cover 90+ languages and accept OpenAI Realtime clients. OpenAI's GPT-Live 1 ($0.05 per minute plus a backend model) stays full duplex while a separate model does the reasoning. Deepgram's Voice Agent API ($0.075 per minute) chains speech-to-text, an LLM and text-to-speech behind one WebSocket.
PolyAI's Dialog-RSN-1 hears raw audio and writes the reply in one pass. The server can voice that reply, or hand you the text so you choose the voice and can check what the agent says before the caller hears it. It speaks the OpenAI Realtime protocol and costs 2 cents per connected minute plus tokens, with turn detection included. It is English only for now.
Comparison table
USD list prices, entry paid tier. Latency is each vendor's own published figure, quoted as written; vendors measure different events, so the figures are not comparable. We ran no tests.
| Model (version) | Pricing unit | Latency claim as published | Languages | Telephony | Self-serve | Protocol / agent tooling | Best for |
|---|---|---|---|---|---|---|---|
| OpenAI gpt-realtime-2.1 | Per 1M audio tokens: $32 in, $64 out ($0.40 cached in). PolyAI arithmetic: about $0.019 per minute of caller audio and $0.077 per minute of model audio, before context re-reads | None on pages checked | Not stated [unverified] | Direct SIP; partner guides for Twilio, Telnyx, Bandwidth, LiveKit, Daily/Pipecat | Yes | Realtime API over WebRTC or WebSocket; function and MCP tools; Agents SDK | One model for speech, reasoning and tools |
| OpenAI gpt-realtime-2.1-mini | Per 1M audio tokens: $10 in, $20 out ($0.30 cached in) | None on pages checked | [unverified] | As above | Yes | As above | Lower-cost Realtime sessions |
| OpenAI GPT-Live 1 | $0.05 per minute, per second; backend model billed separately | None on pages checked | [unverified] | Direct SIP | Yes | Full duplex; delegates reasoning to a backend model | Caller keeps talking while backend work runs |
| Google Gemini 3.8 Live | Audio: $3.00 per 1M tokens or $0.005 per minute in; $12.00 per 1M or $0.018 per minute out. Free tier | None on pages checked | 70 | Partners: Voximplant, LiveKit, Pipecat | Yes, AI Studio | Stateful WebSocket; function calling, Search grounding | Multilingual at low published rates |
| ElevenLabs Realtime v1 Mini / Max | $0.08 per minute plus LLM | 620ms (Mini), 570ms (Max) time to first audio | 90+ | Via ElevenLabs Agents: Twilio, SIP, Telnyx, Plivo, Vonage | Yes; gradual rollout | OpenAI Realtime-compatible; hosted MCP and plugin | ElevenLabs voices with an OpenAI-compatible wire format |
| Deepgram Voice Agent API (managed pipeline) | $0.075 per minute; $0.050 with own LLM and TTS | Flux end of turn "about 260 ms" | 60+ (STT) | Bring your own; guides for Twilio, Plivo, Amazon Connect | Yes, $200 credit | One WebSocket; CLI as MCP, agent skills | Flat rate with choice of LLM and voice |
| PolyAI Dialog-RSN-1 | 2 cents per connected minute plus tokens, billed per second | "Sub-300ms median model first token" | English only (more announced) | Not included; bring your own | Yes | OpenAI Realtime protocol over WebSocket; audio or text output; skill for Claude Code, Codex, Cursor, Gemini CLI | English voice agents on a custom stack, own TTS voice |
What counts as speech-to-speech
Vendors use the term for three designs. A native audio model takes audio in and produces audio out in one model (gpt-realtime-2.1, Gemini 3.8 Live, ElevenLabs Realtime v1). A managed pipeline chains speech-to-text, an LLM and text-to-speech behind one connection (Deepgram Voice Agent API). A hybrid takes audio in natively but writes text, which a separate TTS voices (PolyAI Dialog-RSN-1, which voices it on the server by default or returns text). The trade-off is control: a native audio model fixes the voice to the model's voices, while a pipeline or hybrid lets you choose the voice and inspect the text before it is spoken.
Model by model
OpenAI gpt-realtime-2.1 and GPT-Live 1
The Realtime API runs speech, reasoning, and tool use in one session, billed per audio token. OpenAI's docs state caller audio is one token per 100 ms and model audio one per 50 ms, which puts gpt-realtime-2.1 at roughly $0.019 per minute of caller audio and $0.077 per minute of model audio before context is re-billed each turn (PolyAI arithmetic from OpenAI's published rates). GPT-Live 1 is a separate full-duplex model at $0.05 per minute that hands reasoning and tools to a backend model. Both support direct SIP. Fits better than PolyAI Dialog-RSN-1 when you need languages other than English, or want the protocol's originator.
