Skip to main content
Updated
Speech-to-speech models take audio in and talk back in real time - the engines behind voice agents and live assistants. The catch: the best-sounding, smartest models often aren’t the fastest to respond, and price swings widely. We compared 15 on quality, speed, price, and access.

Best Speech-to-Speech Models

#ModelBest for
The score reflects model quality from the Artificial Analysis Speech to Speech benchmark suite. Time to First Audio is a separate measure of how quickly audio starts, not overall quality - a top scorer can still be slow to answer.

How to Choose

When choosing between these models, consider:
  • Access: First decide whether you want an app, an API, or a model you run yourself, because that changes cost, privacy, latency, and setup work. Most main picks are hosted services. Step-Audio R1.1 supports infrastructure-heavy self-hosting, PersonaPlex is the main high-end local option, and the weaker Moshi also runs on a typical machine.
  • Quality: We use the Artificial Analysis Speech to Speech benchmark suite as the main score - an equal-weighted look at speech reasoning, conversational dynamics, and agentic voice performance. It measures how good the model is, not how fast it responds.
  • Price: We compare USD per hour of input audio. We use Artificial Analysis’s calculated hourly cost where available; otherwise the value is the listed input-audio rate, so output charges may still apply.
  • Time to First Audio: This is how quickly audio starts, averaged across benchmark runs - not total call latency. Lower feels more human. A strong model can still start slowly (Qwen3.5 Omni Plus Realtime and Gemini 3.1 Flash Live both do), which matters a lot for snappy, interactive agents.

Other Models We Considered


Frequently Asked Questions

GPT-Realtime-2. It’s the top scorer and the most reliable at complex, tool-driven conversations, so it’s the default recommendation for demanding production voice agents. The main reason not to use it is cost on very high-volume, simple traffic.
For most new builds, GPT-Realtime-2 if you want peak quality, or Gemini 3.1 Flash Live if you want strong reasoning at a much lower hourly price and can accept a slower start. For high-volume lightweight voice, GPT-Realtime mini or Qwen3.5 Omni Flash Realtime keep costs down.
Step-Audio R1.1. It competes with hosted leaders on reasoning under an Apache-2.0 license, and you can use it through first-party app and API routes. Just know that self-hosting it is infrastructure-heavy, not a local-machine task.
PersonaPlex is our main local pick, and it needs a high-end GPU (A100/H100-class), not a laptop. Moshi runs on a typical machine but is far weaker. For serious quality, a hosted model is still the better route.
Deepslate Opal starts talking faster than anything else on this list, which makes conversations feel close to instant. It’s only mid-tier on reasoning, though, so it’s best where responsiveness matters more than deep capability.
For quality, mostly yes - the score tracks how well a model reasons and holds a conversation. But it says nothing about speed. Always check Time to First Audio too, since a high-scoring model like Qwen3.5 Omni Plus Realtime can still feel sluggish in a live call.
They’re close. GPT-Realtime-2 edges ahead on overall quality and the hardest reasoning, while Grok Voice Think Fast 1.0 posts the strongest agentic, tool-using results. If your agent mainly takes actions and calls tools, test Think Fast; for the toughest reasoning, GPT-Realtime-2.
GPT-Realtime-2 for most cases - it’s more capable than both and much cheaper than GPT-Realtime. If you need lower latency or lower cost within the same realtime family, look at GPT-Realtime-1.5 for speed or GPT-Realtime mini for budget.