Ertas for Voice Agent Fine-Tuning

The Challenge

Voice agents (AI systems that handle phone calls, drive-through orders, appointment scheduling, and customer service conversations) are rapidly replacing traditional IVR systems. The language model is the brain of every voice agent, responsible for understanding caller intent, generating natural responses, making decisions about call routing, and maintaining coherent multi-turn conversations. Yet most voice agent builders rely on generic language models that do not understand the specific domain, vocabulary, or conversational patterns of the business they serve.

The consequences of using a generic model in a voice agent are immediately apparent to callers. The agent misunderstands industry-specific terms, asks redundant questions because it cannot infer context, generates responses that are too long for natural speech timing, and fails to follow the specific call scripts and escalation procedures the business requires. Latency is another critical factor: voice conversations require sub-second response times, and sending requests to large cloud models introduces perceptible delays that make the conversation feel unnatural. These issues compound caller frustration and drive abandonment rates that undermine the business case for voice AI.

The Solution

Ertas lets voice agent builders fine-tune small, fast language models on domain-specific conversational data. In Ertas Studio, teams train on de-identified transcripts of successful calls, approved call scripts and conversation flows that capture the patterns callers expect, and they can write or correct example exchanges in Data Craft, Studio's guided dataset builder. The fine-tuned model learns the business's terminology, follows its call handling procedures, and gives the short, natural replies that work when spoken aloud.

Hub exports the model as GGUF or full safetensors, so it can run on your own GPUs close to your telephony stack. A small model running locally skips the network round trip to a large cloud model, which helps keep turn-taking natural, and its compute cost per conversation is typically lower than API calls to a large hosted model. Served with Ollama, llama.cpp or vLLM, the model becomes the language model behind voice platforms such as Retell, Vapi or Bland, or behind a custom telephony integration. Call recordings and transcripts fall under consent and privacy rules, so de-identify transcripts before uploading them for training; once the model runs on your own servers, live call content stays with you.

Key Features

Example Workflow

A dental practice management company builds voice agents that handle appointment scheduling for 500 dental offices. They collect 100,000 call transcripts from successful scheduling interactions, covering appointment types, insurance verification questions, schedule negotiations and cancellation handling, remove patient names and identifiers, and upload them to Ertas Studio. In Studio, they fine-tune Qwen 2.5 7B on the dental scheduling domain, training it to understand dental terminology (prophylaxis, periodontal maintenance, crown prep), insurance plan names and the scheduling logic for different procedure types, and they train a smaller Llama 3.2 3B in parallel to compare speed and quality. They test both on a held-out set of real calls, checking whether each model books the right appointment type and keeps replies short enough to speak. The chosen model is exported from Hub and served with vLLM on GPU servers in their data center, behind their voice platform. The team measures response latency, end-to-end completion rate and call duration against the generic model they used before, and routes the calls the model handles poorly to staff while they gather examples for the next training run.