Choosing a voice is the most enjoyable part of setting up an agent and the part where the least careful decisions get made. People audition voices on a laptop, in a quiet room, reading a clean sentence — and then ship to a phone network that throws away most of the frequency range they were judging.
Audition on the medium
A telephone call is narrowband: roughly 300 Hz to 3,400 Hz. The warmth and air that make a voice sound premium on a laptop live outside that range and simply do not survive.
What survives is articulation. A voice with slightly harder consonants and less breathiness — which can sound a little plain in a browser — is often clearly better on an actual call. The only way to know is to place a test call and listen on your phone, ideally on speaker in a noisy room.
The setting most people get wrong
Speaking rate. Agents commonly ship slightly slow, on the theory that clarity helps comprehension. On a phone call it mostly makes the agent feel ponderous, and it inflates every call's duration — which you are paying for by the minute.
Conversational English runs around 150 to 160 words a minute. If your agent is at 130, every ten-second reply is taking eleven and a half seconds and you are billing for the difference across every call in the batch. Nudging the rate up is one of the cheapest improvements available, and it usually sounds better rather than worse.
Match the voice to the language, not the brand
If your calls are in Hindi, Telugu or Tamil, the voice has to be built for it — an English voice reading transliterated text is instantly wrong to a native speaker, and it undermines everything else about the call. Pick the language first; the voice list follows it.
For code-mixed speech, which is what most Indian business calls actually sound like, test with a real sentence that switches mid-way. That is where a poorly matched voice falls apart.
What to listen for
- Numbers. Phone numbers, salaries, dates. These carry the meaning and are where synthesis most often goes wrong.
- Your own proper nouns. Company name, product names, the recruiter's name. You will hear these on every call.
- The question mark. A voice that does not lift at the end of a question makes every question sound like a statement, and people do not answer statements.
- The first two seconds. That is where the person decides whether this is worth their time.
The honest ranking
Latency beats voice quality. Barge-in beats both. A slightly plainer voice that responds in under two seconds and stops when interrupted will out-perform a beautiful one that does neither — every time, and not narrowly.