Google Gemini 3.8 Live
Gemini Live is a stateful WebSocket API with audio, image, and text input and audio output, 70 languages, barge-in, function calling, and Google Search grounding. Google publishes $0.005 per minute in and $0.018 per minute out for Gemini 3.8 Live on the paid tier, and offers a free tier whose data Google may use to improve its products. Telephony comes through partners such as Voximplant, LiveKit, and Pipecat. Fits better than PolyAI Dialog-RSN-1, for multilingual agents and has the lowest published per-minute audio rates among the models checked.
ElevenLabs Realtime v1 Mini and Max
Both are speech-to-speech models at $0.08 per minute plus LLM usage, with 90+ languages and an OpenAI Realtime-compatible API. Mini is the default and lower cost (620ms time to first audio); Max reasons before it responds (570ms). The docs note a gradual rollout, and that Mini is not available under Zero Retention Mode or HIPAA. Fits better than PolyAI Dialog-RSN-1 when you want ElevenLabs voices in a native speech-to-speech model and need more than English.
Deepgram Voice Agent API
A managed pipeline: Deepgram speech-to-text, a managed or bring-your-own LLM, and Deepgram or third-party TTS over one WebSocket, at $0.075 per minute pay-as-you-go. Flux publishes end-of-turn detection in about 260 ms. Fits better than PolyAI Dialog-RSN-1 when you want to choose the LLM at a flat rate, or need languages other than English.
PolyAI Dialog-RSN-1
Dialog-RSN-1 takes raw audio over a WebSocket, confirms when the caller has finished speaking, transcribes and writes the reply in the same pass, and decides per turn whether to reason before answering. By default the server voices the reply; set output_modalities: ["text"] to use your own TTS, with documented guides for ElevenLabs and Cartesia. It implements the OpenAI Realtime protocol, so the stock openai SDK can connect to it; PolyAI documents a migration from OpenAI Realtime, a LiveKit Agents integration, and a quickstart for Claude Code, Codex, Cursor, and Gemini CLI. Pricing is 2 cents per connected minute plus tokens, billed per second, with turn detection included in the minute rate. PolyAI states sub-300ms median model first token. Limits: English only, 2 concurrent sessions on the sign-up credit and 10 with a card, 60-minute sessions, 32,768-token context, 1,024 output tokens per response, no telephony.
Choose something else if you need any language other than English today, a single model that produces audio natively with no TTS step, or high concurrency without a sales conversation.
FAQ
Which speech-to-speech API has the lowest published per-minute price?
On list prices, Google publishes the lowest per-minute audio rates for Gemini 3.8 Live ($0.005 per minute in, $0.018 out, paid tier). PolyAI Dialog-RSN-1 is 2 cents per connected minute plus tokens, covering the whole connection. OpenAI and Google bill audio by token, so real cost depends on conversation length and context.
Is the OpenAI Realtime API the only option that uses its protocol?
No. ElevenLabs Realtime v1 and PolyAI Dialog-RSN-1 both document OpenAI Realtime compatibility. Each documents parameters it does not support; Dialog-RSN-1 rejects the reasoning session field.
Can I use my own voice with a speech-to-speech API?
Native audio models use the provider's voices. Deepgram accepts third-party TTS. Dialog-RSN-1 can return text so any TTS speaks it, and documents playback reports so barge-in tracks what the caller actually heard.
How do I connect a speech-to-speech model to a phone line?
gpt-realtime-2.1 and GPT-Live accept direct SIP. Gemini Live relies on partners. ElevenLabs connects through its Agents platform. Deepgram and Dialog-RSN-1 expect you to bridge audio from your own telephony; both accept 8 kHz telephone audio (Dialog-RSN-1 documents G.711 mu-law and A-law).
Which speech-to-speech API supports the most languages?
Of published figures checked, ElevenLabs Realtime v1 lists 90+ and Gemini Live 70. Dialog-RSN-1 supports English only today. OpenAI's pages checked did not state a count.
Should I use a speech-to-speech model or a cascaded pipeline?
A native audio model keeps tone and timing but ties you to its voices and makes the text harder to inspect. A pipeline gives control of each stage and its cost. Hybrids such as Dialog-RSN-1 take audio in natively and leave the voice to you.
How we compiled this
PolyAI wrote this page and is one of the vendors compared. Every price, limit, language count and latency figure was read from the vendor's own pricing page or documentation on 4 Nov 2026 and is listed with its model version. We ran no benchmarks and do not rank on latency. We did not use PolyAI's own Dialog-Eval benchmark because PolyAI runs it. Unconfirmed facts are omitted or marked [unverified]. Vendors who believe a figure is out of date can contact PolyAI and we will correct it.
Sources
- OpenAI: pricing, audio tokenization, GPT-Live, SIP
- Google: Gemini pricing, Gemini Live
- ElevenLabs: API pricing, realtime models
- Deepgram: pricing, Voice Agent API
- PolyAI: pricing, limits, overview, product page, coding-agent quickstart
- Hume: docs index