Choosing a voice, model and language

The Voice & model card has three fields. They get confused constantly because all three change how a call feels, and they fail in completely different ways, which means the wrong diagnosis sends you to the wrong setting.

Model: bigger is not automatically better#

This is the least intuitive thing on the page, and it catches people out constantly.

A larger model reasons better and answers slower. In a chat window nobody notices half a second. On a phone call, that pause is the most conspicuous thing in the conversation, because human turn taking is fast and a gap reads as confusion.

So a bigger model can genuinely make a call feel worse, even while every individual answer gets better on paper.

Match the model to the job rather than reaching for the best one. Reading opening hours out of a knowledge base does not need the same reasoning as untangling a billing dispute, and paying for the latter on the former costs you money on every call.

Voice: judge it by listening#

Voices differ in pace and warmth as much as in pitch, and none of that survives being read as a name in a dropdown.

Use Talk to your agent and have each candidate say your actual welcome message. Ten seconds of that tells you more than any description, because you are hearing the thing your callers will hear rather than a sample sentence.

• Listen to it say your business name. Some voices handle unusual names badly.
• Listen to a number and a time, because those are what a booking agent says most.
• Listen on a phone if you can, not only on laptop speakers. Phone audio is narrower and it flatters some voices and ruins others.

Language: it sets what the agent expects to hear#

This one is misunderstood more than the other two. Language is not only what the agent speaks. It also tells transcription what to expect.

That is why a wrong language setting usually shows up as the agent mishearing people rather than as it replying in the wrong language. The words arrive at the model already wrong, and the model answers the question it was given.

If callers are misheard often, check this before blaming the model, and check the transcript of a call that went wrong. If the transcript does not match what was said, the problem is upstream of anything the model did.

Change one thing at a time#

Changing the model, the voice and the language together and then judging the call tells you nothing about which change mattered. It is the most common way people spend an afternoon and end up back where they started.

Change one. Hear it. Keep it or revert it. Then move to the next.

Every change here moves what a call costs#

Each component is priced separately by the provider doing the work, so the configuration is what decides the price of a minute. Two agents can hold identical five minute calls and cost very different amounts.

The per call cost breakdown in the call log shows this split out line by line, which makes it the fastest way to see what a change actually did to your costs rather than guessing from the total.