Intel / article

Cascaded Vs Fused Models

Source: elevenlabs.io
Author: ElevenLabs editorial sitemap
Date: Published Oct 5, 2026

Context

A cascaded voice agent leaves readable text between its processing stages. Speech recognition transcribes the caller, a language model writes a response, and speech synthesis voices it. In ElevenLabs' architecture explanation, those handoffs give developers places to inspect errors and apply checks before speech is generated.

The company contrasts this modular arrangement with fused models that combine recognition, reasoning and generation within one network. Separate components let teams replace the without replacing the speech models. The tradeoff described in the article is that converting speech to text loses cues such as intonation, which a cascade must recover or model separately. Architecture tradeoffs.

Key takeaways · AI-distilled

ElevenLabs describes voice-agent design as a spectrum of five architectures: basic cascaded, advanced cascaded, hybrid (acoustic fed to the LLM), sequential fused, and duplex fused, with control and auditability falling as you move toward fused.
The core cascaded advantage is readable text between STT, LLM and TTS: can run at the text layer, failures can be localized, and a new frontier LLM can be swapped in without retraining speech models. Its known cost is lost prosody that must be reconstructed on output.
In ElevenLabs' advanced cascaded design, the LLM tells the TTS how to say a line (for example reassuringly or with urgency), turn-taking draws on the same signals, and speech models are co-located to avoid network hops between components.
The authors argue fused models keep prosody naturally but ship with lighter reasoning cores than frontier LLMs, while duplex models that listen and speak simultaneously are the hardest to guardrail and become unstable in longer conversations.
The piece argues for ElevenLabs' own cascaded choice: it concedes fused systems still lead on prosody, but says advanced cascaded systems close that gap every quarter while fused models have made little progress on complex reasoning or trust.

Terms in this piece · Glossary

reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.

Read the source elevenlabs.io