Six things that break Japanese voice agents

1. Numbers and proper nouns

Test amounts, dates and company names before anything else. A misread price gives the caller wrong information, which outranks every other quality consideration. Check whether a pronunciation dictionary exists and how many entries it holds.

2. Repetition inflates minutes

Elderly callers and noisy environments produce repeated clarification. A call estimated at three minutes becomes six, doubling the cost. Design questions that can be answered yes or no rather than open-ended ones.

3. Honorific register drifts

Japanese has levels of politeness that English-trained generation does not hold steady. Mixing registers inside one call sounds wrong. Fix it with three example responses rather than an instruction to be polite.

4. Apologies create records

An automatic apology can read as accepting liability. Separate empathy from admission, suppress the latter, and if the product cannot control this, do not use it for complaints.

5. Recording disclosure

Calls are recorded and transcribed, so callers must be told, normally in the opening announcement. Confirm storage location, retention period and whether the data is used for training.

6. Latency

Transcription, inference and synthesis stack. On a phone call, half a second of silence makes callers ask whether anyone is there. Fill unavoidable gaps with a short acknowledgement rather than leaving silence.

FAQ