Voice Bot Requires Classifier Like Jev for Improved Performance
November 3, 2026 · 2 min read
Also published in Português (Brasil), Français, Nederlands, Norsk, עברית, Deutsch, Svenska, Polski
A voice bot utilizing a large language model (LLM) experiences challenges during real calls, often failing to record clients or interrupting speakers. To address these issues, it is suggested that certain functionalities be delegated to a fast classifier model. Currently, the most discussed classifier is Jev from TypeSafe, which has attracted interest due to its open-source alternatives.
Between 2022 and 2024, before large language models became widely accessible, a classifier named Voximplant Avatar was developed to extract intent from phrases and identify structured data within them. While it performed adequately, the advent of large language models led to a belief that classifiers were no longer necessary. However, practical applications soon revealed that these models struggled with lengthy prompts and topic changes, leading to inaccuracies.
Developers increasingly relied on multiple agents and multi-stage bots to enhance performance, as a single model often failed to maintain context, leading to excessive token usage. The Voximplant Avatar eventually became outdated, prompting a transition to a new architecture utilizing a smaller, faster language model coinciding with the release of Jev.
Jev, a decision model from TypeSafe, was made available in early access in September 2026. It processes input data such as transcripts and scenario steps, returning formatted responses without generating textual output. The specific use of the structured output is determined by the bot developer. Jev's response times vary but are generally reported to be within 500 milliseconds, with some measurements indicating performance as fast as 170 milliseconds.
In addition to Jev, several other classifiers have emerged recently, including Laya from Convai Innovations, Clef from Cloudflare, CMF from Cortiq, Bespoke Nimble from Bespoke Labs, and CLM-8B, a collaboration between Stanford and Nvidia. These models aim to enhance classification tasks and improve the efficiency of voice bots.
The typical scenario in a voice menu involves multiple topics under one number, which can overwhelm average models. To mitigate this, employing a quick classifier to connect specialized agents for each topic is proposed. This approach allows for swift topic changes and reduces reliance on larger models, optimizing both performance and token usage.