Speech-to-speech vs cascaded pipelines

Speech-to-speech models Voiceflint can run

For comparison the balanced cascade is $0.019 per minute in models with a 700–900 ms voice-to-voice band.

Task completion: cascades still lead

Daily's February 2026 benchmark of language models for voice agent use cases ran the same multi-turn tasks through text models and speech-native ones. GPT-4.1 in a cascade passed 94.9% of scenarios; GPT Realtime, the same family in speech mode, passed 86.7%. Ultravox was the first speech-native model to perform well on long multi-turn calls. The gap is closing, and it is still there.

Latency: not the win you expect

Hosted speech-to-speech models take 0.6 to 1.5 s to first audio on the Artificial Analysis board once their thinking modes are included. A cascade with fused end-of-turn detection, a fast text model and a flash voice lands at 0.6 to 0.9 s. Full-duplex research models such as NVIDIA's PersonaPlex show 170 ms is possible if you self-host, at a quality cost.

Cost and control

Speech-to-speech bills the listening minute and the speaking minute at audio-token rates. Grok Voice Think Fast 2.0 works out to $0.060 per minute, against $0.019 for the balanced cascade. In a cascade you can also swap any one line, read the transcript the model saw, redact it, and cache the prompt. In speech-to-speech you get the vendor's voice list and its behaviour.

When to pick speech-to-speech

• Short, emotional or sales conversations where tone carries the outcome.
• Callers who interrupt a lot; native models handle barge-in without a turn detector.
• Languages where your cascade's STT or TTS is weak but the speech model is strong.
• On Voiceflint the speech-to-speech preset runs Gemini 3.1 Flash Live and switches back to a cascade with one click, so you can A/B it on real calls and read both receipts.

Questions

Can I use tools with speech-to-speech models?

With the ones marked yes above. The tool definitions are the same JSON Schema functions you use in a cascade.

Can I mix them?

Not inside one call today. You can route different agents or phone numbers to different modes and compare the receipts.