Ask any Sri Lankan shop owner and they will tell you: a large share of WhatsApp enquiries arrive as voice notes, especially from older customers and anyone typing in Sinhala.
A bot that only reads text quietly fails for a big slice of real customers — and fails silently, because the customer just gets no reply.
How it works
- 1The customer records a voice note as normal.
- 2The agent transcribes it, in Sinhala, Tamil or English.
- 3The transcript goes through the same pipeline as a typed message.
- 4The reply comes back as text, in the language the customer spoke.
The customer does nothing differently. They speak, and they get an answer.
Sinhala transcription is the hard part
Transcribing English is close to solved. Sinhala is not. Older speech models return empty text for Sinhala audio, or worse, transcribe it into the wrong script entirely — Kannada characters instead of Sinhala is a real failure mode we hit while building this.
If you are evaluating a tool, send it a Sinhala voice note before you sign anything. English voice notes prove very little.
When it cannot understand
Background noise, a very short clip or a bad connection will occasionally defeat any transcriber. The right behaviour then is a polite apology and a request to type it — not silence, and certainly not a confident answer to something it did not hear.