Most latency numbers you see for voice agents aren’t what the customer actually feels. Most knowledge setups either bloat every prompt or go stale.
Most TTS engines still mispronounce your own brand name.
Here’s how we changed that.
Latency Now Means What the Customer Feels
Industry reporting typically tracks AI Infra Latency, the sum of STT Latency, LLM Latency, and TTS Latency. That number misses one thing: network latency.
Acefone now reports Perceived Latency- AI Infra Latency + Network Latency. Individual voice agent calls might only exhibit slight noise during calls. At scale, however, network latency stops being just noise. It starts affecting real-life call experiences.
Perceived Latency is now living in the platform Latency tab. It gives teams the number that matches what a customer hears on a live call.
Knowledge That Doesn’t Bloat Every Prompt
Stuffing every reference document into a system prompt gets expensive real fast. Token size balloons, ingestion costs climb, and answers get less reliable.
Acefone’s AI voice agents now run on a two-level Retrieval-Augmented Generation (RAG) system instead. Global knowledge bases apply account-wide, for policy or company-level documents. Local knowledge bases apply to a single product or voice agent.
Documents get chunked, embedded, and stored as vectors. Each query then pulls only the relevant chunk through a semantic search. Prompts stay lean, and hallucination rates drop.
Your Brand Name, Pronounced Correctly
Every TTS provider has a normalization layer, and it doesn’t always get things right. Let’s say you hop on a test drive booking call with “MG Motors” and the voice agent responds, “Milligram Motors”.
TTS Dictionaries let you override that default normalization. Set “MG” to always render as “M-G,” manually or through bulk upload via Lexicon, CSV, or TXT files. It’s a small fix with an outsized impact on how professional a voice agent sounds.
What’s Coming
Acefone’s AI voice agent platform is moving to a full DIY model in September, covering Build, Test, Deploy, and Observe in one place. Voice agent to voice agent transfer is already live.
Testing now spans telephony calls, chat-only LLM testing, and real-time WebRTC streaming. Teams can validate behavior before a single production call goes out.
If pronunciation errors or bloated prompts have cost you calls before, these three fixes were built for exactly that.
